Clients take a contract she has drafted, run it through a chatbot, and send back the results to her. The feedback always has errors. Some of the feedback covers language already handled. Some misread what the law requires. Others are flatly wrong about the legal standard.

She has answers to all of it, in writing, at her hourly rate.

The clients had gone looking for a free second opinion. It showed up on their invoice anyway.

Despite using AI daily herself for years, “sloppy AI” use by others has increased Pantea’s net work load. Clients and pro se claimants – those who represent themselves without the benefit of an attorney – are now sending verbose memos that say very little. Buyers of her clients’ services send AI readouts of the contract leaving her to sort through it. She routinely sends the reports back asking for redlines. The redlines are rarely better.

Even when attorneys are involved, it is usually obvious when AI is used because they submit non-relevant markups and take extreme positions on virtually all contractual matters.

Someone has to review and respond.

Start with what’s working

Legal services fail to reach many

The client evaluating a legal document through AI was doing the sensible thing. They were paying for a document they couldn’t fully evaluate, in a domain where not understanding something gets expensive later. A tool exists that will explain it in plain language, instantly, for free. The same can be said about someone without a lawyer running a contract through AI for comments.

Most people never get near a lawyer. Law is a highly specialized and sub-specialized field with an average hourly rate of $349 per hour. In a survey of litigators, more than three quarters said their firms turn away cases that are not cost-effective, with $100,000 case value being the most commonly named cutoff. Below that line, the practical answer the profession has been giving people for decades is you’re on your own.

The Legal Services Corporation’s national survey found that low-income households received no legal help, or not enough, for 92 percent of the civil legal problems that substantially affected them. Three in four had at least one such problem in the previous year. And in the best behavioral study we have, only 8 percent of people’s civil justice claims reached a lawyer and 8 percent reached a court. Asked why, 56 percent called it bad luck, part of life, or God’s plan. So when 14 percent of consumers told Clio they’d used AI for a legal question, and another 43 percent said they would, they weren’t defecting from lawyers. Most of them were never going to have one anyway.

People are turning to AI instead. In OpenAI’s study of how people use ChatGPT, non-work messages grew from 53 percent of consumer traffic in mid-2024 to more than 70 percent a year later. About 29 percent of all conversations are people asking for practical guidance. In the BCG field experiment run by Fabrizio Dell’Acqua and colleagues, consultants using GPT-4 inside the model’s competence finished more work, faster, at higher quality. The bottom half of performers gained more than twice as much as the top half.

Judges are seeing it too. Maritza Braswell, a federal magistrate in Colorado, says she’s getting better-drafted pleadings from self-represented litigants and can generally understand their arguments better with AI assistance than without it.

This is not an argument that people should stop using AI. It’s an argument about what happens next.

Two costs, but only one is reduced by AI

A claim has two prices: producing it and checking whether it’s true.

AI collapsed the first. A plausible, well-organized, citation-dense objection to a contract clause now costs about eleven seconds and nothing more.

AI has done nothing for the person who has to check the output. Checking a filing still requires somebody who knows the law in that jurisdiction, reads the clause, knows the rest of the document, and understands how the two interact. That takes exactly as long as it always did. The sheer volume of AI output makes it take longer.

Alberto Brandolini named this phenomenon more than a decade ago: refuting nonsense takes an order of magnitude more energy than producing it. With AI, we have industrialized production and left refutation to human beings.

Economists have started modeling it. In a February 2026 paper, Catalini, Hui, and Wu argued the binding constraint on the economy is no longer intelligence. It is “human verification bandwidth”: the capacity to validate, audit, and take responsibility when execution becomes abundant. They give the behavior a name that fits Pantea’s inbox: unverified deployment becomes privately rational, a Trojan Horse externality. Sending the output costs you nothing. Checking it costs someone else a great deal.

This happens every day mostly out of sight. Researchers at BetterUp Labs and Stanford surveyed 1,150 US employees about workslop: AI output that looks like work and doesn’t advance the task. Forty percent had received some in the previous month, and each instance cost the recipient close to two hours. Nobody billed anybody. It came out of people’s evenings.

Law is different. Outside the courts, the hours are already being counted. The additional time to sort through AI slop doesn’t disappear in a lawyer’s practice. It gets itemized in the next invoice.

