Not slightly outdated. Structurally outdated. The band describes a role that no longer requires the same skills, produces the same output, or holds the same value as it did when someone last opened the spreadsheet.

Salary bands get reviewed once a year. Sometimes every 18 months. The review process involves benchmarking against market data, auditing job descriptions, getting sign-off from finance, and publishing an updated range. It takes weeks. Often months. By the time the new numbers are live, the job underneath them has already shifted.

This isn’t a new observation about bureaucracy being slow. Bureaucracy has always been slow. What’s new is that the thing being measured is changing at a pace that makes the measurement instrument useless. A thermometer that updates once a day works fine for weather. It doesn’t work for monitoring a nuclear reactor. And the laws being written right now to enforce pay transparency are anchored to a model of work that is already cracking.

66% Faster Than the Spreadsheet

The premise behind salary bands is straightforward. Define a role. Benchmark it against the market. Set a floor, a midpoint, and a ceiling. Place people within the range based on experience, tenure, and performance. Review periodically. Adjust. The system assumes that a “Senior Data Analyst” this year looks roughly like a “Senior Data Analyst” last year, with maybe one new tool or a slightly updated responsibility.

That assumption held for decades. It doesn’t hold now.

The World Economic Forum’s Future of Jobs Report projects that 39% of existing skill sets will become outdated or substantially transformed between 2025 and 2030. Not over a generation. Over five years. In a single band-review cycle, the role being reviewed may have already lost or changed nearly two-fifths of the skills it was originally priced around.

Think about what that means for the comp team doing the review. They’re opening a job description that says “Proficiency in Excel, SQL, and statistical analysis.” The person in the role is now doing prompt engineering, AI-assisted segmentation, and automated reporting that used to require a junior analyst working two full days. The description and the reality have quietly divorced. The band is pricing a ghost.

And the speed is accelerating, not stabilizing. In 2025, skills in AI-exposed jobs were changing 25% faster than average. Last year it was 66%. Nobody has a model for what next year looks like, because the tool we use to measure these things (annual benchmarking cycles, compensation surveys that take six months to compile and publish) was designed for a world where roles moved slowly.

There’s a question I keep turning over, which is whether collective bargaining frameworks (in some locations and companies) have anything useful to offer here. Union contracts tend to be even more rigid about pay bands than corporate comp structures. Traditional labor agreements specify pay by classification, and those classifications are negotiated years in advance.

If some companies can’t keep up, it’s hard to imagine that a negotiated contract reviewed every three years would do better. But maybe there’s a version of collective action that could respond faster.

The practical consequence is blunt. The salary band is a lagging indicator. It tells you what a role was worth when someone last looked at it. The gap between that snapshot and reality is growing every quarter.

The Job Description Doesn’t Describe the Job

There’s a more philosophical problem underneath the speed issue, which is probably why most comp teams avoid it.

A salary band is attached to a job description. The description defines the role. The band prices the role. Simple chain. But the chain breaks when the description and the actual work diverge far enough.

If 39% of the skills in a role get replaced or augmented over five years, at what point is it a different role? Not a slightly updated version. A fundamentally different job wearing the same title. A “Content Strategist” writing blog posts by hand in 2023 and a “Content Strategist” managing AI content pipelines, editing model outputs, and building prompt libraries in 2026 share a title. They do not share a job.

The band treats them as interchangeable.

A comp consultant from Spain whom I spoke with last spring (she was running late to the call because her dog had escaped into a neighbor’s yard, which she mentioned twice, apologetically, before we actually started talking) described a calibration meeting at a client company where a manager argued that one of her reports had effectively reinvented the scope of her position. The employee was producing roughly twice the output of peers, had absorbed analytical work previously sent to an outside vendor, and had built internal tools that three other departments were using. The manager wanted to move her up a level.

She told me that the HR/TR team blocked it. The job description for the next level required management responsibilities. The system had no mechanism for someone who’d expanded a role horizontally rather than vertically. The band structure could process seniority. It couldn’t process transformation.

I should note that the employee’s situation was unusual. She had a technical background, most people in her title didn’t, and she’d been experimenting with AI tools for a year before the company formally adopted anything. Most people in the same role hadn’t made comparable changes. But the system’s inability to recognize what she’d done, its inability to even categorize it, felt like a symptom of something larger.

The job description is becoming administrative fiction. Useful for legal compliance and organizational charts. Disconnected from the actual work. And when your pay structure is anchored to a fiction, the pay stops reflecting reality. It starts reflecting the last time someone updated a document.

Same Band, Wildly Different Value

This is where the structural problem becomes personal.

Picture someone who spent evenings and weekends learning AI tools. Not because their company mandated it. Not because there was a training program. Because they saw where things were headed and invested their own time. They got faster. They got better. They started producing work that would’ve required two people a year ago. Their manager noticed. Peers started asking them for help.

Their salary didn’t change.

They’re still sitting inside the same range as someone who does the job the old way, takes twice as long, and produces half the output. The band was designed for fairness. People in the same role should be paid within the same range to prevent discrimination and favoritism. That was a good idea. In most ways it still is.

But “the same role” doesn’t mean what it used to.

A 2023 study from researchers at MIT and Stanford tracked customer service agents using an AI assistant. Average productivity rose 14%, with the largest gains among the least experienced workers. That’s the average. The spread within the group was enormous. Some agents barely changed their workflow. Others transformed it.

When the variation in output within a single title gets wide enough, the band stops creating fairness. It starts creating resentment. The high performer looks at the range ceiling and sees a cap that has nothing to do with their contribution. The low performer looks at the floor and sees protection. Both are right from where they’re standing, which is exactly the problem.

The pattern that follows is predictable. The person who adapted, who invested, who made themselves dramatically more valuable, doesn’t get loud about it. They get quiet. They start asking careful questions about internal mobility. They update their LinkedIn profile on a Sunday afternoon. They’re gone within a year, and the company acts blindsided.

There’s a cost to fixing this too, and I want to be direct about it. If you break the band to pay your AI-skilled performer more, you’ve told everyone else in the same title that the range is negotiable, that some people’s contributions count for more than others, and that the system they were told was fair actually isn’t. That’s a real organizational cost, not a hypothetical one. I genuinely don’t know how you manage that tradeoff well. Every approach I’ve seen, read, or heard about so far creates a new problem while solving the old one.

And this conversation assumes every role is actually growing. Not all of them are. Some positions aren't being transformed by AI. They're being quietly hollowed out. The work is shrinking, the scope is narrowing, and the role is drifting toward something a well-configured automation could handle by next quarter. Nobody sends a memo about it.

The job title stays on the org chart, the band stays in the spreadsheet, and the person in the seat keeps collecting the same paycheck for a role that the company no longer needs in the same way, or at all. These aren't people who failed to adapt. They're people whose roles simply stopped offering anything for AI to augment. There's no upskilling path for a job that's evaporating. And the band, predictably, says nothing about that either.

Pull Out the Bottom Rungs and Call It a Ladder

Everything I’ve described so far affects people already inside the system. But there’s a structural issue underneath that I think will matter more over time.

Entry-level hiring in many industries dropped. Half of reward leaders surveyed by Ravio, a compensation data platform, explicitly cited AI automation as the reason for deprioritizing junior roles.

Salary bands are built as ladders and comp teams. Entry, mid, senior, lead, principal. You start at the bottom, you climb. Each rung has a range. The ranges overlap slightly to create a progression path. The architecture assumes a population at every level, with people entering at the bottom and moving up.

What happens to the architecture when the bottom levels empty out?

Five years ago, an entry-level data analyst posting listed SQL, Excel, maybe some basic Python, and “strong communication skills.” Now postings for the same nominal level want experience with AI-assisted analysis, prompt engineering, and “the ability to evaluate and improve model outputs.”

These are skills that used to live at the mid-level. They’ve been pushed down into entry-level descriptions. But the entry-level band hasn’t moved up to match. And in many companies, the posting exists on paper but nobody’s actually filling it.

You can’t have a progression-based pay architecture when there’s nothing to progress from. A company with six levels and nobody in the first two isn’t running a ladder. It’s running a shelf. People arrive mid-career or they don’t arrive at all, and the comp structure below them is vestigial, a set of ranges for people who don’t exist.

I’m not sure what this means for compensation design long-term. The obvious move is restructuring bands to reflect a workforce that enters at a higher skill floor and has a shorter distance to the ceiling. But I haven’t seen anyone do that well yet. Most companies I’ve heard about are still running the old structure and just leaving the bottom levels empty. The architecture assumes a population that doesn’t show up anymore.

Transparency Was Supposed to Fix This

Pay transparency laws are spreading. Colorado, New York City, Washington State, California, Illinois, whole Europe. The direction is toward more disclosure, more posted ranges, and more visibility. The reasoning is sound: if everyone can see the range, companies can’t quietly underpay people based on gender, race, or negotiation skill. That’s a real problem, and transparency is a reasonable response.

But transparency assumes the range means something stable. It assumes that “$XK to $YK for Software Engineer” describes a coherent group of people doing roughly comparable work.

Two people inside that range might be delivering wildly different value. One writes code the traditional way. The other uses AI tools to do what two engineers used to do, has automated half the team’s deployment process, and spends freed-up hours on architecture decisions that used to belong to someone a level above. The range says they’re equivalent. Their output says they’re not.

A Head of People at a Slovakian startup told me something that has been rattling around in my head since I heard it. She said: “We posted our ranges because the law requires it. But internally, the ranges create more problems than they solve. Our strongest people see the ceiling and get demoralized. Our weakest people see the floor and feel safe. And I can’t explain to either group why the system works the way it does, because it was built for a different kind of workforce.”

She’d been considering output-based bonuses as a workaround. Pay the base inside the band, then layer variable comp on top tied to measurable results. I’ve heard this from a handful of other companies too. It sounds reasonable until you try to define “output” for most knowledge work.

What counts as output for an HR business partner? For a program manager? For legal counsel? Sales has quota. Engineering can maybe count shipped features, though even that’s debatable. Most roles don’t have a clean metric, and defining one poorly (which is what happens when you rush) creates perverse incentives that are worse than the original problem.

I don’t have a prediction for how this resolves. The current system is fracturing. Transparency requirements are making the cracks visible to everyone, which was the point, but the cracks are in the structure itself, not just in how it was being applied.

Something will replace this, or at least significantly modify it. But the replacement hasn’t arrived yet, and I’m skeptical of anyone who claims to have it figured out. Most of what I’ve heard about so far is the old model with a new label.

The question that keeps coming back is whether fair pay can even exist when the same title, in the same company, on the same team, now describes two fundamentally different jobs. I don’t think anyone has answered that yet. I’m not sure anyone has fully asked it.

The salary band was never meant to be eternal, but it was built for a world where human capability evolved on human timescales. In the age of AI, the real compensation challenge is no longer how to price a job; it is how to reward the rare combination of judgment, taste, and rapid adaptation that turns a tool into leverage, before the next wave of augmentation renders even that edge temporary.

In case you missed it:

I’m genuinely tired of the jobs debate. Not because it’s unimportant. Because it’s the wrong argument.

The thing AI is actually changing is more uncomfortable than job loss. It’s this: for a lot of people, the thing they’ve been paid to do, the core activity they call their job, doesn’t require human judgment. Never did. AI is just making that visible.

That’s the shift. Not headcount. Not automation. Not even productivity.

The shift is from doing work to owning decisions. And those are not the same thing at all.

The comfortable middle is already gone

Before getting into what changes, it helps to understand what’s already happened. Because the “AI is coming” framing is weirdly premature. The middle of the labor market has been hollowing out for two decades.

Economists David Autor, Lawrence Katz, and Alan Krueger documented the early version of this pattern in the late 1990s, showing how computerization started eliminating routine cognitive work while high-skill and low-skill roles stayed relatively intact.

David Autor extended that work significantly, and by the early 2010s the evidence for what he called “job polarization” was fairly well established: the roles most at risk weren’t the lowest-paid but the mid-wage, routine-heavy ones. Data entry clerks. Bookkeepers. Quality control auditors whose entire job could be described in ten rules.

That pattern didn’t stop. It accelerated.

What AI does is push the boundary upward. The “routine” category now includes things that used to sound complicated: drafting legal summaries, writing first-pass code, analyzing competitive positioning, building financial models, responding to customer escalations based on policy guidelines. These tasks look professional. They’re cognitively demanding in a surface sense. They require training, familiarity with jargon, sometimes years of practice to do quickly.

But they share something with the bookkeeper’s job: given enough examples and a clear success criteria, they can be done without anyone actually making a call.

That’s the tell. Not whether the task is hard. Whether it requires a call.

A junior analyst who spends sixty percent of their week pulling data, formatting it, and summarizing it for a director to interpret is not making calls. The director is making calls. The analyst is producing inputs. AI produces inputs faster, cheaper, and without burnout.

That’s not the uncomfortable part, though. It’s not that the analyst might lose their job. It’s that the analyst’s job was always the director’s job, outsourced downward for cost and convenience, and AI just made that obvious.

MIT economist David Autor noted in a 2022 NBER working paper that middle-skill job displacement has been ongoing since roughly the 1980s, and that what distinguishes the current wave is the speed of capability expansion into tasks previously considered non-automatable. His concern wasn’t apocalypse. It was the compression of the timeline for adaptation.

That compression is what makes this moment different. Not the destination. The speed. And the fact that this time, the capability expansion is hitting roles that have traditionally felt quite safe: analyst, associate, coordinator, specialist. The cushion between “I do knowledge work” and “my role is economically exposed” is thinner than it has ever been before.

What “knowledge work” actually meant

This is happening in nearly every org I know. A team gets built around specialization. You need someone who knows the analytics platform. You need someone who knows the legal compliance requirements. You need someone who knows the presentation layer of the client relationship. And you need someone who can write.

Each of those people becomes the organization’s memory for their slice. The system works. It’s inefficient, but it works. The knowledge lives in people, so the people have to stay.

There’s a mild irony in how much of what we called “knowledge work” was really “organizational memory work.” The knowledge wasn’t insight. It wasn’t judgment. It was familiarity. Knowing where things lived. Knowing which form to use. Knowing that Lucy in legal always takes three days to turn around a contract so you have to plan for that. Knowing the PowerPoint template is actually version 4.2 from February, not the one on the shared drive.

This is not a small amount of what knowledge workers do. In my experience it’s a large amount. Maybe most of it for mid-level roles.

Management consulting, interestingly, built an entire industry around this problem. When consultants come in and spend two weeks just “understanding the org,” what they’re doing is acquiring organizational memory fast enough to be useful. The firms that got very good at this, McKinsey, Bain, BCG, developed structured approaches: process maps, knowledge repositories, interview frameworks. They were not solving intellectual problems. They were solving organizational amnesia problems efficiently.

The consultants who survived long careers were never the ones who could acquire organizational knowledge fastest. They were the ones who could sit with a CFO who’d just gotten contradictory data from three departments and help her figure out what to actually do. Not what the data said. What to do.

AI dissolves the organizational memory problem. That capability, the thing most of those mid-level roles were actually protecting, evaporates.

What remains is the judgment layer. And we have not, as a society, trained people for that in any systematic way.

The difference between doing and deciding

A few weeks ago, I found a National Bureau of Economic Research paper written by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond that I’ve thought about more than almost anything else in this space. They studied the deployment of an AI assistant in a large software customer service operation and found that the workers who improved most were not the most experienced.

They were mid-experience workers who could now perform like experts because the AI was handling the retrieval and pattern-matching work. The experienced workers improved less, percentage-wise, because they already knew the patterns.

The finding I keep returning to is a follow-on implication: if AI lifts average performance toward expert performance in retrieval-heavy tasks, what distinguishes the expert now?

Not faster retrieval. Not more accurate pattern matching. Not a better memory for the organizational manual.

What distinguishes the expert is knowing when the manual is wrong.

That sounds simple. It is not simple. Knowing when a rule should be broken, when a client relationship is about to go somewhere no process was designed for, when a decision looks right on the numbers but feels wrong for reasons that require articulating, when you should escalate versus absorb a risk yourself, when to trust a good outcome that arrived via a bad process.

This is not knowledge. It’s judgment. And judgment comes from having been wrong in consequential ways enough times to have a calibrated sense of where you’re likely to be wrong again.

I’m not fully confident that judgment can be taught or that organizations will reward it properly even when they need it. When I was working in an agency, I’ve seen companies say they want people who “challenge the status quo” and then freeze out the first person who actually does. The gap between organizational rhetoric about judgment and organizational tolerance for the actual exercise of judgment is large. I don’t know how to close that gap. I’m not sure this piece closes it.

What I’m confident about is that it’s the right problem. The roles that will hold value are not the roles where humans beat AI at retrieval or speed. They’re the roles where the human is accountable for a call, where the output requires owning a position, where being wrong has a name and a face attached to it.

Responsibility is becoming the scarce resource. That’s new. Or at least it’s happening faster than it ever has before.

Teams built for specialization will break

Most organizations are still structured as if information is expensive to move.

The departmental model, marketing over here, analytics over there, product somewhere else, legal on a different floor or a different continent, exists because historically you couldn’t give everyone access to everything. So you hired specialists. Put them in groups. Built handoffs between the groups. The handoffs created friction. The friction created coordination roles. The coordination roles created more friction. You end up with an org chart that’s a map of information bottlenecks, not a description of how work actually flows.

AI removes a lot of the reason information is expensive to move.

A generalist with good tools can now perform work that would have required three specialists six years ago. Not in every domain. Not for the hardest problems. But for a meaningful percentage of the problems organizations actually face on a given week.

The structural response to this is not complicated in concept: smaller teams, flatter reporting, roles organized around outcomes rather than functions. You stop asking “what’s your job title” and start asking “what problem are you solving.”

In practice it’s very disruptive. Entire departments whose value proposition was managing the complexity of their own existence become hard to justify. Middle management layers that existed to translate between specialists and executives get squeezed from both sides. The executive can now consume a synthesized briefing from AI tools that would have taken three analysts two days to produce. The specialist can now self-deploy in ways that used to require a manager’s bandwidth.

At this moment, I’m uncertain how fast this actually happens at scale. Organizations resist structural change with more force than they resist almost any other kind of change. The people who benefit from the current structure, and there are many of them, do not make transitions easy.

Regulation may slow adoption in some sectors. Some industries, healthcare, financial services, legal, are subject to oversight frameworks that add friction to every deployment of automated tooling.

The 2-to-4 year horizon is probably right for the early-adopter companies, the ones already organized for speed. For large traditional enterprises? Maybe 5 years, maybe a decade, maybe longer. Who knows. Structural change at a Fortune 500 is not a product sprint.