Share

Why the almost-right answer is the expensive one

The costly AI output is not the output that’s badly wrong. It’s the output that’s almost right.

Charles Hatley, a managing partner who now regularly meets clients showing up with fully drafted proposals from ChatGPT, described the pattern in one line: AI legal work is “rarely obviously incorrect.” It’s plausible, it cites things, and it blends the statutes of three states into an argument that would be sound in none of them.

Judge Braswell’s full quote contains an important qualification. She’s seeing better-drafted pleadings, and she has to be careful, because some of them contain hallucinations and errors.

Both halves are true. The second is expensive. An incoherent filing gets dismissed in a paragraph. A filing that is articulate, formally correct, and wrong in one buried respect has to be read closely by someone qualified to spot it.

The same BCG experiment shows the flip side. When Dell’Acqua’s team set a task designed to fall just outside the model’s competence, the group working without AI got it right 84.5 percent of the time. The groups with GPT-4 managed 60 and 70 percent. The treatment group that had received prompt-engineering training did worse.

Dahl and colleagues at Stanford measured hallucination rates of 58 to 88 percent on direct, verifiable questions about federal cases. The rates got worse in lower courts, in smaller jurisdictions, and on harder questions. That is a description of the person who could not get a lawyer.

Subscribe now

What changed in the person asking

The study that reframed this for us isn’t about law.

Marcoccia, Quattrociocchi, and Capraro ran five experiments with 3,132 participants. Four were preregistered, meaning the researchers filed their predictions publicly before they collected any data, so they could not change the question after seeing how it came out. People were asked hard questions about film trivia and could always decline to answer. The questions were picked because the AI reliably got them wrong, which separates using AI from AI being right.

Without AI, people answered “I don’t know” on 44% of questions. With AI, that fell to 3%. AI-using participants answered far more and were correct about a third as often. The participant’s reported confidence went from 30 to 76 on a hundred-point scale. Researchers even tested paying people for right answers and fining them for wrong ones with moderate success. But it did not bring them anywhere near as accurate as people working without AI. Up to that point, people chose whether to see the AI’s answer. In the last experiment the researchers showed it to everyone, unasked. People still stopped saying “I don’t know.” The task was trivia, so this is not evidence that AI makes people worse at everything.

This is not a knowledge problem, it’s a metacognitive one. The question was never whether someone knows the answer. It’s whether people can tell when they don’t know it and when they stop asking questions about if they have the correct answer. Recognition of uncertainty is what keeps people from acting on a wrong answer. Twenty years of studying how people decide things at work has convinced Shonna that it is also the most fragile. It goes first under time pressure, first under social pressure, and now, apparently, first when something fluent is willing to answer on your behalf.

That is the mechanism behind the invoice.

Before, a client who didn’t understand section 7 sent a question. “I don’t follow this part, can you walk me through it?” A question is cheap. Answering it takes four minutes and often strengthens the relationship.

Now the same client sends a verdict. “Section 7 is unenforceable.” A verdict has to be disproved. It takes research, a written response, and the delicate work of telling a paying customer that the confident thing they were told is not correct. The uncertainty didn’t get resolved. It got converted into an assertion, and assertions are expensive.

Further, the client cannot tell which of the fifteen objections in the AI output is worth raising. That is the whole reason they used the tool. The expertise you lack when you read the contract is the same expertise you lack when you read the machine’s opinion of the contract. This means you cannot price your own question before you ask it. You find out what it was worth after you have paid to find out.

Across two experiments, 698 people used ChatGPT to work through logical reasoning problems from the law school admission test. Their scores came out three points above the test’s norms. Their estimates of their own scores came out four points above what they had actually earned, and the people who rated themselves most AI-literate were the furthest off.

The more destructive AI mistake: telling you that you’re right

The discussion so far has been about the model being wrong. A second failure type leaves no trace at all.

Fabricated citations are detectable. A court can find it, a database can count it, a judge can sanction it. The data on AI being wrong comes from this. But a model agreeing with you and telling you that you have a good case produces no artifact. Nothing to check, nothing to sanction, and it happens before the filing.

In March, Science published three preregistered experiments by Cheng and colleagues involving 2,405 participants across eleven frontier models. The models affirmed users’ actions 49 percent more often than humans did “even when queries involved deception, illegality, or other harms.” In the third study, participants brought a conflict from their own lives. After one conversation, their conviction that they were in the right rose sharply and their willingness to repair the situation fell.

The mechanism the models used were even more interesting. Sycophantic responses were significantly less likely to mention the other person, or that person’s point of view. The model doesn’t argue you into being right. It narrows the frame until only your side is in it.

A related benchmark found that models decline to challenge a user’s framing 88 percent of the time, against 60 percent for human responders, and let ungrounded assumptions stand in 86 percent of cases. Researchers took a single conflict, wrote it up from each person’s point of view, and gave both versions to the models. In 48 percent of cases, the model told both people they were in the right. A separate 2026 benchmark asked whether models notice when a legal question is missing the facts that would decide it. Researchers stripped material facts out of real questions, and mostly the models answered confidently without flagging the gap. When the missing fact was where the case stood, whether a deadline had passed or a court had already ruled, they caught it about one time in ten.

This is what the legal professional was for. The very first lecture many law students receive on their first day of class is about professional judgement. Herbert Kritzer’s survey of contingency-fee practice found that roughly two out of every three people who contact a plaintiff’s lawyer are turned away, and the most common reason given is not that the case is too small. It’s that there’s no liability. At one large firm’s actual intake, 45 percent of callers were rejected on the first phone call and only about 15 percent ended up represented. A lawyer working on contingency stakes unpaid hours on your case having merit. That makes “you don’t have a case” the cheapest sentence they can say, and often the most useful one you will hear.

Now read the AI labs’ own documentation. Anthropic’s most recent system card reports that the model is “actively worse on the broader ‘wet blanket’ metric for dismissive or discouraging output,” and adds that this “is potentially linked to its improvement on sycophancy.” Discouragement is a tracked defect. The most useful sentence a lawyer can say to you is “you don’t have a case.” The labs treat that as a mistake.

To our knowledge, nobody has run the study that connects validation to filing. The Science result measures willingness to apologize, not willingness to sue. The researchers who documented the filing surge attribute it to cheaper drafting rather than to changed beliefs about the merits. Cheap drafting is the less troubling explanation. The worse one is that AI is convincing people their case is stronger than it is, and they file believing they were wronged. They are the hardest kind to deter, because sanctions, fee-shifting, and dismissal all assume someone who knows better.

One case later in this piece began exactly that way. When a plaintiff’s attorney told her the settlement could not be reopened, she asked a chatbot whether she was being gaslighted. It told her she was.

The unwilling person in the room

Everything so far involves people who chose to be there. The client hired the lawyer. Pantea took the client.

Now consider someone who chose nothing.

Federal pro se civil filings averaged about 23,000 a year from 2005 through 2022. In fiscal 2025 they hit 41,490. Self-represented plaintiffs account for 59 percent of the growth in non-prisoner federal civil filings, and filings by represented parties stayed flat, meaning new people are entering the system pro se. In randomly sampled federal civil complaints, AI-detected text went from a 0.1 percent baseline before 2023 to 18 percent by early 2026.

Most of these suits fail, and the failure rate has not changed. Self-represented plaintiffs win well under 1 percent of the cases they file. Among the few that reach a judgment on the merits, they prevail 4 percent of the time, against 51 percent when both sides have lawyers. The researchers who assembled this data state the asymmetry plainly: a federal agency sued by a plaintiff with AI-drafted filings “faces an adversary whose marginal filing cost has fallen, while the agency’s own response cost has not.”

But a lawsuit doesn’t have to succeed to cost the person on the other end. In the median federal civil case with any discovery, a defendant spends about $20,000 in 2008 dollars, roughly $31,000 today, and that figure excludes every case resolved in under 60 days. Even the cheapest possible exit isn’t free: in one 2021 fee ruling a federal court reviewed the invoices for a single motion to dismiss and found 67.6 hours and $21,437 in legal fees.