But the direction is clear. And people who have built their careers entirely inside a specialization, who’ve never been asked to own an outcome from beginning to end, who’ve always been insulated by role definition from having to make a call that wasn’t covered in the job description, a responsibility that wasn’t in their job description, are in a really tough spot.

Not because AI is taking their job. Because the organizational logic that required that job in the first place is being dismantled.

This is not going to feel like progress

I want to end here without the standard move, which is to pivot to “but here’s how you adapt.” I’ll say a few things that are actually useful. But not before saying something honest.

This is going to be hard for a lot of people. Not in a temporary disruption before everything gets better way. In a the path you were on is gone and the new one requires something you haven’t been asked to develop way.

Maybe there won't be a “jobs apocalypse” due to AI, but there will be job chaos, as Gartner predicted.

But workers most at risk are not the people who think they’re most at risk. The junior analyst who’s worried about AI taking their entry-level job is right to be concerned about the market.

But the mid-level manager who assumes their judgment is safe because they’ve been “doing strategy” for ten years should be equally worried.

Especially if the strategy work mostly involved collecting inputs from reports, formatting them into a deck, and presenting a consensus recommendation. That’s not strategy. That’s a well-dressed coordination function.

The people who will come through this in a strong position share a few traits, and none of them are credentials. They can hold ambiguity without freezing. They take positions even when the data is imperfect. They’ve been wrong in ways that were costly and they know it and they adjusted because of it, not despite it. They’re accountable for outcomes in a way that means someone calls them when something fails.

Junior roles are almost certainly shrinking in the near term, the entry-level slowdown among workers in their early twenties is showing up in hiring data pretty clearly. But whether those roles disappear or just mutate into something nobody has named yet, I genuinely don’t know.

The historical record says new role categories emerge. What those categories require and whether the people being frozen out now can access them later is a different and harder question. I keep reading optimistic takes about it and I keep not being fully convinced either way.

Recent McKinsey Global Surveys on AI adoption have highlighted that most large organizations still feel their workforce is not fully prepared for AI-augmented work roles, with many continuing to focus on initial experimentation and training rather than deep integration.

Which is a polite way of saying that eighty-five percent of companies know they have a problem and most of them are still doing workshops about it.

The honest advice is not complicated to say. Be the person who makes the call. Get into positions where you’re accountable for outcomes, not just contributory to them. Learn to be wrong in front of people and recover from it fast, because that’s what building judgment actually looks like.

But I should acknowledge that not everyone has access to those positions.

Not everyone has been given the runway to develop judgment because the orgs they’ve been in didn’t trust junior people with anything real.

And not everyone has the safety net to take career risks. The people who’ll thrive under this shift are disproportionately the people who already had structural advantages.

The shift is real. The direction is clear. But I keep coming back to something that the frameworks and the forecasts don’t really address: the people most exposed to this transition are often the ones who did exactly what they were told. They built expertise inside systems that rewarded specialization. They were reliable. They showed up.

The fact that the system is now changing its mind about what it values is not a personal failure on their part. It’s a structural one.

I don't have a satisfying answer for those people. What I'd offer instead: the discomfort of not knowing where you stand is more honest than the confidence of someone who thinks they do. Start there. It's not nothing.

The three-question test for whether your role survives the next five years

Most of the “is your job AI-proof” frameworks I’ve seen are reassurance dressed up as analysis. They ask things like “do you work with people?” or “do you use creativity?” and then confirm that yes, you’re probably fine. That’s not useful. Plenty of roles involving people and creativity are going to be restructured significantly.

Here’s a more useful frame. Three questions. Answer them about your actual day last Thursday, not about your job description.

Read more

That’s it. No paid tool. No special subscription. No developer skills required. Claude can open job portals, read listings, fill in applications, write tailored cover letters, and hit submit while you’re asleep.

The application used to cost something. Time and mostly the low-grade anxiety of a blank cover letter field, the small act of deciding whether you actually wanted a job before sending your resume into the void. That friction was annoying, and it was also doing real filtering work. It separated people with genuine interest from people who figured “why not.”

That filter is gone. And what’s replaced it isn’t better screening. It’s volume, without the selection that volume was supposed to create.

Sit with that for a second. Not the version of this story where AI is coming for jobs someday in the future. The version where, right now, the act of applying has already been devalued to the point where it says almost nothing about your interest in a role. It says nothing about your fit. It says you had a browser open and knew how to type a prompt.

I ran a coaching session last weekend with a group of job seekers. Several of them were running auto-apply tools around the clock, targeting any remote or hybrid role globally, rotating IP addresses through VPNs to avoid getting flagged as bots.

One person in that group sent over 2,000 applications last month. His thinking was simple: it’s a numbers game, and somewhere in those 2,000 there had to be one company that would write back. Nobody did.

Not because his resume was weak. Not because 2,000 wasn’t enough. Because he needed a visa and relocation. He was optimizing volume on a problem that volume couldn’t solve. The actual blocker was never the application.

This one question that kept coming up on applications: “Do you require visa sponsorship?” They answered no, even when they did need it. Their reasoning was straightforward: they’d figure out the visa themselves, so technically it wasn’t the company’s problem. But it is a company problem and this only shows why some knockout questions stopped working.

The cost of applying just hit zero

For most of its hiring history, applications had a built-in filter. Filling out a form took several minutes. Writing a cover letter took an hour. Customizing a resume meant you actually had to read the job description. The effort was a weak signal, not a strong one, but it existed. If you applied somewhere, you had at least made a decision to apply.

LinkedIn’s Easy Apply changed some of that around 2019. One click, profile attached, done. Application volumes started climbing. Recruiters started complaining. But even Easy Apply required a human hand to click. Some intention survived.

AI removed intention from the equation.

When you can instruct a model to apply on your behalf across dozens of roles with no input beyond your resume and a general preference, the act of applying stops meaning anything at all. I don’t say this to be judgmental about the people doing it. The job market is brutal right now, and if a system can be gamed at zero cost, people will game it. What I’m pointing at is what happens to any system when its cost drops to zero: it gets flooded. The signal collapses.

A talent director at a mid-sized tech company told me in January that they received over 2,400 applications for a senior data analyst role in eight days. She asked me not to name the company. Their previous record for that position type was around 100. Same job boards. Same description. The volume multiplied and kept going.

Her team filtered it to 38 candidates using keyword screening. Of those 38, she guessed maybe a dozen had actually read the job description. One applicant, based abroad with no work authorization for the role, had applied to the same reposted listing three weeks in a row. The cover letter confidently named the wrong company.

That’s not an edge case. That’s a representative Tuesday.

Both sides automated

Companies noticed the flood. So they responded with more AI: automated screening tools, knockout questions, behavioral assessments that happen before a human ever sees a name on a resume. Some built custom filters. Some bought third-party products claiming to detect AI-written applications (terrible idea if you ask me).

Some companies are now using AI to screen candidates and then posting on LinkedIn about how fair their hiring process is. What they don’t realize is that if their AI filter is flawed, they’re just rejecting good candidates faster. Their consistency is not fairness. If the AI consistently penalizes career gaps or non-Western universities, they’ve scaled that bias across every candidate in the funnel. But it’s a great message for candidates about where not to bother applying. Nothing says more about a company’s culture than a LinkedIn post telling everyone they treat candidates like numbers.

This AI vs AI has led to candidates optimizing their resumes for AI, and job descriptions being written for algorithms instead of people. And then we all wonder why the quality of hires isn’t getting better. We’re not fixing hiring. We’re automating the illusion of fixing it.

The signals that used to carry real information have degraded all at once. A well-written cover letter used to suggest care. Following even strange application instructions used to filter out the careless. No typos used to be a small positive mark. All of that can now be produced in 30 seconds. I genuinely don’t know what replaces those signals. I don’t think the field has worked that out yet.

The knockout question trend is a real partial response. Asking applicants to submit a short video or describe a piece of relevant recent work introduces friction that mass applications usually can’t handle. It works, for now. But it’s not a complete answer, and it puts a real burden on candidates who are camera-shy or who simply can’t afford to record ten videos only to never hear back.

And there’s a whole separate conversation about how poorly most applicant tracking systems are configured. That dysfunction predates AI by a decade and is genuinely its own mess.

What referrals actually cost the person giving them

That 70 to 80 percent figure, the one claiming most positions are filled through personal connections, gets cited constantly. I’ve never found a study behind it that I actually trust. It circulates because it’s useful to people selling networking advice, not because anyone audited it carefully. But the underlying reality it points at is real: referrals work, and they work for a specific reason most people skip past.

When you refer someone, you’re not forwarding a resume. You’re attaching your own judgment to their candidacy in a room where your judgment already has standing. If they underperform, you’re the one who vouched for them. If they turn out to be difficult to work with, the people who trusted your read on them will remember that. A referral is a bet, and the person making it has real skin in the game. That’s exactly what makes the signal credible: it’s costly to give, which means it tends to only get given when someone genuinely believes in the candidate.

When the cost of applying drops to zero, application signals collapse. The edge migrates to whoever controls higher-cost signals. The person willing to show up at a hiring event in person. The person who spent a year in a professional community before they needed anything from it. The person whose former colleague sent an unsolicited message to a hiring manager saying “I worked with her for two years and she’s the best person I’ve ever managed.”

I still tell people to optimize their LinkedIn profiles and clean up their resumes, because the resume matters once you’re in a conversation. It rarely gets you there anymore, not through a portal.

What actually helps, and I’ve seen this work: treating a job search like a research project about people rather than a process for submitting documents. Pick a few companies you’d genuinely want to work for. Find out who actually runs the team you’d join, read what they’ve published or posted publicly, and show up at one event where they might be present.

Then send one message, not a template, something that shows you engaged with something they actually put out. That’s a week of real work. It produces more signal than hundreds of applications ever will. The cost is time and real discomfort. Most people won’t do it, which is part of why it still works.

Good recruiters got more valuable right when everyone expected them to vanish

For three years I’ve been watching panels about AI replacing recruiters. Screening automation, scheduling bots, AI-generated job descriptions, skills-matching at scale. The conclusion was usually some version of “recruiters need to adapt or they’ll be irrelevant.”

What’s actually happening is the opposite.

When you have 1,400 applications and an AI that’s nominally ranked them, you still need someone who can look at the top 60 and notice that candidate 52 has a non-linear path making her more interesting than everyone above her on the list. You still need someone who can call a passive candidate who wasn’t looking. You still need someone with enough standing to push back on a hiring manager and say “the job description you wrote is filtering out the people you actually want.” Algorithms return the candidates who most closely resemble what you already thought you wanted, which is frequently not what you need.

The administrative work of recruiting is getting automated. The judgment part, which was always the harder part, is getting more important, precisely because the noise is so loud that you need someone who can hear through it.

The recruiters who’ve adapted are building communities rather than pipelines. They’re showing up in professional Slack groups and Discord servers where the talent they want actually spends time. They’re posting things worth reading, not just job listings, becoming the kind of person a candidate thinks of first when they quietly decide they’re open to something new. That’s a meaningfully different job than reviewing applications after the fact, and most companies haven’t restructured around it yet.

It’s bimodal right now. Some teams are doubling down on inbound and AI screening stacks. Others have basically written off portals for certain roles and told their recruiters to go find people directly. The second group isn’t making a mistake, and I expect more companies will follow them.

What I’m skeptical about is the AI spam-detection tools starting to appear, the ones claiming to identify AI-generated applications and filter them out. They can’t reliably distinguish between a mass-submission from someone who doesn’t know the company exists and a careful application that was polished with AI. A genuinely qualified candidate who used AI to tighten their resume language gets flagged the same way as the noise. Nobody gets a rejection explaining why their AI-polished resume looked the same as the spam. Companies that invest most heavily in that detection will filter out real candidates to stop fake ones, and won’t know they’re doing it. The more you invest in those tools, the worse your candidate pool probably gets.

Where I think this is heading

I wrote a piece (The Job Search Is Broken. Here’s What Could Fix) about two years ago arguing that every company will eventually have an agent(s), and every person will have an agent(s), and those two agents will find each other. Not through a job posting. Through some form of structured, ongoing matching where your agent holds your actual track record, your preferences, your reputation in the market, and the company’s agent knows what the team genuinely needs, not what a job description written by committee says. They negotiate a first conversation before either human has committed to anything.

I still believe that’s coming. There are technical and regulatory problems nobody has fully worked out yet. What data does your agent hold, and who controls it? The matching layer sounds clean in the abstract and will be messy in practice. Whether it reproduces the same bias problems in new forms is something I think about more than I usually say out loud.

In the shorter term: expect more proof-of-work before any human reviews your application. Take-home tasks, short video submissions, project samples. Expect AI sourcing tools to treat your public LinkedIn posts and your published writing as your actual application, whether you filed one or not. Your GitHub matters too, though with the rise of AI-generated code, everyone’s repository is starting to look like a developer’s, which makes that signal harder to read. And expect in-person hiring events, things that felt like relics two years ago, to start mattering again in ways most people aren’t quite prepared for.

My instinct is that the retreat from inbound creates pressure toward local and in-person hiring, because those contexts are where higher-cost signals still function. A conversation at an event, a presentation someone saw, a coffee meeting that eventually turns into a referral. That probably cuts against remote work at the margin, and I don’t think that conversation has fully caught up with what’s happening in hiring. I may be misreading the direction. I don’t have clean data on it.

What I keep returning to is that anyone currently treating portal applications as their primary job search method is playing a game where the rules changed, and the new rules favor people with warm networks. That’s not advice I love giving. Warm networks aren’t equally distributed. Not everyone has spent years building LinkedIn connections or has a colleague willing to put in a good word. The thing that works best right now is the thing that’s hardest to build quickly, and I haven’t found a satisfying answer to that tension.

What I keep coming back to, though, is something someone told me very early in my career. Do your job well. Treat everyone with respect. Help others when you can. Those are your real credentials for whatever comes next. I’ve watched the tools change, the channels change, the whole architecture of how people find work change, and that advice has somehow aged better than almost anything else I was given.

Your reputation in the actual rooms where you’ve worked, among the specific people who’ve seen you handle something hard, is the one signal that hasn’t been devalued. It costs a decade to build and it can’t be auto-filled in a browser tab.

That person was right.

In case you missed it:

He was still watching three or four episodes of something before midnight. He’d sent out six applications in eight weeks and told himself he was being strategic about it. He wasn’t sleeping well, and there was a half-finished cover letter sitting open in a browser tab he kept minimizing.

The wake-up call framing had done exactly nothing except give him a tidy story to carry around.

I’ve been thinking about Marek’s situation a lot lately. He’s a composite of maybe five or six people I’ve observed closely enough to have real information, not one specific person. His situation captures something real about how low points actually work, which is usually not the way we describe them. We talk about burnout, breakups, and rejection as if the bad event is the agent of change. The thing happens, and then transformation follows, the way a match lights a fuse.

That’s not what happens.

What actually happens is that something breaks. There’s pain, sometimes genuine grief. And then there’s a window. Not a long one. It opens in the space between when the worst of the acute distress has lifted slightly and when your old patterns have fully rebuilt themselves.

During that window, you have more genuine flexibility than usual. The familiar routines, the ones that helped create the conditions for burnout or kept you in the wrong situation, are momentarily offline.

Most people spend that window waiting for the pain to motivate them.

It doesn’t. Pain is a condition. It’s not a force that propels you toward a better life, and treating it like one is how people miss the only period when change is actually available to them.

The Mythology of Rock Bottom

There’s a persistent belief, especially in self-improvement writing, that rock bottom is reliable. Hit it hard enough and a rebound is guaranteed. The logic is almost physical: compress something far enough, it springs back.

This belief does something specific. It encourages passivity during the exact period when action matters most. If the low point is the catalyst, then you don’t need to do anything except endure it long enough. The change will come.

Richard Tedeschi and Lawrence Calhoun at UNC Charlotte spent years studying what they called post-traumatic growth, the genuine positive change that some people report following serious crisis. Things like a deeper sense of meaning, stronger relationships, a clearer sense of what they actually want. Their work, beginning in the mid-1990s, found real evidence that crisis can produce this. The concept is legitimate and well-documented.

But here’s what their research also showed, and what almost nobody who cites it bothers to mention: post-traumatic growth and post-traumatic stress are not opposites. The people who grew were often also the people still carrying real damage. Growth didn’t replace suffering. It coexisted with it.

And a significant portion of people who went through comparable crises simply returned to baseline. They recovered. They didn’t grow. The crisis happened, they endured it, and then they reconstituted more or less the same life they had before, same patterns intact, same unexamined assumptions in place.

Recovery is not transformation. It just requires less effort.

There’s also something I genuinely don’t have a satisfying answer for. The relationship between the size of the blow and the size of the response is wildly inconsistent. Some people seem to reach their window from relatively minor setbacks, a mid-level rejection, a mediocre annual review, while others survive catastrophic losses and emerge more or less unchanged.

I’ve watched this happen enough times that I stopped expecting a pattern. Anyone who claims there’s a clean explanation for it is probably working backward from the cases that confirmed their theory.

Pain is Necessary But Not Sufficient

The standard script for how low points produce change runs like this: bad thing happens, person hits bottom, discomfort becomes unbearable, person changes to escape the discomfort. This is a behavioral model. It treats humans as things that will eventually move away from whatever is burning them.

Sometimes that’s true. But it gets the mechanism wrong.

Barbara Fredrickson at UNC Chapel Hill has spent decades researching positive emotions and their effect on cognition. Her broaden-and-build theory, published in a 2001 paper in American Psychologist, argues that positive emotional states expand what she calls your thought-action repertoire.

They widen the range of behaviors you can actually see as possible options. Negative states narrow that range. Which has real survival value in acute emergencies. Terrible value when you’re trying to redesign your professional life or your relationship patterns.

What this means in practice: when you’re in serious pain after a major setback, the mental conditions required to imagine a genuinely different future are partially unavailable. You’re in tunnel mode. You can see the next safe step, maybe two. Wholesale reconstruction of your identity is not something you can do from inside acute grief. That’s not a character failing. It’s how stressed cognition works.

So the window matters not because the pain eventually disappears, but because there’s a brief period, usually a few weeks to a few months, when the worst of the acute distress has lifted slightly but the old habitual patterns haven’t re-established themselves. Your previous routines are momentarily offline. Slightly more flexibility than usual. Decisions made during this period tend to stick.