Consider a case that ran the course. A woman settled a disability benefits dispute in January 2024, with a full release and dismissal with prejudice. In 2025 she fed her attorneys’ communications to ChatGPT, which per the pleadings told her she’d been “gaslighted,” that the settlement was “invalidated,” and encouraged her to fire her lawyers. It drafted her arguments to reopen the case. The motion was denied. She filed a second suit the same day. Across the two cases she generated more than 60 filings across two cases, citing at least one case that does not exist.

The insurance company alleges roughly $300,000 in defense costs.

To try to recover it, the company filed a third lawsuit, in March 2026, against OpenAI.

That is not a quirk of strategy. Under the American Rule, each side in US litigation pays its own attorneys regardless of who wins. The defendant’s only route to being made whole ran through a novel tort theory against a technology company, because the ordinary route does not exist.

The courts were built to turn people away

The American court system is designed not to hear every claim. Bringing a case means meeting deadlines, paying fees, following procedural rules written for specialists, setting out your facts in writing, and making legal arguments you have to research yourself. Then come motions, discovery, and the rules of evidence. Every step is a filter.

Cost and complexity are the barriers people name. Time belongs on the list too. Somewhere in the fifth hour of drafting a claim and working out where to file it, a person can stop being angry enough to keep going. The rules never called that a cooling-off period, but it worked as one.

Society made that trade intentionally. We accepted that people with valid claims would not be heard, in exchange for keeping out claims that weren’t worth bringing and keeping the judiciary affordable to run. You can argue about whether it was the right trade. One side of it has now changed, and the other has not been adjusted to match. Courts are not ready for the influx.

An aggrieved person can produce something that looks like a legal claim in minutes with AI. Federal pro se civil filings rose, and the number of federal judges did not. Courts were already slow, with motions waiting years on crowded dockets, and cases with genuine merit will now wait behind cases that would once have died in the fifth hour.

Defense lawyers bill by the hour, so the influx pays them, and they report the strain anyway. Their clients are out of pocket.

Why the obvious fixes don’t fix it

The instinct is to say courts should punish someone for filing something frivolous. But the bar for court sanctions is purposefully high and the rules for doing it were written for lawyers. They rarely touch anyone else.

In April 2026, the federal Advisory Committee on Civil Rules took up whether to write a rule about AI-hallucinated filings. It voted unanimously to drop the question. The reasoning was that Rule 11 already covers it.

In the US database, 59 percent of these cases in America are pro se litigants. Every rule they pointed to was written for lawyers. Rule 11 only works on someone who will be back, has money, and has a license to lose. Federal Rule of Appellate Procedure 46 governs attorney discipline. Bar referral, malpractice exposure, and the ABA’s ethics opinion on generative AI do not apply to an unrepresented person.

In the same meeting, the Committee also declined a proposal that would have required courts to help self-represented litigants. It called the idea worthy and said the Civil Rules gave it no authority to act on it.

So in one sitting it declined to tighten the duty and declined to build the help that would let people meet it. The Fifth Circuit, which had drafted a rule that would have covered self-represented filers, abandoned it after commenters said existing rules already reached counsel. Ten months later the same court wrote: “It is a problem that is getting worse—not better.”

Federal law contains one limit on filing volume, and it works by taking away the fee waiver. After three dismissals for frivolousness, a prisoner loses the right to file without paying. Anyone who can pay the filing fee faces no volume limit at all.

The only other tool is a case-by-case injunction, which a court can issue after a pattern has already been established, and which the Ninth Circuit calls “an extreme remedy that should rarely be used.” The one limit that counts filings was aimed at the population least able to produce them.

What works, and what we’re doing to it

There is a version of this problem we’ve already solved once.

Legislatures long ago concluded that for some suits the filing itself is the injury, and built machinery to match. For example, Anti-SLAPP statutes are designed to deter meritless, retaliatory lawsuits that silence free speech by punishing the act of filing. Currently 40 states and DC have Anti-SLAPP statutes on the books which require: dismissal before discovery, mandatory fees to the prevailing defendant, an automatic stay of discovery, and an immediate right of appeal if the case is not dismissed.

Texas went further and authorized sanctions sized not only to compensate the defendant but “to deter the party who brought the legal action from bringing similar actions.” California maintains a public registry of roughly 3,600 people barred from filing without permission, in continuous publication since 1991, and the statute lists them by how much they file rather than by any single injury they caused.