The people who use low points well tend to act inside that window before the numbness fully arrives. Not on the first terrible day, but before they’ve recovered enough to retroactively decide the thing wasn’t that bad, or concluded that a minor adjustment is all that’s required.

I’ve watched a few people miss this entirely by spending the window processing. Daily journaling, long conversations with friends about what happened, therapy, which is genuinely worth doing regardless.

The processing is necessary. But processing and deciding are not the same event. Someone can spend three months in therapy working through why they burned out and still, at the end of it, take the same kind of job at a different company under slightly different conditions.

This is also around the time when “processing” became a culturally dominant frame for responding to hard things, somewhere in the last fifteen years, and I sometimes wonder whether that framing gives people permission to stay in reflection mode longer than is actually useful. I don’t have data on this. It’s just a thing I notice.

What the Growth Research Actually Showed

Post-traumatic growth, in the research, is not spontaneous. Tedeschi and Calhoun describe a process involving what they call deliberate rumination (the voluntary, active process of thinking about a stressful event or personal challenge to understand it, find meaning, and foster growth.)

The person actively works to make sense of the event, tests their previous beliefs against what happened, and revises their understanding of themselves and the world.

The operational version of that is more specific than it sounds. It requires that the person actually allow the crisis to challenge their prior beliefs. Not just feel bad about the outcome, but genuinely question whether the framework they were operating in was wrong.

That’s much harder than it sounds.

Klára gets laid off from a company she poured herself into for four years. The diagnosis she lands on: the company was poorly managed, the leadership made bad decisions, the market shifted. All of which might be true.

What she doesn’t examine, because it’s more uncomfortable, is why she stayed that long when she’d known something was wrong, or what that intensity of investment was compensating for, or what she’d been avoiding thinking about by staying that busy.

The situation-level diagnosis leads to a situation-level fix: find a better company, vet the culture more carefully next time. The system-level diagnosis leads somewhere harder and more interesting.

I’m not saying everyone who experiences a low point is responsible for engineering it, and I’m not saying looking inward is always the right move. Sometimes a job just ends badly and the healthiest thing is to process it and move forward without extensive excavation. I’ve seen people dig and dig and find nothing useful, just old discomfort that doesn’t connect to anything actionable. There’s no universal prescription.

But for the people for whom a low point does produce real change, what I consistently observe is not that the pain motivated them. It’s that they questioned something they hadn’t previously been willing to question.

The Window Closes Faster Than People Expect

Most people come out of a serious low point one of two ways. They either change more than they expected to, or they recover more fully than they expected to.

Recovery is the default. The brain is genuinely good at it. Wilson and Gilbert's research on affective forecasting, developed across a series of studies through the 1990s and 2000s, documented what they call the impact bias: people consistently overestimate both the intensity and the duration of their emotional reactions to negative events.

The part people miss is duration. You predict you'll feel bad for longer than you actually do. You adapt back toward baseline faster than you expect, and the adaptation is partly driven by what they call immune neglect, your failure to anticipate how quickly your psychological coping mechanisms will kick in.

That adaptation mechanism is useful in many contexts. It’s why you’re not still devastated by every professional rejection you’ve ever received. But it also means the window for serious change closes faster than people typically assume.

Six months after serious burnout, many people have rebuilt enough psychological comfort that the urgency is gone. The conditions that produced the burnout are quietly reassembling themselves. The unexamined pattern that kept them in the wrong relationship has started operating again under a different name.

If you’ve been through something hard recently and you’re waiting to feel stable before making real decisions, you may be waiting until the window has already mostly closed.

What to Do With the Wreckage Before It Rebuilds Itself

I want to be careful not to make this sound simpler than it is.

The practical advice in this category clusters around two poles: “lean into the discomfort” (which usually means, vaguely, feel your feelings) and “take action immediately” (how people make impulsive post-breakup decisions they spend years correcting). Neither is right.

What I’ve seen work, in myself and in people I’ve watched closely enough to have real information, involves something more specific.

First: write down what you were pretending wasn’t true before the low point. Not what the bad thing was. What you already knew was wrong that you weren’t looking at directly. This is uncomfortable enough that most people skip it or do a lite version where they acknowledge something vague, “I knew I wasn’t happy,” rather than something specific, “I knew I was staying because leaving felt too complicated and I was afraid I couldn’t do better.” The lite version produces lite decisions.

Second, and this one is harder: once you have that list, make at least one concrete change that addresses something on it before you feel ready to. Not while you’re in the worst of the acute pain, but during the window, while you’re still unsettled enough that your normal patterns haven’t re-solidified. This is where most people stop. They’ve identified the thing. They’ve told a few people they’re going to change it. Then they wait for the right moment, which turns out to be a moment they never quite arrive at.

I did this badly after a period of serious burnout in my life. I spent about four months in careful reflection mode: reading, having long conversations with people I respected, filling notebooks. I identified something real about how I was working and why it wasn’t sustainable. Then I made an extremely hedged version of a change that preserved most of what I’d been doing before. I called it a transition.

It was basically the same life with a slightly different description. About eighteen months later I was back in the same position, slightly more tired, slightly less optimistic about my ability to fix it.

What I didn’t do was make a decision that had actual cost to it. Not self-punishment, not dramatic disruption, just something that required giving up something I was comfortable with. I kept looking for the version of change that didn’t require a real loss.

There isn’t one.

A third thing, and take this one with some skepticism: the people who follow through tend to have told at least one specific person what they’re planning to do. Not announced it broadly. Told one person whose opinion of them matters, someone who will actually notice if they don’t do it. The mechanism is not accountability in the motivational-poster sense. It’s simpler and less flattering than that. When someone you respect knows what you said, the path of least resistance shifts slightly. Quietly abandoning the thing costs more than it did before you said it out loud.

I’ve seen people make real changes with no witnesses at all. So I don’t want to oversell this. But for the people I’ve watched struggle to act during the window, the ones who did tend to have said something specific to someone specific, rather than just held it privately.

The Specific Question That Separates Processing From Deciding

Most people who make this list and sit with it experience one of two things.

Read more

I wasn’t stupid about it. I had a reason: if I could automate the twelve-minute task, I’d have twelve extra minutes every time it came up. Multiply that across a week, a month, a quarter, and suddenly you’re recovering real time. The math checked out, and the motivation was legitimate. By 6 pm, I had a workflow diagram, a webhook that kept timing out, and the original task still sitting in my queue because I’d been too busy building the system to actually do the work.

My wife asked what I’d gotten done. I showed her the diagram.

She nodded politely and said nothing. That doesn’t surprise me. I get really excited about the diagrams and things I create, like EditText.App, but I’m often the only one among my family and close friends who feels that way.

While I was looking for ways to improve my AI agent’s sourcing abilities, I found a study published in Harvard Business Review that suggests that rather than making work easier, AI may be causing what researchers term “brain fry.”

That got me thinking about how I was spending my time with the latest tech stack, you know, all those AI agents like OpenClaw and others. Instead of letting the machines do the work, I was spending a lot of time babysitting them, making sure they didn’t hallucinate, misunderstand instructions, or just make a bigger mess than necessary.

Over the past few months, I’ve tested at least 15 AI agents. Different tools, different use cases, some designed for research and summarization, some for content generation, some that were supposed to handle the kind of low-level decision-making that eats up thirty minutes a day in email triage.

I went in as a believer. I still am, in some long-run sense. But somewhere around agent number eight I started to notice that the time I was “saving” kept getting absorbed by something I couldn’t quite name.

The workflows multiplied. The actual client work didn’t change. My output, measured by things with deadlines attached to them, stayed roughly flat while my Notion workspace got increasingly elaborate. I kept optimizing the system that was supposed to free me up to do the work, instead of doing the work.

I’m not sure when I noticed the gap. Probably sometime around the time, when I realized I’d spent an entire Sunday morning configuring an agent to help me write faster content for websites and produced nothing writable by noon.

The screenshot is already the product

There’s a psychological mechanism called effort substitution, and it’s not a fringe idea. Research published in the Journal of Consumer Psychology in 2011 by Moty Amar and colleagues found that people who planned to accomplish a goal reported meaningfully lower motivation to actually pursue it afterward. The planning itself discharged some of the psychological pressure. You got partial credit from your own brain without doing the thing.

AI tools didn’t create this problem. But they gave it a much nicer interface.

Before, performing productivity looked like: color-coded calendars, elaborate to-do apps, journaling about your goals in a Moleskine. Those things could at least pretend to be preparatory.

The AI version looks like: a multi-agent pipeline that summarizes your inbox, drafts your LinkedIn posts, and schedules follow-ups based on email sentiment analysis. It’s harder to dismiss as mere performance because it involves real technology that does real things. The tool is actually running. Code is actually executing. You’re not just drawing boxes in a notebook; you’re watching logs scroll by in a terminal.

Which is exactly what makes it more dangerous as a substitution mechanism, not less.

LinkedIn rewards the aesthetic of process. I’ve seen screenshots of workflows that have to take two hours to build, but those people received three times their usual engagement of posts describing actual results they spent three weeks producing.

The platform isn’t optimizing for you sharing what you accomplished. It’s optimizing for the thing that makes other people feel like they’re behind. A complicated-looking agent diagram does that. A business outcome usually doesn’t photograph well.

There’s a whole conversation about why LinkedIn specifically amplifies this that I’m not going to get into, but I’ll say this: it’s not just LinkedIn. Reddit productivity threads, Hacker News threads about someone’s personal OS, the YouTube genre of “my complete 2026 productivity system” - they all reward the same thing. They all reward the showing of the work, not the work itself. AI just made the showing more impressive-looking than anything we had before.

And here’s something I’ve been sitting with and can’t fully resolve. I wonder how much of this predates AI entirely. The productivity genre has been selling systems over outcomes for decades. Getting Things Done came out in 2001 and spawned an entire sub-economy of people who spent more time refining their GTD implementation than they spent doing the tasks the system was supposed to capture.

Maybe the AI agent trend is just the latest iteration of that. Or maybe the speed and sophistication of current tools have crossed some threshold where the substitution effect is qualitatively different. I genuinely don’t know. But I notice the problem more now than I did five years ago, but I can’t tell if that’s because the problem is bigger or because I’m paying more attention.

What effort substitution actually does to a brain mid-task

The Jean Baudrillard reference in a lot of online productivity discourse is a bit overstated, but it’s pointing at something real. A simulacrum, in Baudrillard’s framework, is a representation that has replaced the thing it was supposed to represent. The map becomes the territory. In this context: the workflow system becomes the work.

Your brain doesn’t experience this as deception. That’s important. You feel productive because you are doing something. The agent is configured. The prompt is engineered. The integration is live. These are real actions that required real effort.

The prefrontal cortex doesn’t have a clean way to distinguish between “effort that produces an outcome” and “effort that produces the appearance of an outcome,” especially when the tool in question generates convincing output.

A 2014 study from the University of Michigan by Ethan Kross found that self-distancing, the act of reflecting on yourself in the third person, reduced emotional intensity and helped people analyze difficult decisions more clearly.

I think about this whenever I watch myself evaluate my own workflows. I can’t fully self-assess here. I don’t have a good read on whether my agent stack is saving me time in any net sense, because the act of building and maintaining it feels like legitimate work even when it isn’t producing legitimate results.

The specific failure mode I see most: I’ll set up an agent to summarize emails I should be reading directly. Not because I get too many emails to read, but because reading emails I might need to respond to creates a kind of low-level anxiety that the summarization layer delays. The agent isn’t saving me time. It’s managing my discomfort. There’s a difference, and I kept pretending there wasn’t.

A friend of mine, Tomáš, who runs operations for a mid-sized logistics firm in Brno, told me something similar last spring. We were at a dinner that had nothing to do with any of this, talking about something else entirely, and it came up sideways. He’d spent six weeks building a dashboard that aggregated all his key metrics in one place.

Beautifully designed thing, real-time data, color-coded thresholds. Then he told me he mostly didn’t look at it. He was getting updates from his team verbally instead. He’d built the dashboard to feel in control, and once he felt in control he didn’t need the dashboard. I don’t know if this proves anything. His situation is unusual because he had the resources to build it properly, which most people don’t. But the dynamic stuck with me.

Babysitting the thing that was supposed to free you

This is what actually happens when you run agents at the level most people are describing online: They break. Not dramatically, not in ways that are easy to diagnose. They drift. A workflow that ran correctly for three weeks takes a random path on a Wednesday morning and produces output that’s 80% right, which is somehow worse than output that’s obviously wrong because 80% right takes longer to catch and fix.

I’ve had a research agent start hallucinating citations that looked plausible enough that I almost posted them. I caught it because I happened to read closely that day. And let's face it, we don't always read things closely, especially on days we're tired.

One of the agents I tested for about six weeks was built for lead research. The idea was simple: give it a list of company names, it pulls relevant context from a few sources, formats a brief for each one. First two weeks, worked great.

Then it started pulling information from the wrong companies, confusing similarly-named entities, and returning confidently formatted briefs that were factually backwards. I didn’t notice for four days. In those four days, I had two calls where I walked in with the wrong prep. The agent cost me more credibility than the time it saved.

I want to be fair here; that might have been a prompt engineering problem. Maybe better guardrails would have caught the entity confusion earlier. But that’s sort of my point. The tool requires ongoing investment just to maintain basic reliability, and the investment isn’t visible on the dashboard where you’re tracking your “hours saved.”

So you watch, you review outputs. You check that the tool did what you told it to do. And the watching and reviewing takes time, except it’s worse than doing the task yourself would have been because now you’re doing two things: the oversight work and the mental load of trusting a system you’re not sure you can trust.

I want to push back on something I see constantly online. The people claiming they have hundreds of agents running autonomously and generating significant revenue without meaningful supervision are mostly either lying about the supervision part or working in extremely narrow, well-defined domains where the failure modes are small and bounded.

Autonomous content generation at scale without human review produces confidently wrong material. Autonomous outreach at scale without human oversight produces messages that annoy people on your behalf without your knowledge. The “agentic future” some people are describing isn’t here. The models aren’t reliable enough for the level of autonomy being claimed.

A 2025 report from Stanford's Human-Centered AI group found that AI coding agents complete multi-hour task benchmarks with roughly 50% reliability. More striking: a randomized trial included in the same report found that experienced open-source developers took about 19% longer to complete work when they had access to frontier AI tools than when those tools were taken away. Not 19% faster, longer.

I should say: I’m not sure the coding context maps cleanly to the kind of business workflow automation most people are actually running. Coding tasks are more measurable than most agent use cases, which makes them easier to benchmark but also a bit cleaner than real-world conditions.

The 50% reliability figure has stayed in my head anyway, because it matches what I’ve experienced, and the developer slowdown finding is harder to dismiss.

The overhead compounds in ways that are hard to see in the moment. You spend forty-five minutes configuring the agent. Another thirty debugging when it fails the first time. Twenty minutes a week on monitoring. An hour here and there on prompt adjustments as the underlying model updates change behavior slightly.

Annualized, you may have spent more time on the infrastructure than you’d have spent doing the task the old way. The time savings are real in theory and often negative in practice, and the negative in practice part is invisible because it shows up as “agent maintenance” rather than “time I could have spent working.”

I’m not saying don’t use them. I’m saying the math you’re doing in your head when you adopt one is probably wrong.

What the twelve-minute task was actually about

The question I find myself sitting with isn’t “which agents are worth it.” That’s solvable with enough testing. The question is harder: if I stripped out every tool that makes me feel productive and only kept the ones that produce verifiable outcomes, what would be left?

I’ve been afraid to actually run that audit. That’s data.

There’s no crisp takeaway here about which workflows to cut or how to structure your agent oversight process. Plenty of people will sell you that. What I keep coming back to is the older, less satisfying question that the AI productivity conversation keeps stepping around: what are you actually trying to avoid doing, and is the thing you’re building making it easier or harder to avoid doing it?

For me it was the twelve-minute task on that Sunday in February. Not because twelve minutes was too long, but because doing it meant looking at a set of numbers I didn’t want to look at. The automation wasn’t about efficiency. It was a very elaborate way of not opening a spreadsheet.

There’s a version of this that’s genuinely hard to answer. Some tasks really do deserve automation. Some friction is real friction, not psychological avoidance dressed up as friction. I don’t have a reliable method for telling the difference in the moment.

What I notice is that when I’m excited about building a workflow, that excitement is sometimes about the problem being solved and sometimes about not having to confront what happens after it’s solved. The second kind of excitement has a slightly different texture, a little more restless, a little quicker to check how the tool looks rather than whether it’s working. I haven’t figured out what to do with that observation beyond noticing it.

The models will get more reliable. The tooling will catch up. Some of the oversight burden will come down as the agents get better at knowing when to ask for help versus when to proceed. I’m genuinely optimistic about the three-to-five-year picture.

But right now, most of what’s being called “autonomous” is just latent human supervision wearing a different costume, and most of what’s being called “productivity gains” is the pleasant feeling of configuring something that might, eventually, save you time.

The twelve-minute task is still in my queue, by the way. I did it last Saturday. Took eleven minutes. The numbers were fine.

Do not believe the hype; it is full of bugs.

The agent audit I ran after wasting six months on automation

The exercise I’m about to describe took me about three hours the first time and produced results I didn’t show anyone for two weeks because they were, lets say… embarrassing.

Read more

The steam engine came. Factory work replaced agricultural work. Electricity arrived and reorganized entire industries overnight. Computers showed up in the 1980s and everybody panicked, then the economy somehow absorbed them.

The internet killed travel agents and video rental clerks and classified ad salespeople, but it created software engineers and logistics specialists and about forty job categories that didn’t exist in 1995. The pattern is real. It’s even reassuring, if you look at it from the right angle, at the right distance.

But ask yourself who gets to tell that story. Not the handloom weavers in early 19th-century England whose incomes collapsed over two decades while factories gradually came online. Not the generation of workers who reached middle age in exactly the wrong moment, too old to retrain and too young to retire. Not the mill towns that hollowed out in the 1970s and 1980s when manufacturing automation and offshoring hit simultaneously, and which in some cases still haven’t recovered.

History, written from a distance, looks like adaptation. Lived from inside, it sometimes looks like getting left behind while someone else benefits.

I’ve been following the AI and employment debates for a few years now, reading the research, paying attention to what’s actually happening in different industries, and I find myself in a genuinely uncomfortable place. I think the historical reassurance is real and I think it might not be enough. Those two things can coexist.

The Luddites get invoked a lot in these debates, usually as a cautionary tale about people who were wrong. What gets left out is that the Luddites weren’t wrong about what was happening to them.

They were skilled textile workers whose specific expertise was being devalued faster than they could adapt. Their assessment of their immediate situation was accurate. They just couldn’t see the industrial economy that would eventually absorb their descendants into different kinds of work.

The fact that things worked out over a century doesn’t mean their experience wasn’t real.

The pattern that’s supposed to reassure you

Technology has always displaced jobs; new jobs always appear, and humans always find a way. The people telling you to panic about AI are the same type of people who worried about the printing press or the tractor.

Or, for that matter, the people in the 1960s who were convinced that automation was going to produce mass unemployment within a decade. President Johnson actually convened a commission on it in 1964. The commission concluded that automation would not cause permanent unemployment. They were right, in the aggregate.

There’s something to this. In 1900, roughly 40 percent of American workers were in agriculture. Now it’s under 2 percent. That’s an enormous displacement, and yet there was no permanent unemployment catastrophe.

The economy absorbed those workers over generations, first into manufacturing and then into services. If you look at the broad employment numbers from 1900 to today, the story is one of continued high employment through repeated technological disruption.

David H. Autor at MIT has done some of the most careful work on this. His research tracks how automation tends to hit what he calls “middle-skill, routine” jobs first: bookkeepers, assembly line workers, data entry clerks. Jobs that follow a predictable set of steps.

Meanwhile, low-skill manual work (cleaning, construction, elderly care) and high-skill cognitive work (management, creative work, medicine) were harder to automate. His work through the early 2010s mostly supported the “adaptation” story, though he’s been more cautious since, particularly in papers he’s written about the 2016-2020 period.

The reassuring version of history is accurate about outcomes over long periods. What it’s less honest about is the distribution of those outcomes across time and across people.

I should say: I find myself persuaded by the historical pattern, most of the time. Then I spend a day reading about what’s happening in customer support departments right now, and the persuasion wobbles a bit.

A company I know of reduced its customer service headcount by about 60 percent over 18 months after deploying AI tools (chatbots). The overall economy might eventually absorb those workers. The individual people had a very rough few months, some of them even more than a year.

What steam and electricity didn’t touch

What truly sets AI apart, or at least what feels different, is something I often reflect on, though I’ll admit my confidence fluctuates depending on which week you ask me.

Previous technologies replaced physical labor first. The steam engine did what human muscles did, but faster and cheaper. Electricity automated what human hands did, but at scale. Industrial robots in the 1970s and 1980s replaced workers who moved and assembled things.

Even early computers replaced clerical work, which was physically performed by people sitting at desks processing paper.

Cognitive labor was largely untouched. The person who analyzed the data the computer processed still had a job. The person who wrote the legal brief based on the research the paralegal compiled still had a job.

The person who designed the product the factory produced still had a job. If your work involved thinking more than doing, previous waves of automation generally left you alone.

AI doesn’t respect that division. A current large language model can draft legal briefs, analyze medical scans, generate code, produce marketing copy, design certain categories of graphics, and do passable financial modeling. Not perfectly. Not without human oversight.

But well enough to change what “human oversight” means in practice. Fewer people supervising more AI output, rather than people doing the work themselves.

There’s also something different about the speed of improvement. The steam engine took decades to develop from a novelty into something that genuinely restructured an economy. The systems doing cognitive work today are meaningfully more capable than the systems from three years ago, in ways that feel faster than the typical technology adoption curve.

I’m not sure how much weight to put on this, because the history of technology is full of moments where people thought change was accelerating and then it plateaued. But I don’t want to dismiss it either.

A friend of mine who runs a small design agency told me in 2022 that AI image tools were a toy. In 2023 he said they were useful for mood boards. In early 2025 he laid off one of his three junior designers. He didn’t say it was because of AI. But I noticed, and he later admitted it was because of AI.

The relevant question isn’t whether AI can replace a lawyer. It’s whether AI can replace three-quarters of the billable hours a junior lawyer at a mid-sized firm currently performs. Those are different questions with different answers.

And the same version of that question applies to radiologists, financial analysts, junior software engineers, and a fairly long list of roles that, until very recently, seemed safely in the “cognitive work” category that automation doesn’t touch.

The gap nobody talks about

In 1878, telephone operators were mostly men. Within a decade, the industry had switched almost entirely to women, partly because telephone companies decided women had better temperaments for the work and partly because they could pay women less.

The men who’d held those jobs went somewhere else. The women who replaced them held those jobs for several decades, until automated switching systems began replacing them in the 1920s and 1930s.

I’ve been thinking about that story lately, though I can’t quite make it prove what I want it to prove. The operators did lose their jobs eventually. But the industry grew so fast that even as individual exchanges automated, total employment in telecommunications kept rising for decades. The operators who got displaced in one place often found work at the companies building the new equipment or managing the new systems.

I was having lunch with a friend who does HR consulting last spring, somewhere around the time the bill came and we started arguing about who got it, and she said something that stuck with me. She said the companies she works with aren’t thinking about whether to use AI. They’re thinking about how many fewer people they need to hire this year. Not layoffs, mostly. Just smaller teams. Slower backfills. The kind of headcount reduction that doesn’t make headlines.

Does the telephone operator story mean retraining works? Partly. Does it mean the workers who got displaced in 1930 were fine? Not necessarily.

What actually worries me about the current moment isn’t whether new jobs will exist. They probably will. It’s the speed. The transition from handloom weaving to factory work took roughly fifty years. The transition from human telephone operators to automated switching took maybe thirty.

These are long enough timelines for a workforce to gradually shift. The displaced workers age out, the younger generation enters a different labor market, the pain gets distributed across time rather than concentrated on a specific cohort.

If AI capabilities keep compressing that timeline, the question isn’t whether the economy adapts. It’s whether the economy adapts faster than individual working lives can absorb. A 52-year-old paralegal whose skills become less valuable over three years doesn’t have fifty years to wait for the new equilibrium.

Neither do the thousands of entry-level programming jobs that used to serve as the training ground for senior engineers. If junior roles shrink because AI can handle that work cheaply, where do the senior engineers of 2035 come from? That’s a question I don’t see asked often enough.

There’s an entire argument here about what governments and companies should do about retraining. I don’t find myself particularly hopeful about it, based on how previous government retraining programs have gone, but I also don’t have good data on what actually works at scale, and the people who study this disagree sharply.

I don’t see obvious reasons why an AI retraining program would work dramatically better, though I’d genuinely like to be wrong about this.

Same fear, different century

In 1589, William Lee invented a knitting machine and brought it to Queen Elizabeth I hoping for a patent. She refused, reportedly saying she feared the machine would put her subjects out of work. Lee eventually took his invention to France.

People have worried about this for a very long time. And for most of that time, they were wrong, at least in the aggregate.

There’s a psychological mechanism at work here that’s worth naming, though I don’t know the exact study. Humans are good at imagining what gets destroyed and bad at imagining what gets created.

When the internet was arriving in the mid-1990s, it was easy to see that travel agents would lose business. Nobody imagined the app economy, or that there would someday be a profession called “influencer” or roles that would employ millions of people globally.

New job categories are, almost by definition, unimaginable before they exist. If I’d asked someone in 1990 to name the skills they’d need to become a successful “social media manager,” they would have had no idea what I was talking about.

This cuts both ways, though. If we’re bad at imagining the new jobs that AI will create, we might also be bad at imagining just how many existing jobs it erodes. The imagination failure works in both directions.

A 2013 paper by Carl Benedikt Frey and Michael Osborne at Oxford estimated that 47 percent of US jobs were at high risk of automation over the next decade or two. That figure got repeated everywhere. It was also somewhat contested, and the follow-up research was messier.

A 2016 OECD study applying similar methods to European countries got numbers closer to 9 percent. The spread between 9 percent and 47 percent tells you how uncertain the methods are, not how confident we should be in either number. Both studies used similar approaches to answering the same question and got results that differ by a factor of five.

The prior history of panics-that-turned-out-fine is real evidence. It just isn’t conclusive evidence. And I notice that the people invoking history most confidently tend to be people who are well-positioned in any scenario: academics with tenure, senior knowledge workers, people whose specific skills are among the last things AI will touch.

That doesn’t make them wrong. But it’s worth noticing who’s most comfortable with the historical argument.

Which humans, exactly

I keep coming back to a distinction that doesn’t get made clearly enough in these arguments.

When people say “technology creates new jobs,” they usually mean: new jobs exist in the economy after the technology arrives. They don’t necessarily mean the same humans who lost their old jobs get the new ones.

Sometimes that happens. Often it doesn’t, or not completely, or not for the people who are in the hardest circumstances: older workers, workers in specific geographies where the old industry was concentrated, workers without the educational background to make the transition.

The industrial revolution created enormous wealth and eventually raised living standards dramatically. It also produced, in the short and medium term, some genuinely miserable conditions for a lot of working people. Economic historians debate how long that adjustment period was.

The range of estimates is sobering. Some put it at 30 years before living standards for average workers began clearly improving, some put it longer. A generation that got ground up in the first decades of industrialization didn’t live to benefit from the prosperity it helped create.

What I can’t tell you, and what nobody can tell you honestly, is which category the current AI transition falls into. Is this another electricity moment, where the displacement is real, but the new opportunities are large enough and fast enough to absorb most workers?

Or is something genuinely different happening, where the cognitive nature of what’s being automated changes the math in ways the historical pattern doesn’t capture?

I’ve spent time reading people who do labor economics professionally, and the ones I find most credible tend to express genuine uncertainty rather than confident predictions in either direction. The confident predictions I’ve seen have an unfortunate tendency to be wrong within a few years, in both directions.

The people who were certain in 2012 that AI was nowhere near replacing professional knowledge work look bad now. The people who were certain in 2020 that GPT-3 was about to eliminate most white-collar jobs also don’t look great.

There’s a version of this story where the next twenty years look like a slightly rougher version of the last forty: some workers lose, more workers gain, the overall employment numbers stay relatively healthy, and we debate afterward about distribution.

There’s another version where cognitive automation hits fast enough, broadly enough, that the labor market can’t absorb it cleanly. There might be a third version we can’t see yet, involving job categories that don’t exist and won’t for another decade.

Both of those first two versions are consistent with the evidence we have right now. The honest answer is that we’re running an experiment without a control group, and the results won’t be clear for another decade at least.

What’s particularly strange about this moment is that the people building the technology often acknowledge they don’t know what the labor market effects will be, while simultaneously continuing to build as fast as possible. I don’t have a strong view on whether that’s irresponsible or just the nature of how technologies get developed, but it’s worth noting.

The question isn’t whether humanity survives this. It probably does. The question is which humans get to surf the transition and which ones get caught underneath it, and right now I don’t think anyone has a confident answer to that, including the people who sound most confident.

The people being most definitive in both directions, the “this is fine” camp and the “civilization-ending disruption” camp, seem to be working from the same incomplete evidence and reaching opposite conclusions because of their priors, not their data.

Though I’ll say this: the historical record suggests we’re better at helping the people who lose when we actually think hard about them in advance, rather than assuming the market will work it out on its own in due course.

Whether we’re actually doing that this time is a different question, and the conversations I hear and policy circles suggest we’re mostly still in the assumption phase, hoping the historical pattern holds and the adaptation happens on its own schedule without much active intervention.

Right now, I feel like a textile worker watching the first steam-powered looms arrive on the factory floor. Is this actually it? Or am I just too close to these AI tools and people using them heavily every day as power users, which makes me overestimate what they can do and catastrophize their impact on the broader job market?

I’ll have my answer soon enough. I hope I’m wrong.

In Case You Missed It:

Platforms have been running this kind of soft removal for years on content they don’t like. This practice is known as shadow banning. The difference now is they’re about to run it on people they can’t verify.

That’s where this is going. Not a ban. A filter.

A recruiter on LinkedIn searching for candidates clicks a toggle: “Verified profiles only.” They’re not trying to exclude anyone on principle. They’re drowning. In some hiring categories, AI-generated applications already account for the majority of inbound volume.

A person who has submitted one job application is sitting in the same pile as a bot that submitted four thousand, and the recruiter has no reliable way to tell the difference from the outside. The toggle isn’t malicious. It’s a desperate reaction to a system filled with fake AI-generated resumes.

But the person on the other side of that toggle who hasn’t verified for whatever reason, maybe they’re skeptical of biometric data collection (not surprised there), maybe they simply haven’t gotten around to it, doesn’t get a rejection. They don’t get anything. Their profile continues to exist. It just stops being visible to the people who would act on it.

The invisible category grows quietly until someone points at the numbers. By then the filter is default behavior, not an option. And opting out of verification looks increasingly like opting out of the platform itself, even if the terms of service never say that.

And it’s not just LinkedIn. All the big platforms like Facebook, Instagram, and X (Twitter) are pushing verification in one way or another, promising things like better visibility for your posts and improved filtering.

The flood nobody planned for

The volume of AI-generated content online has crossed some threshold where the old trust signals don’t work anymore. Profile pictures mean nothing. Work histories can be fabricated in seconds. Endorsements can be purchased or generated.

The signals that used to tell a platform “this is a real person with genuine intent” have been so thoroughly mimicked that they’ve lost most of their signal value.

There’s research on this. A 2023 joint report from Georgetown’s Center for Security and Emerging Technology1, OpenAI, and the Stanford Internet Observatory examined how language models could be weaponized for influence operations at scale.

The authors argued that AI-generated content would make existing campaigns harder to detect, specifically because text generation tools produce original output each time they run, removing the copy-paste patterns that researchers had previously used to identify synthetic activity. The threat isn’t just volume. It’s that the old fingerprints stop working.

The recruiting side of this is even more acute. Resume parsing tools have been weaponized. There are now AI services, some running openly, that allow a single person to submit applications at scale, auto-customize cover letters, inject keywords for ATS systems (they do not work, btw.), and generate portfolio samples on demand.

A recruiter at a mid-size sales company in Germany told me last fall, while we were both sitting through a very slow panel discussion on AI ethics that neither of us wanted to attend, that her team had started using a private spreadsheet to track which names appeared more than twice in their inbound pile in a given week. One name had appeared 47 times across different job listings.

That’s not an edge case anymore. That’s the environment recruiters are operating in, and they’re going to reach for whatever tool reduces that noise, including identity verification, even if they haven’t thought hard about what that tool actually does under the hood.

There’s a whole conversation about what platforms could do differently at the algorithmic level to identify synthetic volume without requiring individual biometric compliance. I’m not going to get into that here, partly because I am not an expert on this topic, and I do not think this will go anywhere politically likely in the next few years. However, some form of verification is on the horizon.

Verification is not what the brochure says

When platforms say “verification,” most users picture a blue checkmark. A quick email confirmation. Maybe a government ID scan, the kind you do at a car rental counter and forget about. The reality of what’s being built is considerably more involved.

Persona, one of the dominant identity verification providers used by companies like Coinbase, Brex, and a growing list of HR platforms, runs what the industry calls “identity graph” verification. You submit a government ID. You submit a selfie. The system maps your facial geometry against the ID photo and checks both against a database of known fraud patterns. It also, and this is the part the onboarding flow doesn’t emphasize, retains that biometric data as part of its fraud detection infrastructure.

Persona’s privacy documentation does disclose this. The issue isn’t that they’re hiding it. The issue is that almost no one reads it, and the platforms integrating Persona often don’t surface it clearly in their own UX. You click through a verification modal on LinkedIn or whatever platform has integrated the service, and you’re not thinking about what you’re consenting to. You’re thinking about getting past the gate.

I did this myself last year with a LinkedIn platform that required Persona verification to unlock basic features. I went through the whole flow in about four minutes on a Tuesday morning before I’d had coffee. I didn’t read the privacy policy. I have years of experience thinking about data infrastructure, and I didn’t read the privacy policy.

That bothers me more than I want to admit. But at that time, I thought that I had an old passport, so after the verification, I would still get a new one.

But the biometric data you generate during verification doesn’t just belong to the platform you were trying to use. It lives in a third-party system with its own retention policies, its own breach surface, and its own business model.

If you request the removal of your data, you’ll receive a response stating: “For LinkedIn-related verifications processed through Persona, we apply an automatic redaction policy to personal data collected on our platform. After these timeframes, data is automatically deleted in accordance with our retention practices.” However, they do not specify what these timeframes actually are.

Some of that data will outlast your account on the platform that requested it. Some of it may be shared with other clients of the verification provider for cross-platform fraud detection. Whether you think that’s a reasonable tradeoff depends on who you are and what you’re protecting.

This article, “I Verified My LinkedIn Identity. Here’s What I Actually Handed Over,” did a good job showing where our data goes. Solid article overall, but a few things I disagree with: like the $50 liability cap, sounds scary but it’s unenforceable against EU residents. GDPR Article 82 gives you an independent right to compensation that no US ToS can override.

And that Anthropic, OpenAI, Groq “processing your passport,” being listed as a subprocessor for “Data Extraction and Analysis” almost certainly means they’re used for OCR and document parsing via API, not that your passport is training data for ChatGPT or Claude. Big difference, still it's good to know where your data is going.

Persona, biometrics, and a problem with no clean edges

Biometric data is different from other personal data in a specific way that doesn’t get discussed enough in the context of verification: you can’t change it. If your password is compromised, you change your password. If your email address is leaked, you make a new one. Your facial geometry, your fingerprint, the vein patterns in your iris, these are not things you can rotate.

A data breach involving biometric records is permanent exposure. Not exposure until you update your credentials. Permanent.

There have been significant biometric breaches already. The BioStar 2 platform, used by banks, the UK Metropolitan Police, and defense contractors, exposed over a million fingerprints and facial recognition records in 2019. The data was sitting in an unprotected database. Researchers found it in less than a day. The people whose fingerprints were in that database had no way to remediate the exposure. They still don’t.

Researchers looking into Discord’s age-verification system found an exposed frontend linked to Persona, the identity verification service Discord uses.

What I’m describing here is a structural risk in the category of service they represent, not an accusation about their particular security practices. The risk exists whether or not any specific provider has been compromised yet.

There’s also the question of who audits the verifiers. Platforms audit their vendors to varying degrees. The verification providers are audited by their enterprise clients, who are mostly checking for uptime and compliance certifications, not for the downstream handling of the biometric data they’ve collected on behalf of thousands of users.

I don’t have a clean answer to what that oversight actually looks like in practice. I’ve asked people who should know and gotten answers that were confident but vague.

That question, who watches the watchers, stays with me. I haven’t been able to resolve it.

Invisible by design

This is how it actually works in practice, the invisibility part.