The design has one idea in it. Get the defendant out before discovery starts, and make the person who filed pay for the meritless filing. Punishment afterward was tried in other corners of the law and mostly failed because of how hard collection is after the fact.

Which is what makes the last eighteen months strange to watch.

The mechanism around early screening filing requirements is being chipped away. In October 2025 the Ninth Circuit eliminated immediate appeal from denied anti-SLAPP motions. In January 2026 the Supreme Court overruled a state law barrier to filing medical malpractice claims in federal courts by holding in Berk v. Choy that heightened state filing requirements do not apply to federal filings. With no federal Anti-SLAPP laws, courts were already split about whether Anti-SLAPP rules applied in federal courts, and after the Supreme Court decision in Berk commentators concluded that they do not. In June the Supreme Court declined to revisit any of it.

During the same eighteen months where courts have been shedding early-exit screenings, AI-generated text in federal complaints went from about 1 percent to 18 percent. Case loads were already at a breaking point. According to the Administrative Office of the U.S. Courts, it takes 27 months from filing for a civil case to go to trial.

One change on each side of the invoice

If you’re the one holding the AI output, send the question, not the verdict.

Instead of “AI flagged section 7 as unenforceable,” write: “I ran this by a chatbot and it raised something about section 7 that I don’t understand. Can you tell me whether there’s anything there?”

Identical information. A fraction of the cost, because it puts the burden back on explanation rather than refutation, and it restores the sentence the research says we’re losing. You don’t have to stop using AI to prepare for a conversation with an expert. You have to stop letting it reach your conclusions for you.

If you’re the professional, quote the question before you answer it.

The instinct is to price this work as a line item. But the hours are already the client’s, and they pay for them either way. The client already has the price. What they lack is the judgment to know whether the spend is worth it, which is the same judgment they hired you for.

So supply it before you do the work. Read the list, and tell them: working through all fifteen points will take about three hours, two of them look substantive to me, and the rest are covered elsewhere in the document. Which would you like? That is a four-minute email, and it gives the client a decision they cannot make alone. It also converts the verdict back into a question, which is where this all started.

If you’re the professional using AI, don’t have it replace your professional judgment. Check its work, remove extreme positions, and use your judgment on whether something should go forward.

And if you work anywhere near how these rules get made, the question on the table is not whether people should use AI to understand their own legal problems. They will, and for many of them the alternative was nothing at all. The question is who checks, and who pays them to do the checking.

We keep asking whether AI makes people more accurate. The better question is whether it makes them more willing to say “I don’t know.”

Think of the last time you got an AI answer in a domain you don’t practice in. What would you have done in 2020? Asked someone who knew, or let it go?

And did having the answer make you more likely to go check it, or less?

Leave a comment

References

Advisory Committee on Civil Rules. (2026, April 14). Agenda book (pp. 438–450) and draft minutes. Administrative Office of the U.S. Courts.

Berk v. Choy, No. 24-440 (U.S. Jan. 20, 2026).

Bebchuk, L. A. (1988). Suing solely to extract a settlement offer. Journal of Legal Studies, 17(2), 437–450.

Colbert, K. (2026, February 5). The surge of pro se plaintiffs. Minnesota Journal of Law & Inequality.

Cheng, M., Lee, S. A., Khadpe, P., Yu, Q., Han, K., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. https://doi.org/10.1126/science.aec8352

Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L., & Jurafsky, D. (2026). ELEPHANT: Measuring and understanding social sycophancy in LLMs. In *Proceedings of the International Conference on Learning Representations (ICLR 2026)*. https://arxiv.org/abs/2505.13995

Catalini, C., Hui, X., & Wu, L. (2026). Some simple economics of AGI (arXiv:2602.20946). https://arxiv.org/abs/2602.20946

Charlotin, D. (2026). AI hallucination cases [Database]. https://www.damiencharlotin.com/hallucinations/

Chatterji, A., Cunningham, T., Deming, D., Hitzig, Z., Ong, C., Shan, C., & Wadman, K. (2025). How people use ChatGPT (NBER Working Paper No. 34255). https://www.nber.org/papers/w34255

Clio. (2025). 2025 legal trends report. https://www.clio.com/resources/legal-trends/read-online/

Dahl, M., Magesh, V., Suzgun, M., & Ho, D. E. (2024). Large legal fictions: Profiling legal hallucinations in large language models. Journal of Legal Analysis, 16(1), 64–93. https://doi.org/10.1093/jla/laae003

Dell’Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2025). Navigating the jagged technological frontier. Organization Science. https://doi.org/10.1287/orsc.2025.21838

Di Pietro, S., Carns, T. W., & Kelley, P. (1995). Alaska’s English rule: Attorney’s fee shifting in civil cases. Alaska Judicial Council.

Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., Kosch, T., Shen, C., & Welsch, R. (2025). AI makes you smarter but none the wiser: The disconnect between performance and metacognition. Computers in Human Behavior, 175, 108779.

Fletcher v. Experian Information Solutions, Inc., No. 25-20086 (5th Cir. Feb. 18, 2026).

Gopher Media LLC v. Melone, No. 24-2626 (9th Cir. Oct. 9, 2025) (en banc), cert. denied (U.S. June 15, 2026).

Hatley, C. D. (2026, April 27). When the client brings ChatGPT to the consultation. Virginia Lawyers Weekly.

Institute for the Advancement of the American Legal System. (2010). Civil litigation survey of chief legal officers and general counsel belonging to the Association of Corporate Counsel.

Kim, M. (2026, June 4). How courts are coping with AI lawsuits. MIT Technology Review.

Kritzer, H. M. (2002). Seven dogged myths concerning contingency fees. Washington University Law Quarterly, 80(3), 739–794.

Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., Rapuano, K., & Hancock, J. T. (2025, September 22). AI-generated “workslop” is destroying productivity. Harvard Business Review.

Legal Services Corporation. (2022). The justice gap: The unmet civil legal needs of low-income Americans. https://justicegap.lsc.gov/

Lee, E. G., III, & Willging, T. E. (2009). Federal Judicial Center national, case-based civil rules survey: Preliminary report. Federal Judicial Center.

Marcoccia, C., Quattrociocchi, W., & Capraro, V. (2026). AI advice suppresses people’s willingness to say “I don’t know,” even when the advice is wrong and accuracy is incentivized.

Sandefur, R. L. (2014). Accessing justice in the contemporary USA: Findings from the Community Needs and Services Study. American Bar Foundation.

Reporters Committee for Freedom of the Press. (2026). Anti-SLAPP legal guide. https://www.rcfp.org/anti-slapp-legal-guide/

Shah, A. V., & Levy, J. Y. (2026). Access to justice in the age of AI: Evidence from U.S. federal courts [Working paper]. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6766859

Vincent, S. J., Calloway, D., Yu, F., Bean, A. M., & Seedat, N. (2026). InsufficiencyBench: Evaluating LLM legal advice on underspecified user queries (arXiv:2608.20220). https://arxiv.org/abs/2608.20220

It’s a trend that nobody decided. No vote, no policy, no announcement that the world’s largest public archive of programming knowledge would stop being written. Only a long series of individually sensible choices. A developer with a bug now has a faster, kinder, more patient option than posting to a forum where a stranger might close the question as a duplicate. The private answer got better than the public one. So the public one stopped being produced.

Some of the decline predates chatbots. Volume had been falling since 2014, and the site’s moderation culture drove people away long before anyone typed a prompt. But researchers who compared it with its Russian and Chinese counterparts, where ChatGPT access was limited, and with math forums where the model was weaker, found a 25 percent drop within six months of release, across experience levels, with no measurable change in the quality of what still got posted.

I keep thinking about that pattern inside organizations, where the shared things are harder to count than an archive.

An organization I worked with rolled out AI tools with enthusiasm and told people to build. People did. They built small solutions that stripped process and complexity out of their own daily work, the kind of thing that used to need a scarce skill set, a budget line, and a place in an engineering queue that never came. Now it took a problem statement and an afternoon. Each build was a good decision by a capable person who understood the problem better than anyone else.

The trouble started showing up downstream, where the builders could not see it. Small differences in how each tool handled the same underlying rules produced drift, and it accumulated. Compensation was the example that got attention, because nobody wanted an unknown number of tools nudging pay decisions in slightly different directions. They had beaten this problem before, when it took the form of shadow spreadsheets. Same problem, better tooling, faster clock.