Platforms don’t need to mandate verification. They just need to create systems where unverified users receive slightly less distribution, slightly fewer impressions, slightly lower placement in search results. Nothing the user can point to. No policy they can cite in a complaint. Just a gradual dimming.

This is not speculation. Content moderation researchers have documented this approach across multiple major platforms going back to 2018, when Facebook and Instagram became the first to publicly acknowledge algorithmically reducing engagement with what they called “borderline content.”

Tarleton Gillespie, who has studied platform reduction policies in depth, argues that demotion exists precisely because removal creates conflict. Banned users complain, appeal, and sometimes organize. Users whose reach quietly drops by 80% mostly just assume their content isn’t connecting and post less.

The problem is that “targeted account types” expands over time. It starts with obvious spam. Then it includes accounts with incomplete profiles. Then accounts without verified contact information. Then accounts without verified identity. Each step feels like a minor trust improvement. Each step also slightly raises the barrier for participation.

Sometimes I think about how completely we’ve accepted the idea of the velvet rope. Not the good clubs with the velvet rope, which most people can’t get into and know they can’t get into, but the mundane version: the slightly better seat on the airplane, the faster security lane, the separate boarding queue.

These things were aggressively contested when airlines first introduced them. Now they’re invisible to most people in the same way the price of a checked bag is invisible until you’re at the counter. Not because resistance succeeded but because resistance failed, and then the next generation grew up thinking this was just what flying was.

Maybe that’s not a perfect parallel to digital verification, but you get my point.

Where I can’t follow my own argument

The part I keep getting stuck on is this: if AI agents really can simulate human behavior at scale, and the evidence suggests they increasingly can, then what exactly is the alternative to some form of human verification?

The answer I want to give is “better detection on the platform side.” But I’ve talked to enough people building detection systems to know they’re in a genuine arms race, and the offense is currently ahead of the defense.

Generating convincing synthetic behavior is cheap. Detecting it reliably is expensive and error-prone. The false positive rates on current AI content detection are bad enough that deploying them at scale would catch a lot of real people in the net.

So verification starts to look like the less-bad option. Not the good option, less bad.

I’m genuinely uncertain whether the people resisting identity verification are protecting something meaningful or acting out a preference that’s already becoming impossible to sustain. Perhaps they’re smarter than the rest of us who were foolish enough to click the “verify” button. My instinct is that the resistance is right on principle and losing on the ground. That’s not a comfortable place to land.

The version of this argument I can’t dismiss is the one made by people in countries where digital identity infrastructure has been used for suppression. India’s Aadhaar system was sold as financial inclusion. In states like Jharkhand, biometric match failure rates ran close to 50%, meaning nearly half of enrolled residents couldn’t authenticate to receive food rations. Human Rights Watch documented cases where that failure translated directly into starvation. The exclusion wasn’t always intentional. Sometimes it was just a fingerprint reader that didn’t work. That almost makes it worse.

China is the sharper version of this concern, the unified social credit score or well-documented cases in Xinjiang: iris scans, DNA collection, facial recognition tied to movement controls, used specifically against the Uyghur population. That’s not a hypothetical misuse. It happened, and the identity infrastructure made it possible at scale.

I’m writing primarily about professional platforms in Western contexts, where the power dynamics are different. But the infrastructure, once built, doesn’t stay in one context.

That’s probably worth thinking about more than most people doing the verification flow on a Tuesday morning actually do. More than I did, anyway.

The title of this article isn’t a metaphor. It’s a description of a mechanism that already exists and is being quietly extended. You won’t receive a notice. There’s no form to contest. The platform will continue to host your account, your posts will technically be public, and your profile will load if someone types in the direct URL.

The disappearance happens upstream, in the filters and ranking signals and recruiter toggles that determine whether anyone who doesn’t already know you exists ever finds you. That’s a different kind of removal than anything platforms have done before, and it requires a different kind of attention to notice.

The people most affected won’t be the ones who refused verification on principle after reading the privacy documentation carefully. They’ll be the ones who never got around to it, who didn’t have the right ID format, who found the flow confusing, who verified once and had the data rejected for reasons the system didn’t explain.

Quiet exclusion disproportionately lands on people who already have fewer resources to navigate bureaucratic friction. That’s not an accident of design. It’s usually a feature of how these systems scale.

In Case You Missed It:

This isn’t a personality flaw. It’s negativity bias, and it’s universal enough that researchers have tracked it across cultures, age groups, locations and contexts. A 2001 study by Baumeister and colleagues analyzed hundreds of psychology experiments and found that bad events produce more emotion, more persistent thoughts, and stronger impacts on behavior than neutral or good events of similar magnitude.

The ratio isn’t subtle. In relationship research, it takes roughly five positive interactions to balance one negative interaction. In memory studies, people recall negative images more vividly than positive ones even weeks later.

Your brain evolved this way for a reason. Ten thousand years ago, missing a piece of good news (ripe fruit in a distant tree) cost you a meal. Missing bad news (predator in the brush) costs you your life.

Natural selection didn’t optimize for happiness, and it didn’t care about your comfort. It optimized for not dying. The humans who survived long enough to become your ancestors were the ones whose brains treated every ambiguous shadow as a potential threat.

The problem is that the environment changed faster than the wiring. You’re not scanning for predators anymore, at least not for those that can eat you, unless you are living somewhere in the wilderness. So you’re scrolling X, Insta, TikTok, reading performance reviews, checking your bank balance.

But your threat detection system hasn’t updated its assessment. It still treats social criticism like physical danger, uncertainty like imminent attack, and any negative information as something that requires immediate, sustained attention.

The world got safer, but your brain didn’t get the memo. And now that mismatch is running in the background of every interaction you have.

Nine compliments vanish, one criticism sticks

When your brain gets nine positive comments and one negative one. The positive comments get processed through your brain’s reward system, which is designed to notice good things and then move on. Dopamine spikes briefly, and you feel pleased. Then the system resets because, in evolutionary terms, contentment is dangerous. If you’re satisfied, you stop looking for threats.

The negative comment goes somewhere else entirely. It hits your amygdala, a small, almond-shaped structure within the brain’s temporal lobe, which doesn’t reset. The amygdala’s job is threat detection, and it has a much longer memory than your reward system. It’s also louder. A 2003 study (A PET Activation Study) using fMRI scans found that negative images produced significantly more activation in the amygdala than positive or neutral images, and that activation lasted longer.

You can see this in hiring, too. Managers routinely overweigh negative information about candidates. One mediocre reference outweighs three strong ones. One gap in a resume gets more attention than five years of solid performance. I still remember a case in 2010 where hiring committees spent 40 minutes discussing a candidate’s single weakness and five minutes on their strengths, then claimed they gave both “equal consideration.”

You can see this in relationships, too. Partners routinely overweigh negative interactions. One cold response after dinner outweighs three weeks of warmth and effort. One forgotten anniversary gets more attention than five years of showing up consistently.

The bias compounds when you’re making decisions under uncertainty. If you don’t have complete information, your brain fills in the gaps with threat assumptions. Silence from your boss becomes evidence of disapproval. A colleague’s distraction becomes a signal they’re upset with you. An ambiguous email gets read in the most negative plausible tone. This is exactly how our brain works.

And the pattern accelerates in environments designed to exploit it. Social media platforms don’t show you a representative sample of human behavior. They show you whatever generates engagement, and negative content generates more engagement than positive content. Outrage spreads faster than joy.

A 2018 MIT study “The spread of true and false news online” tracked 126,000 news stories on Twitter (now X) and found that false news reached 1,500 people six times faster than true stories, largely because false news tended to be more negative and more novel.

Your feed isn’t a window into reality; it’s sadly a window into what your amygdala can’t ignore.

Gratitude journals, forced smiles, and strategic avoidance

The standard advice for dealing with negativity bias falls into three categories, and all three make the problem worse if you follow them literally.

Fix one: Ignore negative information.

The logic here is that if your brain overweights bad news, you should consciously underweight it. Treat criticism as noise and focus on the positive. This works until you encounter negative information that actually matters. Your bank account is overdrawn, your project is behind schedule, or a relationship is deteriorating. Ignoring real problems because you’ve trained yourself to dismiss negative signals is how small issues become catastrophic ones. We’ve all been there, ignored a small issue that spiraled into a total disaster, we still think about years later.

I learned this when I got my first car, a Hyundai Getz. I really loved that car back then, even though my friends teased me about the brand. And I stayed loyal to this brand over the years.

One day, the “check engine” light came on, but the car seemed fine. I told myself it was probably just a faulty sensor. I chose to ignore the warning and focused on how well it was running. A few weeks later, I was stranded on the side of the highway with an overheated engine. A simple fix had turned into a major, expensive repair. That day taught me that while staying positive is good, ignoring clear warning signs can lead to much bigger problems. Some problems will never go away, no matter how positive your mindset is.

Another thing I learned back then was that I really need to know more about the cars I’m driving.

Fix two: Force positivity.

Gratitude journals, affirmations, and deliberate reframing of every setback as a learning opportunity. The research on this is mixed at best. Some studies show benefits from gratitude practices, but they tend to be small and they fade. A 2023 meta-analysis found that gratitude interventions produced effect sizes around 0.22 for positive well-being, meaningful but not transformative. The effects tend to be larger in populations that were initially depressed, smaller in healthy populations trying to boost baseline happiness.

More importantly, forced positivity creates a secondary problem. If you’re trying to feel grateful and you don’t, you now have two problems: the original negative event and your failure to properly appreciate your circumstances. At the end, you feel bad about feeling bad.

Fix three: Numb out.

Reduce exposure to negative information entirely. Quit social media, stop watching the news. Ideally, avoid any difficult conversations you might have. This works in the short term. Your ambient anxiety drops.

But you’re not fixing the threat detection system, you’re just removing the triggers. And the system doesn’t get less sensitive when you remove stimuli. It gets more sensitive, sometimes way more sensitive. The longer you avoid something, the larger it looms. When you finally encounter negative information again, your response is bigger than it would have been if you’d maintained exposure.

I’m not saying these strategies never work. Gratitude practices can be helpful for some people, and there were times in my life when I’d say they worked for me. Reducing doomscrolling is probably net positive for most. I can tell you from experience, it really worked for me. When I took a break from Facebook and Insta, my anxiety went down, and my mood and focus got a lot better.

But they’re treating the symptoms, not the mechanism. Your brain is doing exactly what it evolved to do: prioritize threats. The question isn’t how to convince your brain to stop. The question is how to give it a task that actually helps.

Your amygdala still thinks you live in 8000 BCE

The mismatch between ancient wiring and modern environment shows up most clearly in what psychologists call “probability neglect.” Your brain is not good at calibrating risk. It treats low-probability, high-severity events (plane crashes, terrorism) as more dangerous than high-probability, low-severity events (car accidents, poor diet). It overweights vivid, recent threats and underweights abstract, distant ones.

This made sense in the ancestral environment. If you saw a lion once, you needed to remember it forever. The base rate didn’t matter; you didn’t need to know that lions only attacked humans in 0.01% of encounters. You needed to know that lions are dangerous and to avoid them. A simple survival instinct.

But in a world where most threats are probabilistic rather than binary, this creates systematic errors. You worry more about stranger danger than car accidents, even though your kid is vastly more likely to die in a vehicle. You worry about terrorism more than heart disease. You treat market volatility as more dangerous than sustained low returns.

And because modern media is optimized for engagement, not accuracy, you’re constantly being fed the vivid, recent, high-salience threats. The algorithm doesn’t show you “nothing happened today in 99.9% of places.” It shows you the plane crash, the mass shooting, and the financial collapse. Your amygdala catalogs each one as evidence that the world is getting more dangerous, even when aggregate statistics show the opposite.

There’s a version of this that applies to personal threats, too. You remember the person who criticized you more vividly than the ten people who didn’t. You remember the job you didn’t get more than the ones you did. You remember the relationship that ended badly more than the quiet years when things were fine.

Your brain is building a threat model from exceptions, not base rates. And then using that model to guide your behavior.

The feedback loop that creates what it fears

Negativity bias doesn’t just make bad things feel worse. It creates feedback loops that generate more bad things to feel bad about.

You get critical feedback at work, and your brain flags this immediately as a threat. Threat responses narrow your focus. You start monitoring your boss’s behavior more closely. Looking for signs of disapproval. And of course you find them, because confirmation bias kicks in once you’ve decided what you’re looking for. Your increased vigilance makes you more anxious. The anxiety affects your performance. Your performance drops, you get more critical feedback. The loop tightens.

Or, you see negative news about the economy. Your brain catalogs this as a threat to your financial security. You start checking your portfolio more often. Frequent checking exposes you to more volatility, which increases anxiety. Anxiety makes you more likely to sell during a downturn. Selling during a downturn locks in losses. Losses confirm the original threat assessment. You check more often.

The pattern shows up in relationships too. You interpret your partner’s distraction as evidence they’re upset with you. You withdraw or become defensive. They notice the shift in your behavior and actually become upset. Your original interpretation is now validated, but only because you acted on it first.

I’m not convinced this pattern is universal. There are people who seem less prone to these spirals, and I can’t tell if that’s because their baseline negativity bias is lower, because they’ve developed better circuit breakers, because their DNA is different, or because they’re in environments that don’t activate the bias as strongly. Maybe they were influenced by their parents or grandparents, but that’s just a guess. But in the cases where the loop does form, it’s self-reinforcing.

And the modern environment is full of systems designed to activate it. Performance management systems that focus on gaps rather than strengths.

Social media platforms that surface conflict and anything that creates emotion. News cycles that prioritize threats, financial systems that encourage constant monitoring. Each one is tugging on the same evolutionary wire.

Building a counterweight, not fighting the bias

The fix that actually works isn’t about reducing negative input or increasing positive input. It’s about building a deliberate counterweight to the bias that’s already running.

Your brain automatically amplifies negative information. You can’t stop that, but you can build a parallel process that deliberately amplifies positive information in a way that doesn’t feel forced or fake.

This is different from gratitude journaling. Gratitude journaling asks you to notice good things. That’s fine, but noticing isn’t the same as amplifying. Your brain already noticed the good things. It just didn’t file them in long-term storage because they weren’t threats.

What works is building rituals that force elaboration on positive events. Elaboration is what creates strong memories. When something good happens, your brain’s default is to notice it briefly and move on. The intervention is to interrupt that process and make yourself elaborate: What specifically happened? What did it feel like? What does it suggest about the person who did it? What might it predict about future interactions?

There’s research on this from 2006 by Sheldon and Lyubomirsky. They tracked people doing various positive psychology interventions and found that the interventions only worked when people varied the timing and elaborated on the details. Rote repetition of “I’m grateful for X” showed no sustained effect. Detailed elaboration on why X mattered and what it meant showed modest but persistent effects.

The mechanism matters. When you elaborate on a positive event, you’re not lying to yourself or forcing optimism. You’re correcting for a known bias in your storage system. Your brain stored the negative event automatically with high detail and high salience. You’re manually doing the same thing for the positive event.

In practice, this looks like this: Someone compliments your work → Default response: brief pleasure, then move on. → Intervention: Stop. What specifically did they compliment? What effort did you put in that they might have noticed? What does their noticing suggest about how they view your work more broadly? Spend 60 seconds on this.

Or, a project goes well. → Default response: relief, satisfaction, move to next project. → Intervention: What specifically went well? Which decisions contributed to the outcome? What does this suggest about your capability in this domain? Store it with the same detail you’d use if you were analyzing a failure.

This isn’t about self-esteem. It’s about building an accurate threat model. If your brain only stores failures in high detail and successes in low detail, your model of your own capability is wrong. You’ll systematically underestimate what you can handle and overestimate the risks you face.

The intervention works because it matches the way your brain actually operates. You’re not fighting negativity bias. You’re building an equal and opposite process that runs in parallel.

Does it eliminate anxiety? No. Should it? I’m not sure. Some level of threat vigilance is probably adaptive. The goal isn’t to feel good all the time, the goal is to have a threat model that reflects reality instead of evolutionary history.

And let's face it, that is quite hard to do because our brain still thinks that we are chased by predators, so it’s just trying to help us survive.

When the amplification ritual becomes the problem

There are four patterns I’ve seen repeatedly where the standard advice breaks down, and what that suggests about when you should and shouldn’t try to correct for bias.

Most of the advice I’ve heard over the years on managing negativity bias treats it as a universal problem with a universal solution.

But there are situations where your threat detection system is giving you accurate information, and trying to amplify positive signals just adds noise:

Read more

Several mentioned specific details from her profile, her company’s recent investment, and the fact that she’d spoken at a conference two weeks earlier.

Every single one of them was trying to pitch their recruitment agency services with a hyper-personalized email.

“Five years ago, this would’ve meant something,” she said.

Now she deletes them in batches without reading past the first line. Because she knows. They all know the messages are good, too good, and far too many. Written by tools that can scrape a profile, analyze posting patterns, and generate something that sounds like it came from a human who actually read her work and thought about it for ten minutes.

The crazy part isn’t that the messages are obvious spam. It’s that most of them aren’t, and they all pass her company and Gmail filters without any problems.

Inbox looked normal until it wasn’t

The shift happened sometime in late 2024, though nobody I’ve talked to can pin down an exact month. Messages started getting better. Not in the “Hey {FirstName}, I noticed your company” obvious mail-merge way we knew for years, but they actually got better. They referenced specific posts, used natural language, made relevant observations, and included details that suggested the sender had spent real time on research.

Jana said the first few felt flattering. Someone at a vendor company sent her a message about a talk she’d given, mentioned a specific framework she’d used, asked a thoughtful follow-up question. She responded to those messages, as she thought a human had written them. They even had a decent exchange.

Then she noticed the same person had sent near-identical “personalized” messages to six other recruiters she knew. Same structure, different details swapped in. It was a Tuesday afternoon, she remembers that because she had a candidate interview at 3pm that she almost missed because she was too annoyed by the realization.

The email volume accelerated. Ten personalized messages a week became thirty. Then fifty, often sent not only to her company email but also to her LinkedIn, and then her WhatsApp started getting them too. All thanks to AI and multichannel outreach. If there is one thing I personally do not like, it is this.

She started getting text messages even on her personal phone number, which she’s pretty careful not to share publicly. Each one had enough context to feel legitimate. Not all of them were selling something obvious. Some were just... conversations. Networking attempts, people “reaching out” because they’d “been following her work.”

She stopped responding to anything that felt too smooth. If a message had zero typos, perfect grammar, and hit every psychological trigger in the relationship-building playbook, it went to trash.