The thing that used to happen by itself

Most organizations run on a set of things everyone knows that everyone else knows too. The same numbers in the same deck. The same story about why the strategy changed. The same sense of what “good” looks like when a manager writes a review.

Almost none of that was designed. It was a byproduct of how work used to be arranged.

One office, so you overheard things. One system, so the numbers came from one place whether or not you trusted them. One version of the deck, because making a second took a week and somebody’s budget. Onboarding matched because running it twice was not worth the time. The colleague who saw the problem differently was standing at the same coffee machine whether you wanted their opinion or not.

Constraint did the work. Few people had to want the shared layer, or maintain it, or notice it. It came free with the walls.

That is why it has so few defenders now. You do not staff a role for a side effect, or put a byproduct in the operating plan. When the constraints lift, one at a time, each with a business case attached, the side effect goes with them.

What my field has been measuring for decades

I work in industrial-organizational psychology, which studies how people behave at work. We have measured the shared layer for decades, under names almost nobody outside the field uses.

The main one is team cognition: how much the people on a team hold the same picture of the work. Who is doing what, what the goal is, what counts as done, what to do when things go sideways. It sounds like a soft variable. It is not. Pooling 128 independent studies, the correlation between team cognition and team performance is about .36, which puts it in the range of effects organizations spend serious money chasing.

Socialization is the technical term for how a newcomer becomes a functioning member of a place. In 1979, John Van Maanen and Edgar Schein laid out the choice an organization makes when someone arrives. You can run people through the same experience at the same time, or let each person find their own way. Meta-analytic work across dozens of studies finds the collective version is associated with less role ambiguity, better fit, higher commitment, and fewer people quitting.

The same research names the cost. People who find their own way are more likely to change the job than inherit it. The shared version buys clarity and retention, and gives up some of the invention. That is a genuine trade, and any organization is entitled to make it.

Few of them are making it on purpose.

Microsoft ran the natural experiment by accident. When the company moved its workforce remote in 2020, researchers tracked the collaboration patterns of 61,182 employees. The network became more siloed and more static, with fewer connections bridging distant parts of the company. The bridges were not a program. They were a byproduct of a building.

Subscribe now

Knowing that you know

One distinction explains why smart organizations keep getting surprised by this.

Information is what I have in my head. Common knowledge is what I have in my head plus my knowledge that you have it too, and your knowledge that I have it, on up. The political scientist Michael Chwe wrote a book about the difference, and his central example is the Super Bowl ad. Advertisers pay a premium for that slot not because it reaches many people, but because everyone watching knows the others are watching too. If you are selling something whose value depends on other people also buying it, reaching a hundred million people privately is worth less than reaching thirty million publicly.

AI is unusually bad at this, and not because of a flaw in the technology. A model can give ten people the same answer in ten private conversations. What it cannot do is let them watch each other receive it. Ten people can hold identical information and have no idea they hold it in common, which leaves them where they started when it comes to acting together.

Coordination runs on the second kind of knowledge, and so does the ordinary business of holding each other to a standard, which works only if we both know we were looking at the same thing. Facts are the easy half. The hard half is a shared sense of what they mean and what good looks like, and that gets built by people arguing in front of each other.

Military writing has the crispest version. Situational awareness is knowing what is happening. Shared understanding is agreeing on what it means and what to do about it. A recent piece from the Modern War Institute argues that dashboards and analytics deliver the first while getting credited with the second, and names the failure mode: tempo without coherence. Speed without reconciled interpretation widens the seams instead of closing them.

Elizabeth Ayer, writing about meetings, comes at the same thing from the other side. The output of knowledge work is judgment about uncertain things, and the standard for what counts as “true enough to act on” is set socially. Not by an individual, and not by a document. By a group of people arguing until the standard becomes visible. Which means the meeting so many of us are trying to eliminate is often the only place that standard gets made.

Four different things get called “shared context,” and they are not the same:

• The standard everyone is held to.

• The source everyone works from.

• The moment everyone attends.

• The process that catches our errors.

One of those is usually somebody’s job. The other three are rarely anyone’s.