The irony is that now she probably deletes real messages from real people who just happen to write well. But the math doesn’t work anymore. When responding to one message costs five to ten minutes (if you are not using AI), and the hit rate on genuine humans is maybe one in twenty, you stop playing.

The economics of fake intimacy

Personalization used to be expensive. Writing a good cold email meant research, time, and having the right skills. You had to read someone’s work, understand their context, find a genuine connection point. That took time. And as we know, time is money. So personalization was reserved for high-value targets. When you got a truly personalized message, it meant something because it meant someone had invested scarce resources in reaching you specifically.

AI makes that investment free. Or close enough to free that it rounds to zero. You can now generate a hundred personalized messages in the time it used to take to write one. Each message can reference specific details, mirror communication styles, adapt to the recipient’s industry, and do it all without human intervention past the initial setup.

There’s interesting research on this, though not specifically about AI-generated outreach. Paul Rozin at Penn studied something he called “contagion” in the 1980s, looking at how people value objects. A sweater worn by a celebrity is worth more than an identical unworn one. But if you tell someone the celebrity sweater was duplicated perfectly, the value collapses. Not because the quality changed. Because scarcity did.

The same mechanism breaks personalization. When everyone can generate messages that look hand-crafted, the hand-crafting loses value. The signal degrades. A thoughtful email used to mean “this person cares enough about reaching me to invest twenty minutes of focused effort.” Now it means “this person has access to ChatGPT and a LinkedIn scraper.”

The flood is already here in sales and recruiting. It’s coming for every other channel. Fundraising, political campaigns, and especially in scams. It's even coming for personal communication between people who actually know each other, because why wouldn’t you use AI to help craft a better apology, a clearer explanation, a more thoughtful thank-you note?

I’m not sure where the line is between “using a tool to communicate better” and “automating communication to the point where it stops being yours.” I suspect that line is fuzzier than we want it to be, and I also suspect we’re going to cross it repeatedly without noticing until the damage compounds.

It feels a bit weird when someone reaches out on LinkedIn saying they admire your work, but you can tell they haven’t even visited your profile, and the message is AI-generated, thanks to LinkedIn AI. And what do you do? You reply with a message also generated by AI, or at least use AI to fix the grammar.

Throughout this whole process, the human part was pretty small. All we had to do was click “generate” to create the content and then “send” to share it.

Spoofed voices and cloned faces aren’t edge cases anymore

My friend Petr’s mother called him two months ago. She was crying. She said she needed money immediately; there'd been an accident, she couldn't explain on the phone, and asked him to send 2,000 euros to this account number right now. He was about to do it, because the voice was perfect; he told me that the crying sounded real.

He called her back on the same number to confirm the account. She answered that she was fine, just watching television at home. Had no idea what he was talking about. Someone had cloned her voice, probably from a video she’d posted to Facebook a few years ago.

This isn’t theoretical anymore. Voice cloning tools are widely available; ElevenLabs and many others can do that within minutes, sometimes in seconds. Video deepfakes are getting harder to spot. The technology to spoof a video call exists and is getting cheaper every month. I read about a case in Hong Kong where a finance worker transferred $25 million after a deepfaked video conference call with people who looked exactly like the company’s CFO and other executives. That was February 2024; if you check new versions of Kling or Seedance, these tools are way better now.

What happens when the base assumption shifts from “if I hear someone’s voice, it’s probably them” to “I can’t trust audio or video without secondary verification”? We don’t have cultural norms for that yet. We’re still operating on pre-AI threat models where sophisticated attacks were rare enough to be newsworthy.

A professor of Psychology at Rider University, John Suler, identified six factors that make people behave differently online versus in person. Anonymity was one, but so were invisibility, asynchronicity, and what he called “dissociative imagination1,” the sense that online interactions aren’t quite real. That was 2004. Now add perfect voice mimicry, video that looks real, and AI that can hold a conversation in real-time with almost no latency, while pretending to be someone you trust.

The question isn’t whether this will be used maliciously. It already is, but the question is how fast the cultural immune system adapts. I often wonder, if its even possible.

What makes something feel real when everything can be simulated

We rely on dozens of small signals to determine authenticity: voice timbre, word choice, and response latency. The way someone structures an explanation, typos in predictable places, even though AI is now able to fix any text. References to shared context that would be hard for an outsider to fake.

All of those signals can be mimicked now. Not perfectly, not yet, but well enough to fool most people most of the time. And the error rate drops every few months. Just take Instagram or TikTok videos, a few months ago you would spot an AI video immediately, now you wonder, is that person real or another AI?

When authenticity can be faked perfectly, trust shifts from content to metadata. Not what someone says, but how they say it and through what verified channel. We’re already seeing this with two-factor authentication, blue checkmarks, verified badges. But those systems break too. Verification becomes a new attack surface.

There’s probably a version of this where decentralized identity systems solve the problem. Cryptographic signatures, web of trust models. But I’m quite skeptical about these things. Not because the technology can’t work, but because adoption requires changing behavior for billions of people, and the threat model is invisible until it personally affects you. Most people won’t implement security measures until after they’ve been burned. By then the damage is done.

I am always shocked by how many people do not even have 2FA for their emails, unless their email provider forces them to use it.

Safe words, verification codes, and…

I’ve started hearing about families creating verbal passwords. My friend Alice, who lives in London, told me that her parents decided on a code phrase after reading about voice-cloning scams. If anyone in the family calls asking for money or claiming there’s an emergency, they have to use the phrase first. She said it feels paranoid and also completely reasonable, which is a weird combination.

Companies are testing similar systems. Some are implementing secondary verification for any request involving financial transfers, even if it comes from a known internal email address. Call back on a verified number, confirm via a different channel.

The cost is time and annoyance. Every verification step slows down legitimate communication. If you’re in a role where fast decisions matter, adding three verification layers to every high-stakes conversation could be the difference between closing a deal and losing it. But not adding those layers could be the difference between keeping your job and getting fired for falling for a scam that should’ve been obvious in retrospect. Perhaps I’m being overly cautious, but this is the reality we’re facing.

I don’t think we’ve figured out the right tradeoff yet. We’re still in the phase where most people don’t implement any protection because the threat feels abstract or something you are reading about on the internet.

That changes the first time someone they know gets hit. Then the pendulum swings too far in the other direction, everything gets locked down, and productivity craters. Eventually, we find some middle ground. But there’s a transition period that’s going to be messy.

The other problem is that authentication systems create exclusion. If you require cryptographic verification for all communication, you’ve just cut off anyone who doesn’t have the technical literacy to set that up. All my friends use WhatsApp, I am trying to avoid it as it's a plague, but getting them to Signal is almost impossible.

The security layer becomes a barrier to connection, especially for elderly relatives. Maybe that’s an acceptable tradeoff, I’m not sure.

Friction as a signal

There is something strange I keep noticing. The most credible messages I get now are the ones with mistakes. Not obvious spam mistakes, but human ones. Typos that autocorrect didn’t catch, sentences that wander. An apology for taking three days to respond because they got busy with something unrelated.

A client sent me a message last week that started with “Sorry, just saw this, my kid had a thing at school and then I forgot to check email until now.” That sounds real. AI doesn’t forget to check email. AI doesn’t have kids with school events. AI would generate a response within seconds or wait a calculated amount of time to simulate busyness, but it wouldn’t explain the delay with a specific mundane excuse that adds no strategic value to the message.

I realize this is temporary, eventually AI will learn to add strategic imperfection. You can do it now with the right prompting, but it still does not feel real, at least to me. The next generation of tools will insert deliberate typos in statistically appropriate places. Add random delays, generate plausible excuses for slow responses. But we’re not there yet (at least I don't think we are), and in the gap between now and then, sloppiness becomes a weak signal of authenticity.

The long-term equilibrium probably isn’t “imperfection equals authenticity.” That’s too easy to game. More likely we end up in an arms race between authentication and spoofing, with verification costs rising over time until most communication happens through gated channels that require some form of identity proof.

Email becomes unusable, direct messages become unusable. Maybe we all retreat to small, verified networks where everyone knows everyone else and strangers can’t get in without a referral, something like Facebook or LinkedIn, but for people we really know.

That has costs too, network effects matter. Open communication enabled a lot of valuable things. Weak ties connect disparate communities. Serendipitous connections happen because someone you don’t know can reach you. Close all that down and you lose something real.

I don’t have a clean answer here. The flood is already happening and it’s going to get worse. Most of the proposed solutions either don’t scale or create new problems. We’ll probably muddle through with some combination of verification systems, behavioral adaptation, and resigned acceptance that a certain percentage of communication is now just noise you have to filter.

What I do know is that the thing I used to value about personalized outreach, the sense that someone had invested time in understanding who I was and what I cared about, doesn’t work anymore. The signal is degraded past the point where I can trust it. And I don’t think we’re going back to a world where personalization means scarcity. The tools exist, and people will use them.

Everything these days also feels like a numbers game. I still remember the time when recruiters spent time writing personalized emails, and now? It’s easy to reach 5,000 people in minutes; the only thing you need to do is hit one button and add as many recipients as you’d like.

Maybe five years from now we’ll have cultural norms that make this manageable. Maybe we’ll look back and laugh at how we used to trust email addresses and phone numbers as identity proof. Or maybe we’ll be stuck in an equilibrium where every conversation requires so much verification overhead that we only communicate with people we already know, and the cost of establishing trust with a stranger is prohibitively high.

I still check my messages. I still respond to some of them. But the bar keeps rising for what counts as credible enough to warrant attention, and I’m probably missing real opportunities because I can’t tell them apart from generated noise.

Articles you shouldn’t miss:

On Monday, one tiny thing broke. A button stopped saving settings after a refresh. Nothing dramatic, just annoying.

We jumped on a call. It was late, someone’s microphone kept rubbing against a hoodie string, and I was eating the kind of sad desk snack that leaves salt on your fingers. He pulled up the code and started scrolling like he was looking for a familiar street sign in a city he’d visited once.

He built almost all of it with an AI assistant (Claude 4.5).

When the bug appeared, he had no mental map. Not “I forgot the exact line,” but “I don’t know what this part is supposed to do.” He could describe the UI. He could describe the feature. He couldn’t explain the flow. He couldn’t predict where state lived, or what got cached, or what re-ran when the page loaded.

So he did what most people do now. He pasted the error into the model and waited for a fix.

It suggested a change. He applied it, and he got a new error.

Another suggestion. Another patch, another new error. And now the tool still didn’t work, plus the code was drifting into a shape neither of us recognized.

I’ve been around software teams, analytics teams, and recruiting teams long enough to recognize it. Not incompetence, not laziness, but something else entirely. It’s that specific kind of competence that ships, but can’t explain itself.

AI makes you faster at producing outputs. It also makes it easier to skip the part that turns work into skill.

People don’t notice because everything looks fine until it doesn’t. The first time you need to debug, extend, teach, defend, or rebuild without the helper, you find out what you actually learned.

What You Lose Before You Realize It

The problem isn’t that AI is “too good.” The problem is that it’s good in a way that changes your posture.

Before, you wrote a thing, and your brain had to keep track of cause and effect. You held a few constraints in working memory, you tested a guess, you got it wrong, you felt the friction, and then you adjusted. You built a messy internal model as you went.

Now the workflow is different. You prompt. You receive. You skim. You accept or reject. The role quietly shifts from “builder” to “editor.”

That sounds fine, because editing still feels like thinking. Sometimes it is, often it isn’t.

There’s research on what happens when humans rely on automated aids. Parasuraman and Riley1 framed it as misuse and disuse of automation, including over-reliance and reduced monitoring, which can lead to errors being missed even by capable operators.

That’s the part most people associate with planes and medical devices. The same pattern shows up in everyday knowledge work, just in a less dramatic costume.

You accept output you didn’t generate.
You stop checking parts that used to be effortful.
You lose the ability to notice what’s off.

It gets worse, because AI is fluent. It produces confident prose, plausible code, tidy summaries. Fluency tricks the brain. You feel like you understood because you recognized the words and the structure. Recognizing something is easy, but recalling it from scratch is hard.

And retrieval is the expensive step that builds the wiring.

One reason learning sticks is that forcing yourself to pull information out of your head changes memory. Roediger and Karpicke2 showed that taking tests improves later retention, even without feedback, compared to repeated studying, especially when you measure days later rather than minutes later.

That’s not a motivational poster idea. It’s a finding that shows up again and again, including in applied settings like medical education, where repeated testing can improve long-term retention compared to restudying.

AI-heavy workflows remove a lot of those forced retrieval moments.

You don’t have to recall the syntax.
You don’t have to wrestle the structure.
You don’t have to generate the explanation.

You just judge the output.

There’s another angle here that makes me uneasy, because it’s not only about memory. It’s about where your brain expects knowledge to live.

Sparrow, Liu, and Wegner’s3 “Google effects on memory” paper found that people tend to remember where to find information rather than the information itself, and that thinking about hard questions can prime thinking about computers.

That’s not “internet bad.” It’s more specific: when external access is reliable, the brain adapts by shifting effort away from internal storage. You can argue that this is rational. You can also admit it changes you.

A quick digression that I can’t fully defend, except that I keep noticing it: I’m seeing younger people get surprisingly calm about not knowing basics. Not embarrassed, just calm. Like “why would I store that in my head.” Maybe it’s fine. Maybe it’s healthy. I don’t know. I just don’t think we’ve sat with what it means when the default is “I can fetch it” for everything, including the parts you used to need in order to think.

The Change You Don’t Feel Right Away

If you want the short explanation, it’s this: skill comes from producing, not approving.

When you generate an answer yourself, even a bad one, you expose gaps. You get immediate error signals. You also encode the path you took. That path is often more valuable than the final answer.

There’s a classic finding called the generation effect. People remember information better when they generate it themselves rather than just reading it. Slamecka and Graf4 demonstrated this across multiple experiments in 1978.

When you use an AI assistant for first drafts, first solutions, first explanations, you skip generation. You get the final product without the cognitive work that makes it “yours.”

The result looks like competence. You can deliver. You can talk about it at a high level. You might even sound smarter because the phrasing is cleaner than what you’d write on your own.

Then you hit the moment that requires an internal model.

Debugging is a good example. You can’t debug by vibes. You need a chain of reasoning about what should happen, what did happen, and where the divergence likely lives. That chain comes from having built similar chains before.

People assume the danger is “hallucinations.” That’s real, but it’s the loud risk. The quieter risk is that you stop doing the kind of thinking that makes you resilient.

I’m not arguing for some romantic version of suffering. A lot of struggle is wasted time. I’ve wasted plenty of it.

I’m saying there’s a category of friction that’s doing a job, and AI removes it by default.

Why Smooth is a Warning Sign

It’s hard to accept, but when you’re focused on output, feeling like the learning process is going smoothly is often a bad sign.

When studying feels easy, you’re usually rereading, recognizing, nodding along. That produces comfort, not capability. When studying feels harder, you’re often retrieving, generating, or spacing practice. That discomfort is part of why it sticks.

The testing effect research is one example. Another is spacing. Another is interleaving. They all share the same annoying feature: the practice feels worse in the moment.

AI makes work feel smoother. It reduces pauses. It reduces dead ends. It reduces the “wait, why isn’t this working” loop.

That loop is where a lot of learning lives.

Automation research also points to a monitoring problem: when a system usually works, humans monitor it less, and they get worse at noticing when it fails. Parasuraman and Riley’s misuse concept includes complacency and reduced vigilance.

Put those together and you get a nasty combo:

• Less generation

• Less retrieval

• Less monitoring

• More fluency

• More trust

I’ve seen this on small teams and side projects, not because people are careless, but because they’re busy. They’re rewarded for shipping. They’re not rewarded for being able to rebuild the whole thing from scratch a month later.

Do not get me wrong, I like these AI tools. I use them every day. I’ve shipped faster because of them. I also notice that if I let the model do the first draft of my thinking too often, I get lazier about holding the whole system in my head.

I catch myself doing it with writing.

I’ll ask for a paragraph, think “yeah that’s basically it,” paste it, and move on. Then someone asks me a sharp question about a claim inside it, and I have that awful moment where I realize I agreed with a sentence I didn’t fully earn.

That’s on me.
It also feels like a predictable outcome of the tool.

One more messy story, I was building a simple salary calculator with a friend, just a weekend thing, nothing tied to my job. We were half arguing about whether the rounding should happen before or after deductions. It was raining. Someone kept sending memes in the group chat like we were not actively trying to finish. I used the model to generate the tax logic for one country I didn’t know well.

It produced something that looked correct. Variable names, comments, even edge cases.

A week later a user emailed: “Your net pay is off by about 70 euros.”

We traced it back. The model had assumed a threshold that used to be true but wasn’t anymore. I should’ve checked the source law table. I didn’t. I saw tidy code and felt safe.

That mistake doesn’t prove AI makes you dumb. It proves something smaller and more annoying. That makes it easier to skip verification because the output looks finished.

And yes, I’ve also seen the opposite, where AI helped someone learn faster because they used it like a tutor and forced themselves to explain each step. So the effect isn’t automatic, but it depends on posture.

My Current System (Imperfect but Real)

Don’t ban AI, but constrain its use. On purpose. In a way that keeps the learning loop intact.

Here’s a system I’ve used, and still fail at sometimes.

Read more

Many people think the gap between where they are and where they want to be is a motivation problem. They watch TED talks, like me, listen to podcasts about finding their “why,” and consume endless content about discipline and drive.

The reality is different. You have responsibilities. You have limited time. You have real constraints.

But you can transform your output without needing to “get motivated” at all.

You just need to understand what motivation is actually compensating for.

A pattern I’ve noticed in people who consistently execute: they don’t have more motivation than you. They don’t wake up more inspired. They don’t possess some magical reservoir of enthusiasm that you lack.

They’ve eliminated the invisible friction that makes starting feel impossible.

There are several ideas I want to share with you about resistance, action, and why everything you’ve been told about motivation is backwards.

This took me years to figure out. I couldn’t stay motivated no matter what I tried. I read book after book about motivation. Nothing worked. Maybe 1 out of 20 tips actually helped.

The tips below worked for me, and who knows, maybe they might work for you too.

Why Motivation Feels Necessary (But Isn’t)

“The impediment to action advances action. What stands in the way becomes the way.”
Marcus Aurelius

Your brain is an energy management system optimized for survival.

Every action you consider taking gets filtered through a cost-benefit analysis that happens faster than conscious thought. Your brain calculates: How much energy will this require? What’s the guaranteed outcome? What’s the risk?

Your brain performs constant cost-benefit analyses faster than conscious thought

That should terrify you. It's terrifies me.

Because here’s what that means: your brain is actively looking for reasons NOT to act. It’s searching for uncertainty, difficulty, complexity, and using those signals to justify inaction.

This isn’t a bug. It’s a feature.

For 200,000 years of human evolution, conserving energy kept us alive. Unnecessary movement meant wasted calories. Uncertain outcomes meant potential death. Your ancestors who hesitated before running into the dark forest survived to pass on their genes.

You inherited a brain designed to avoid action unless absolutely necessary.

Think about the last time you needed to start something important. A work project. A difficult conversation. A lifestyle change.

Your mind generated a list:

• Maybe I should research the best approach first

• I’m not in the right headspace right now

• I should wait until I have more time to do it properly

• Let me just handle these smaller things first

• I need to be more prepared before I begin

• I…. add any of your excuses.

That’s not you being lazy.

That’s 200,000 years of evolutionary programming running its optimization protocol.

The problem is that modern life requires the opposite of what kept your ancestors alive. Success today demands action in the face of uncertainty. Progress requires starting before you feel ready. Achievement means moving when the outcome isn’t guaranteed.

Your brain experiences this as a threat.

So it does what it’s designed to do: it generates resistance. Doubt. Hesitation. A vague sense of “not yet.”

And here’s the trap: motivation is what you use to overpower that resistance.

You aren’t failing to act because you lack motivation.

You’re burning motivation just to overcome the invisible barriers your brain has constructed.

The Three Types of Resistance

Before we can eliminate resistance, we need to understand what we’re actually fighting.

Most people experience resistance as a single, monolithic force. That vague sense of “I don’t want to do this right now.”

But resistance has structure. It has components. And each component requires a different solution.

Activation Energy Resistance

This is the gap between intention and initiation. The moment where you know you should start, but you’re stuck scrolling, reorganizing your desk, or suddenly remembering seventeen other tasks that feel urgent.

Physics has a term for this: activation energy. The minimum energy required to start a chemical reaction.

Your brain treats every new action as a reaction that needs sufficient energy to begin. The more complex the task appears, the higher the activation energy required.

Psychological Reactance

This is your brain’s allergic reaction to feeling controlled. Even when you’re the one creating the expectation, even when you genuinely want the outcome, part of your mind rebels against the obligation.

You’ve experienced this: the project you were excited about on Friday suddenly feels oppressive on Monday morning. The commitment you made with enthusiasm now feels like a constraint.

Your brain is protecting your sense of autonomy by making you resist your own decisions.

Uncertainty Paralysis

This is what happens when your brain can’t predict the outcome clearly enough. You have twelve possible approaches. You’re not sure which is correct. You don’t know if you have the skills required. You can’t visualize the end state.

So your brain selects the safest option.

Which is nothing.

But here’s what most productivity advice gets wrong: it tells you to overcome these forces with willpower, discipline, and yes, motivation.

That’s like trying to drive a car with the parking brake engaged. Sure, if you press the accelerator hard enough, you might move forward. But you’re burning fuel fighting friction that shouldn’t exist.

So what’s missing?

The answer lies in the fact that you don’t need more force.

You need to remove the brake.

What I Learned Studying Professional Writers

I’ve always had some form of aversion toward hustle culture’s obsession with motivation.

The Instagram quotes. The morning routine videos. The productivity gurus promising that if you just find your “why” deeply enough, action becomes effortless.

It wasn’t until I started studying how professional creators/writers actually work that I understood why this advice never worked for me.

Ernest Hemingway wrote every morning, whether he felt inspired or not. He stopped mid-sentence so starting the next day required minimal activation energy.

Haruki Murakami runs before writing, using physical routine to bypass his brain’s negotiation phase entirely.

Steven Pressfield calls resistance “the most toxic force on the planet” and treats showing up to his desk like a soldier reporting for duty, regardless of how he feels.

These aren’t people with superhuman motivation.

They’re people who’ve systematically eliminated the need for it.

Here’s what I discovered when I reverse engineered their approaches: they don’t fight resistance. They don’t try to overpower it with inspiration or willpower.

They’ve built systems that make resistance irrelevant.

Hemingway’s mid-sentence stop? That’s activation energy reduction. He’s not starting from zero each morning. He’s completing a sentence, which requires almost no energy.

Murakami’s physical routine? That’s psychological reactance management. He’s not making a decision about whether to write. He’s executing a sequence that includes writing.

Pressfield’s militaristic approach? That’s uncertainty elimination. There’s no question about IF he writes or WHEN he writes. Those decisions don’t exist.

The secret they share: they’ve identified their specific resistance patterns and built countermeasures.

Not motivation. Not discipline. Not inspiration.

Strategic friction removal.

And here’s the uncomfortable truth: the reason you haven’t done this is because it requires honest self-diagnosis. It’s easier to watch a motivation video and feel temporarily energized than to map the exact psychological mechanisms preventing you from starting.

But that temporary energy is expensive.

It’s finite. It depletes. And every time you burn it to overcome resistance, you’re training your brain that action requires massive force.

With that, the best approach to consistent action does not come from finding better reasons to start.

It comes from removing the reasons not to.

I used to think I just lacked discipline. Turns out I was just making everything harder than it needed to be.

What’s Actually Happening in Your Brain

There are a few moments in my life that I remember vividly.

One was sitting in my car in a parking lot, needing to make a phone call I’d been avoiding for three weeks. The call would take five minutes. I had the person’s number. I knew exactly what to say.

But I sat there for forty minutes, cycling through elaborate justifications for “not right now.”

That experience taught me something: resistance isn’t rational. It’s neurological.

The Anterior Mid-Cingulate Cortex

Your brain has a region called the anterior mid-cingulate cortex (aMCC). Think of it as your brain’s “difficulty sensor.”

It activates when you face challenges, make difficult decisions, or push through discomfort. Research suggests this region may show structural changes in people who regularly do hard things

This region of our brain doesn’t distinguish between physical difficulty and psychological resistance. To your aMCC, the activation energy required to start a difficult email feels identical to the effort of climbing stairs.

Most people experience resistance and interpret it as: “This task is too hard right now. I need to be more prepared.”

But that sensation isn’t proportional to actual difficulty.

It’s proportional to perceived friction.

The Dopamine Prediction Error

Your brain runs on a neurotransmitter called dopamine, which most people misunderstand completely. They think it’s about pleasure or reward.

It’s not.

Dopamine is about prediction. Your brain releases it when reality exceeds expectation. When you expect nothing and get something. When you predict difficulty and experience ease.

Every time you scroll social media instead of starting your task, your brain gets a tiny dopamine hit.

Expected effort (zero) matched by actual effort (zero). Perfect prediction.

Every time you reorganize your desk instead of beginning your project, same thing. Low effort predicted, low effort experienced. Dopamine confirms your choice.

But when you finally force yourself to start the task you’ve been avoiding? Your brain predicted massive difficulty. And sometimes, once you begin, it’s not as bad as you thought.

Dopamine should flood your system, right?

Except it doesn’t work that way when you’ve been avoiding something. Because you’ve trained your brain that THIS specific task requires enormous activation energy. The prediction error is too large. Your brain doesn’t trust it.

The Salience Network Hijacking

Your brain has a network called the salience network that determines what deserves your attention. It’s supposed to highlight important, relevant information and filter out noise.

But here’s what happens: every time you avoid a task, your salience network marks it as threatening. Not important. Threatening.

The next time you consider starting it, your network triggers the same neural pattern as encountering a predator. Your attention gets hijacked. You suddenly notice everything else that seems more urgent, more safe, more immediately rewarding.

You’re not weak, you’re not undisciplined. Your brain has been trained to perceive your most important work as a threat.

And motivation is the emergency override system you’re using to temporarily shut down that threat response.

What I learned over the years is that you can retrain these systems. Not through willpower. Through systematic resistance reduction.

The Friction Elimination Protocol

You need a system.

Not a motivation ritual. Not a morning routine. Not a productivity hack. They never worked for me.

A diagnostic and elimination protocol for the specific resistance patterns sabotaging your action.

I start every meaningful project by identifying friction points before they become obstacles. Most people wait until they’re stuck, then wonder why they can’t push through.

Step 1: The Resistance Mapping Exercise

Become brutally aware of three things:

What are you avoiding right now? List every task, project, or action you know you should take but haven’t. Be specific. “Work on business” is useless. “Write the first draft of the client proposal” is useful.

What story are you telling yourself about why? For each avoided task, write the exact thought you have when you consider starting it. “I need to be more prepared.” “I don’t have enough time to do it properly.” “I should wait until I’m less tired.”

What would make it stupid-easy to start? For each task, identify the smallest possible first action. Not the full task. Just the entry point. “Open the document” not “Write the proposal.”

Keep in mind that you’re not analyzing tasks. You’re analyzing your psychological response to tasks.

The task itself might be objectively simple. But if your brain has marked it as high-friction, it doesn’t matter.

Include the “anti-resistance” exercise: Write out every single reason your brain generates for not starting. Every excuse. Every justification.

Every “not yet.” Get them on paper where you can see them as the patterns they are, not the truths they pretend to be.

Step 2: The Activation Energy Audit

Your brain treats decision-making as effort. Every choice about how to start, when to start, or what to do first burns cognitive energy before you’ve done actual work.

The solution: eliminate decisions before you need to make them.

For your highest-friction tasks, create a pre-decision protocol:

• Exactly when you’ll start (time-based trigger, not feeling-based)

• Exactly where you’ll work (same place every time removes spatial decision)

• Exactly what the first action is (open laptop, open document, write one sentence)

• Exactly how long the minimum session lasts (15 minutes works, “until it’s done” doesn’t)

This isn’t about discipline. It’s about removing the micro-decisions that give your brain opportunities to negotiate.

Step 3: The Certainty Builder

Your brain resists uncertainty more than difficulty. If you’re stuck, it’s often because you don’t know what “done” looks like.

The best way to eliminate uncertainty paralysis is to define completion before you start.

For every avoided task:

• What does “finished” look like specifically? (Not “better” or “good enough”, actual completion criteria)

• What’s the absolute minimum that counts? (The floor, not the ceiling)

• What information do you actually need vs. what you’re using research as procrastination?

Then start with the minimum. Not because you’re aiming low. Because your brain needs proof that completion is possible.

Once you finish the minimum, you’ll have momentum. Once you have momentum, your brain recalculates the effort prediction. Once the prediction drops, resistance drops.

The minimum isn’t the goal. It’s the mechanism for bypassing your brain’s threat response.

Step 4: The Environment Forcing Function

Every environment you’re in is either reducing friction or creating it. Your phone is nearby. Your desk is cluttered with twelve half-finished projects. Your comfortable chair invites distraction.

If your environment requires willpower to stay focused, you’re going to lose that battle eventually.

Environmental design should make the right action easier than the wrong action.

For deep work:

• Phone in different room (not just silent, physically separated)

• Single project visible (everything else closed or out of sight)

• Uncomfortable enough that you want to finish and leave (not punishing, but not luxurious)

Note: Giving up my phone was hard for me and sometimes it’s still is, so I bought myself a plastic box with a timer. I’d lock my phone inside, set it for an hour, and have to wait out the full hour to get it back. It was the only way I could limit the constant checking.

For avoided tasks:

• Pre-open the exact file/tool/document you need

• Close every other application

• Set a timer for the minimum session you defined

That is the only piece of productivity advice you need: make starting require less effort than avoiding.

Our anti-resistance frame is composed of:

• Mapped psychological patterns (you know what you’re fighting)

• Eliminated decisions (no negotiation opportunities)

• Defined completion (no uncertainty paralysis)

• Engineered environment (friction works for you, not against you)

This creates a tight feedback loop that encourages your brain to recalculate effort predictions, which reduces activation energy, which makes starting easier, which proves to your brain that resistance was disproportionate.

That’s the foundation. That’s what eliminates the need for motivation.

Why Resistance is Actually Information

You’re supposed to feel hesitation before starting something important.

You’re supposed to experience doubt when facing uncertainty.

You’re supposed to notice the gap between where you are and where you’re trying to go.

The problem isn’t that you feel resistance. The problem is that you’ve been taught to interpret resistance as evidence that you’re not ready, not capable, or not motivated enough.

What in the world did you expect to happen when you decided to do something that actually matters to you?

Did you think your brain, optimized for 200,000 years to conserve energy and avoid uncertainty, would just cooperate? That you’d wake up inspired every day, resistance-free, flowing effortlessly toward your goals?

The most successful people don’t flinch at resistance.

They expect it. They recognize it as confirmation they’re attempting something difficult enough to be worthwhile.

There are levels to how you engage with action:

Level 1 is waiting until you feel motivated. Letting your emotional state determine your behavior. This guarantees inconsistency because emotions are variable by design.

Level 2 is using discipline to override motivation. Forcing yourself to act despite not feeling like it. This works, but it’s energetically expensive. You’re spending willpower to fight your own resistance.

Level 3 is removing the need for both motivation and discipline by eliminating the resistance that made them necessary.

Many people are stuck between Level 1 and Level 2, burning enormous energy trying to become more disciplined, more motivated, more consistent.

They never consider that the game is removing friction, not increasing force.

Think of your brain like a self-driving car learning to navigate. For months, maybe years, it has received input that certain tasks require massive effort. Every time you’ve hesitated, procrastinated, or needed motivation to start, you’ve reinforced that pattern.

But patterns can be retrained.

Every time you start despite resistance and discover it wasn’t as bad as predicted, you’re updating the model. Every time you reduce activation energy and make starting easier, you’re teaching your brain a new pattern.

Eventually, the tasks that once required motivation become automatic. Not because you’ve become more disciplined, but because your brain has recalculated the effort required and stopped generating resistance.

I wish I could tell you that everyone reading this can eliminate all resistance immediately.

But if that were the case, the ability to act consistently would lose all meaning.

You must go through the process of mapping your resistance, eliminating your friction, and proving to your brain that action doesn’t require the force it’s been predicting.

The difference between people who act and people who wait for motivation isn’t internal strength.

It’s systematic friction removal.

And you already have everything you need to begin.

I still catch myself sometimes, sitting at my desk at work or home, feeling that old familiar weight. The sense that I need to be “ready” before I start. That I should wait until I feel different.

Then I remember that I’m not waiting for a feeling. I’m looking for friction I haven’t removed yet.

Articles you shouldn’t miss:

Not whether you’ll be employed, but whether the work you do will still mean something.

Last week, you spent three hours writing a campaign brief. You saw a colleague generate something 80% as good in four minutes using an AI agent (Claude, Gemini, ChatGPT…). Maybe 90% as good if you’re being honest.

You still have your job. But you can feel it shrinking around you.

The problem isn’t that the robots are coming. It’s that you don’t know what you’re supposed to be good at anymore. That Excel expertise you built over five years? Automated. Your ability to research competitors and synthesize findings? There’s an agent for that. Your skill at writing clear project updates? Gone.

You’re losing your professional identity faster than you can rebuild it. And nobody’s telling you what comes next.

The Three Things Everyone Tries That Don’t Actually Work

When you feel your value eroding, you do what seems rational. You adapt, you learn, and you try to stay relevant.

First, you learn to use the AI tools better. You take courses on prompt engineering. You master ChatGPT, Claude, whatever new platform launches next week and the week after. You become the “AI person” on your team. You think that if I can’t beat them, I’ll use them better than anyone else.

This fails because you’re still competing on execution speed. You’re just a faster horse. And execution is exactly what’s being commoditized. Six months from now, the tools will be easier to use. Your “expertise” in prompting becomes worthless the moment the interface improves. You’ve learned to use the shovel better, but the backhoe is coming anyway.

Second, you double down on your existing expertise. The accountant learns more advanced tax code. The designer masters more software. The analyst builds more complex models. You will have the same thought as many others, “I’ll go so deep they can’t replace me.”

This fails because depth in a disappearing domain is a trap. You’re building a fortress in a flood zone. Agents aren’t just matching human expertise at the median level anymore. They’re rapidly approaching expert-level performance in narrow domains. Your specialized knowledge becomes a liability because you’ve invested everything in something that’s actively being automated. You’re becoming the world’s best telegraph operator in 1995.

Third, you try to “stay human” through soft skills. You lean into creativity, empathy, relationship building. You go to workshops on emotional intelligence. You focus on being irreplaceably human. You might think that what makes us human can’t be automated.

This fails because it’s too vague to be actionable. What does “be creative” actually mean when an AI can generate 100 ideas in 10 seconds? How do you monetize empathy when your job is to produce reports? The advice feels right but provides no compass. You end up doing the same tasks you always did, just with more anxiety and a vaguer sense of purpose.

The real issue with all three approaches is that they’re reactions, not redesigns. You’re trying to adapt your old role to a new reality. What actually works is building an entirely new role that didn’t exist before.

But nobody’s teaching you what that looks like.

The Economic Logic Working Against You

This isn’t happening to you because you’re failing to adapt. It’s happening because the economic incentive structure is perfectly designed to create this problem.

The mechanism is simple, companies profit immediately from adopting AI agents. Every task automated results in cost reduction. The CFO sees the spreadsheet, where one AI subscription replaces 40% of a mid-level employee’s work. The math is simple, and the decision is obvious.

Many people hate to hear that. But if they owned the company or sat in leadership, they’d do the exact same thing. Companies exist to drive profit, just as employees work to drive higher salaries. That’s how the system has worked for centuries.

But companies don’t profit from retraining you for a higher-order role that doesn’t exist yet.

Why? Because that new role is undefined, unmeasured, and uncertain. You can’t put “figure out what humans should do now” on a quarterly earnings call. You can’t show ROI on “redesign work itself.” Short-term incentives win. Long-term strategy loses.

Nobody invests in the 12-24 month process of discovering what your new role should be because there’s no immediate return on that investment.

We’re in a speed mismatch. Agent capabilities are compounding at 6-12 month cycles. Human adaptation through traditional systems operates on 2-5 year cycles.

Universities can’t redesign curricula fast enough. They’re teaching skills that will be automated before students graduate. Companies can’t retrain fast enough. By the time they identify the new skills needed and build a program, the landscape has shifted again. You can’t pivot fast enough. Career transitions take time. Mortgages don’t wait.

We’ve never had to do this before.

Previous automation waves happened in manufacturing. You could see the factory floor. You could watch jobs disappear and new ones emerge. There was geographic and temporal separation.

This is different, knowledge work is being automated while you’re still at your desk. The old role and new role exist simultaneously in the same person, the same company, the same moment.

And nobody has an economic incentive to solve it. Companies maximize value through cost reduction, not workforce transformation. Educational institutions are too slow and too far removed from real-time market needs. Governments don’t understand the problem yet. You’re too busy trying to keep your current job to redesign your future one.

The system isn’t helping because it isn’t designed for continuous, rapid role evolution; it is designed for stability.

We’re using industrial-era institutions to solve an exponential-era problem. That’s why you feel stuck.

Your Experience Just Became Worthless (The Timeline)

Let me tell you a story of my friend, let’s call her Jane (Her real name is Katřina, but the Czech diacritic is tricky for many). She was a senior research analyst at a mid-sized consulting firm. Ten years of experience. Her job was provide answers to the client companies, who would ask questions like “What’s our competitor doing in the Asian market?” and she’d spend 2-3 weeks gathering data, reading reports, interviewing experts, synthesizing findings, and creating presentations.

She was good, clients loved her work, and she billed at $250 an hour.

The firm deployed an AI research agent (first version) in Q2 2024. Not to replace her, but as they said, to “augment” her. Management said all the right things about human-AI collaboration.

The agent could do Jane’s initial research in 90 minutes, it would scan thousands of sources, identify patterns, generate a first-draft report.

Month one: Jane was relieved and thought she could focus on high-value synthesis work. She’d take the agent’s output and refine it, add strategic insights, make it client-ready.

Month three: A partner asked her, “Why does this take you a week now? The AI gives us 80% of what we need in an hour. What’s the other 20% worth?”

Jane couldn’t answer clearly. Because sometimes the agent’s output only needed light editing. Sometimes her “strategic insights” were things the agent had already identified, just worded differently.

Month six: The firm restructured. They didn’t fire Jane, they changed her role to “Quality Reviewer.” She now oversaw the AI’s output for 6-8 projects simultaneously instead of owning 2-3 end to end.

Her title stayed the same. Her billing rate dropped to $150 an hour. Her ten years of experience felt worthless.

Jane tried everything. She took an AI prompt engineering course. She tried to go deeper into specialized research methodologies. She emphasized her client relationships. None of it mattered because the firm had already made the economic calculation.

One AI subscription costs $50 a month. Jane’s salary: $140K a year. The agent didn’t need to be perfect; it just needed to be 70% as good at 5% of the cost. But it was fast, faster than her.

The part that illustrates the systemic problem, you often hear from AI vendors that, thanks to their AI tools, people can focus on higher-value work. But when pressed on what that meant specifically, they’d go vague. Strategic thinking, client relationships, creative problem solving.

Nobody could define what higher-value work actually looked like in practice. Nobody could describe the new role. So they defaulted to the only thing they could measure: cost reduction.

Jane left six months later. The firm hired two junior analysts at $65K each to do what she did. With the AI, they’re 85% as effective as Jane was.

Jane’s still trying to figure out what she’s supposed to be good at. Last anyone heard, she’s thinking about leaving the industry entirely.

Stop Trying to Be Better at Your Current Job

The people who are winning aren’t trying to be better at their current job. They’re building new jobs that combine human judgment with agent capability.

Not becoming prompt engineers, not becoming AI experts. Becoming orchestrators who use agents to do what was previously impossible at their level.

Marcus (real name Zdeněk) was a marketing strategist at a retail company. When AI tools emerged, he didn’t try to write better marketing copy than the AI. He started running 50 campaign variations simultaneously. Something that would’ve required a team of 12 people before.

He’d use agents to generate the variations, test them, analyze results, and iterate. His job was to design the testing framework, analyze the patterns the agents uncovered, and make strategic decisions based on data no human could manually process.

Within six months, his campaigns were outperforming competitors by 40%. Not because he was better at any single task. Because he could operate at a scale that was previously impossible.

This is a pattern that works. Find the constraint in your domain that exists because of human limitations. What doesn’t get done because it takes too long? What questions don’t get asked because analysis is too expensive? What experiments don’t get run because you’d need a team of 20?

Then use agents to remove that constraint, not to speed up your current tasks. To do things that were previously impossible.

Then build expertise in the judgment layer. What experiments should we run? Which patterns matter? What do these results mean for strategy? When should we override the agent’s recommendation?

This isn’t vague strategic thinking. It’s specific, you’re the decision maker orchestrating a capability that didn’t exist before.

You’re not competing with the agent. You’re creating a new capability that requires both you and the agent. You’re not defensible because you’re better at the task. You’re defensible because you’ve built something that only exists with you orchestrating it.

This requires letting go of your identity as “the person who does X.” Marcus doesn’t write copy anymore. That bothered him at first. He liked writing. But he likes being valuable more.

Here’s what you can do this month:

Week one: Identify one thing in your job that you’d do 10x more if it didn’t take so long. Customer research? Competitive analysis? Testing variations? Data modeling?

Week two: Use AI agents to do that thing at 10x volume, even if quality drops to 70%. See what becomes possible.

Week three: Find the patterns. What insights emerge at scale that you’d never see doing it manually? What new questions can you answer?

Week four: Pitch this as a new capability to your boss. Not “I’m more efficient now.” But “We can now do this specific thing we couldn’t do before, which creates this specific business value.”

People who do this aren’t getting squeezed. They’re getting promoted or poached. Because they’ve made themselves the linchpin of a new capability, not the executor of an old task.

One critical caveat, this won’t work forever in its current form. Eventually, agents will get better at orchestration too. But it buys you three to five years. And in that time, you’ll see the next evolution coming.

The meta-skill is this: learning to spot what becomes possible when a constraint disappears, then building your value around that new possibility.

Most Strategic Thinking Was Actually Just Thoroughness

Many people currently doing “strategic” knowledge work aren’t actually that strategic.

When agents started handling the execution layer, everyone assumed humans would naturally move up to higher-order thinking. Strategy, judgment, and vision.

But a different reality is emerging, many senior people with years of experience can’t actually operate at that level. Their expertise was mostly pattern matching and process execution dressed up in strategic language.

“We thought Lisa was a strategic thinker because her analyses were thorough. Turns out the thoroughness was the skill. When an agent can be thorough in three minutes, we’re discovering Lisa doesn’t actually have strategic insights to add.”

This isn’t that these people are bad at their jobs. They were excellent at their jobs. The job required diligence, attention to detail, process mastery. They delivered exactly what was asked.

But the industry sold them on the idea that experience equals strategic capability. That putting in the hours would naturally develop judgment. For some people, it did. For many others, they got really good at execution and called it strategy.

Here is what one CEO (he didn’t want me share his name) of a mid-sized company in Canada told me: “We’re discovering that our senior people and our junior people are equally lost when we ask them what we should do, not just how to do it. The seniors are just more articulate about their uncertainty.“

The agent economy isn’t just automating tasks. It’s revealing who was coasting on the appearance of strategic thinking versus who actually possesses it.

And there’s no gentle way to tell someone, but you’ve spent 15 years building a career, and we’re just now realizing the thing you were good at wasn’t what we actually needed.

Nobody says this publicly because it suggests the problem isn’t just technological adaptation. It’s that our evaluation systems were broken all along. We promoted people for the wrong reasons. We confused “does the work well” with “thinks strategically about the work.”

Admitting that means admitting we don’t actually know how to identify or develop real strategic capability. We’ve been guessing, using credentials and years of experience as proxies.

The Only Durable Strategy Is Spotting What Just Became Possible

You’re not going to solve this by being better at your current job. That job is dissolving under you in real time.

You’re not going to solve it by learning the tools better. The tools will get easier to use without you.

You’re not going to solve it by going deeper into your specialty. That specialty is being automated.

What works is this: become the person who spots what just became possible and builds your value around that new capability. Use agents to remove constraints that previously limited what you could do. Become the orchestrator of scale that didn’t exist before.

This isn’t a permanent solution. In three to five years, you’ll need to do it again. The meta-skill is learning to continuously spot the next evolution and position yourself at the edge of what’s newly possible.

This will separate people who were genuinely strategic from people who were just thorough. There’s no way around that. The system that rewarded thoroughness is breaking down. The new system rewards the ability to see what constraints just disappeared and build something new in that space.

You still have time, but not much. The speed mismatch between agent capability and human adaptation is real. The companies won’t save you because they’re optimized for short-term cost reduction, not long-term workforce transformation. The educational system won’t save you because it’s too slow.

You have to save yourself. And the way you do that is by stopping trying to defend your current role and starting to build the role that didn’t exist six months ago.

Monday morning will keep coming. The question is whether you’re still wondering what you’re supposed to be good at, or whether you’ve already built the answer.

Article you shouldnt missed:

You sit down. You do the work. The same work that used to define strong performance.

And nothing happens.

Not because you got worse. Not because you stopped caring. But because the gap between effort and result quietly widened, and no one told you the standard moved.

Five years ago, a thoughtful email stood out. Today, your inbox is full of messages that look just as polished, many written or assisted by AI. What once felt personal now blends into noise.

Speed used to create advantage. Being first mattered because early meant visibility. Now information spreads instantly. Being fast no longer creates leverage. It simply keeps you from falling behind.

Volume lost its meaning too. Producing hundreds of outputs once signaled effort and drive. Today, automation tools can produce thousands. When activity becomes cheap, it stops being impressive. Your manager hears the numbers and shrugs, because motion is everywhere but progress is not.

You tried the obvious fixes. You doubled down on output. You bought better tools. You personalized harder. None of it restored the old baseline.

That is when the question starts to surface, usually late at night when you are still working: Am I losing my edge?

You are not.

The standard moved.

Doubling Down Makes the Problem Worse

The first instinct is volume. If the work is not landing, do more of it. Send more emails. Take more meetings. Process more requests. Push harder.

The logic makes sense. More attempts mean more chances to break through. If reply rates dropped, increase the denominator.

But volume increases noise, and noise lowers trust.

When you send more, you also send more average work, even if you try to maintain quality. People start to feel like they are part of a campaign, not a conversation. Response rates drop, so you compensate by sending even more. Now your calendar fills with follow ups. Your task list becomes a blur. And you spend less time on the few moves that actually change outcomes, like clarifying priorities, tightening decision criteria, or aligning stakeholders.

The system becomes a treadmill. More output, same or worse result.

What actually breaks is signal. You stop learning what works because everything becomes activity. You cannot see patterns when the pattern is chaos.

The second common move is tools. If execution feels hard, buy software that makes it easier. The pitch is appealing. Automate repetitive work. Get insights faster. Scale your impact.

The problem is that tools raise your throughput, but they also raise everyone else’s.

If the tool is widely available, it cannot be your edge. It just resets the baseline. You get faster at producing the same mediocre outcomes.

Tools, like AI, also create a false sense of progress. Dashboards fill up. Automation runs. Metrics look healthy. But the actual constraint is often human. Unclear goals. Slow decisions. Misaligned expectations. Weak follow through. The tool optimizes the wrong part of the system.

What breaks here is accountability. Teams start blaming the market, the platform, or the process instead of fixing root causes like vague priorities or delayed feedback loops.

The third trap is hyper personalization, often assisted by AI. If generic does not work, go specific. Reference details. Show you did your homework. Make it feel custom.

Surface personalization does not equal relevance.

Adding a line about someone’s recent project or alma mater is easy now, and people know it. It can even feel manipulative when it is hyper specific but still misses the point. The real driver of engagement is usually fit and timing, plus a clear reason why this matters to them right now.

When personalization focuses on flattery instead of substance, it reads as templated charm. People disengage or reply with skepticism.

What breaks is trust. You think you are increasing warmth. They experience it as scripted.

All three approaches share the same flaw. They try to outperform machines at machine work.

That is a losing game.

The Floor Rose for Everyone at Once

The real problem is not effort. It is leverage.

Five years ago, performance was limited by human speed. How fast you could research. How many drafts you could write. How thoroughly you could analyze. Effort scaled linearly. If you worked harder or were more skilled, you produced more output than others. That gap created advantage.

AI broke that relationship.

The system shifted from effort based leverage to tool based leverage. When a capability becomes automated, it stops being a differentiator and turns into infrastructure. Like email. Or spellcheck. Or search engines.

Once everyone has access, the advantage disappears.

This raised the floor for everyone at the same time. Not gradually. All at once.

Tasks that once required experience, training, and judgment can now be executed instantly at acceptable quality. Not perfect quality, but good enough. And in most systems, good enough is enough to compete.

The market no longer rewards execution alone. It rewards judgment about what to execute and why.

Here is what that looks like in practice.

Around 2019, a strong performer on a global team was known for one thing: response rate. She consistently hit 40 to 45 percent replies on outreach. That number mattered. Leaders tracked it weekly. High response rate meant strong communication skill. Low response rate triggered coaching.

It worked because writing good messages was hard. You had to understand context, read carefully, and phrase things like a human. Most people could not do that at scale, so the metric separated top performers from average ones.

Then automation and AI writing tools entered quietly.

Within a year, almost everyone’s messages looked better. Grammar improved. Tone improved. Structure improved. Reply rates briefly went up across the team.

Then something strange happened.

By the following quarter, reply rates dropped for everyone, including the best performers. Not a little. In some cases they fell by more than half. The same person went from 45 percent replies to under 20 percent, using messages that were objectively better written than before.

Nothing was wrong with her skill.

The signal collapsed.

People started assuming every message was automated. Even when it was not. They began scanning for relevance, not polish. If the context or timing was not right in the first two lines, they ignored it. Quality writing no longer bought attention.

The team reacted predictably. They rewrote templates. Added more personalization. Tested longer messages. Then shorter ones. They ran A/B tests. They added tools that promised smarter outreach.

None of it restored the old baseline.

Eventually leadership removed response rate as a primary performance metric. Not because it was unimportant, but because it no longer reflected individual skill. It reflected market saturation and tool access.

What replaced it was different. Speed to meaningful conversation. Conversion after first contact. Stakeholder satisfaction with outcomes. Fewer metrics, but closer to results.

That moment made the shift visible. A metric that once defined excellence became useless, not because people got worse, but because the system learned around it.

Competence Became Infrastructure

The standard moved because leverage moved.

That shift did not happen in isolation. It reflects a deeper change in how capability gets distributed.

For decades, work followed a predictable pattern. Skills were scarce. Training took time. Experience created advantage. Organizations competed by hiring people who could execute well.

Technology changed that gradually at first, then suddenly.

Email made communication instant. Spreadsheets made calculation accessible. Search engines made research democratic. Each wave automated a task that used to require skill, turning expertise into infrastructure.

AI accelerated that pattern.

The difference is scale and speed. Previous tools automated narrow tasks. AI automates entire categories of cognitive work. Writing. Analysis. Research. Design. Code. Not perfectly, but well enough to compete.

When competence becomes cheap, it stops being rewarded.

That is not a moral statement. It is structural.

Another force at play is asymmetry of scale. One person with AI can now operate at the output level of several people from five years ago. That compresses differentiation. When one person can do the work of three, volume stops signaling value. It only signals access to tools.

There is also signal inflation. When everyone can produce polished content, polished stops meaning competent. Clean deliverables, articulate communication, and fast responses used to signal care and skill. Now they mostly signal assistance. The signal still exists, but its credibility dropped.

At the same time, decision makers adapted. Audiences learned to filter. Managers learned to distrust surface quality. Leaders learned that activity metrics lie. When perception adjusts, standards rise automatically.

No one decided to raise the bar. The system did.

The counterintuitive truth is this: AI did not make work easier. It made basic competence cheaper. And when competence becomes cheap, it stops being rewarded.

That is why yesterday’s excellence feels invisible today.

Stop Competing on Execution

The people who still create separation did not get better at doing more. They changed where they apply judgment.

That is the new advantage.

Here are four adaptations that actually work, and why.

• First, move upstream closer to problem definition. Average performers focus on completing tasks. Strong performers shape them. They challenge vague briefs. They force clarity on priorities versus distractions. They pressure test assumptions, timelines, and constraints before work starts.
  
  This creates leverage because AI cannot fix a badly defined problem. If the goal is wrong, no amount of execution will save it. People who influence the input get advantage before the system even runs.
  
  In practice, this means asking harder questions earlier. What are we actually trying to solve? What would success look like if we had no constraints? What is the real blocker here? These questions feel slow at first, but they prevent rework later. And rework is where time dies.

• Second, optimize for decisions, not activity. Stop asking how much did we produce and start asking how fast can we get a real yes or no. Shorten feedback loops. Chase clarity harder than completion. Design processes that reduce waiting. This creates separation because most delays are internal, not external. AI cannot speed up indecision. Humans can.
  
  The shift here is subtle but critical. It means treating speed to resolution as more valuable than speed to output. A fast no is often better than a slow yes. Waiting costs more than most people calculate.

• Third, build credibility, not just reach. Top performers invest in being known before they need something. They show up consistently. They educate their audience. They create familiarity. When they make a request later, it does not start cold.
  
  This works because trust compounds while automation does not. AI can scale messages, but it cannot build reputation at the same depth. Familiarity lowers resistance before the conversation even begins.
  
  This does not mean posting constantly or networking aggressively. It means being useful over time. Sharing what you learn. Helping without asking for anything back. Making small deposits that add up.

• Fourth, measure what AI cannot fake. Track outcomes that require human alignment. Quality of decisions. Stakeholder satisfaction. Time lost to confusion or rework. Conversion rates that depend on trust. These metrics expose thinking, not typing. When you measure judgment, people improve judgment. When you measure activity, people game activity.

The pattern underneath all four is the same.

Execution got automated. Judgment did not!

People who keep trying to outperform machines at machine work lose. People who redesign the work around human judgment create separation again.

The Advantage Shifted, Not Disappeared

This is not a story about decline. It is a story about recalibration.

The skills that built careers five years ago were real. They mattered. They still matter in many contexts. But the system changed faster than most organizations adapted their expectations.

What used to be scarce is now abundant. What used to signal excellence now signals access. That does not mean effort is worthless. It means effort alone is no longer enough.

The gap you feel between working hard and getting results is not personal failure. It is structural friction. The old playbook stopped working because the game changed.

The good news is that judgment, context, and alignment cannot be automated yet. Maybe they never will be. Those are still human advantages. The people who win in this environment are the ones who stopped trying to be faster typists and started being better thinkers.

That shift is uncomfortable. It requires letting go of metrics and habits that once defined competence. It means accepting that some of what made you valuable before is now table stakes.

But it also means the ceiling is higher than it used to be. When execution becomes infrastructure, the leverage moves to strategy, prioritization, and influence. Those are harder to build, but they compound differently.

The standard has moved. That is real. But so is the opportunity to build an advantage in a new place.

You just have to stop looking for it where it used to be.