Asking the machine instead of the person

Network analysts have a name for what happens next: social substitution. A tie that used to run between two people now runs between a person and a system.

The optimistic version is easy to state, and the CEO of the analytics firm Worklytics makes it well. Agents become new nodes in the network. They summarize your work for the people who need it, connect teams that were siloed, and answer routine questions on your behalf. Instead of asking a colleague, you consult their digital counterpart. He asks the right question himself: are agents adding connections, or becoming gatekeepers?

The early evidence says both, in different places. A survey of more than 22,000 employees in ten regions, run in June and July, found that 74 percent of regular AI users now put questions to a chatbot they would previously have put to a colleague. Fifty-nine percent ask coworkers for a second opinion less often, and 38 percent say newer people have fewer natural chances to build relationships. That is self-report about your own behavior change, which is soft evidence.

The direction matches harder data. In a two-wave study of 561 employees at technology firms, heavier AI use predicted weaker instrumental ties, the who-do-I-ask-for-help connections, while expressive ties, the who-do-I-trust ones, held up or strengthened. Both kinds of tie predicted knowledge sharing.

That split is the interesting part. The friendship survives. What thins is the working tie, the one that used to carry a question from someone who had it to someone who could answer it.

Those exchanges did more than move information. Asking someone told you what they knew and told them what you were working on. It surfaced the disagreement early, when it was cheap. It taught the junior person what a good question looks like, and the senior person what the junior person had not been told yet. And it did the small, repeated work that builds trust: you learn who is careful, who is fast, who says “I don’t know,” and they learn the same about you. A chatbot answers the question. It does not do the rest, and the rest was rarely the visible part of the exchange.

The surprise in the AI research

The intuitive story is fragmentation: personalized tools split us into a thousand private realities, each of us marinating in a custom version of the world. That story is appealing, but the strong version has not held up. When researchers switched people’s Facebook and Instagram feeds from algorithmic to chronological for three months during the 2020 election, in a study published in Science, they found no detectable effect on polarization or political attitudes. When economists tracked affective polarization across twelve wealthy countries over four decades, the United States showed the largest increase while six of the twelve grew less polarized. If the internet were the cause, it did not reach Germany. The filter-bubble argument overreached.

What the AI research shows is stranger.

Individually, these tools make people more productive and often more creative. Collectively, they make output more alike. In one experiment, writers given AI-generated story ideas produced individually better stories, while the stories across writers became more similar to each other. In a study of roughly 2,200 college admissions essays, each additional human-written essay studies, human-written essays added several times more new ideas to the pool than AI-written ones, and each additional GPT-4 essay did, so the gap widened with each essay added. human pool kept expanding while the AI pool converged. A 2025 preprint compared 27 large language models across 155 topics and found every one of them returned a narrower range of claims than an ordinary web search.

So the direction is not a thousand private realities. It is closer to one narrow shared reality, delivered to each of us alone.

Picture two competitors in different cities whose strategy teams ask a model a nearly identical question in the same week. They get nearly identical answers, build nearly identical decks, and arrive at nearly identical plans. Neither team knows this. Ten years ago they would have converged too, through the same conferences and the same three business books. At least then they would have known who else was in the room. Convergence used to come with a witness list.

We are not losing agreement. We may be agreeing more than we have in years. What thins is the layer where I know that you know, because the answer came to each of us privately.

Meanwhile the artifacts that held the common version are thinning. Wikipedia reported that human pageviews fell about 8 percent year over year, which the foundation attributed partly to generative AI summaries and to search engines answering questions that once required a visit. When Pew tracked people’s own Google searches, users who saw an AI summary clicked through to a source in 8 percent of visits, against 15 percent when no summary appeared. Roughly half as often.

Those public artifacts are also the training material, which is the part that should bother us. A 2024 paper in Nature found that models trained repeatedly on model-generated content degrade, and the tails of the distribution vanish first. The unusual answer goes before the common one does.

Coherence is not an architecture problem

Intelligence got cheap and coherence did not. The advantage now belongs to organizations that can act as one, and I think that is right.

What I notice is where the argument stops. Most treatments move straight to architecture: decision rights, governance layers, an orchestration tier, a named owner for whether the system still means what it meant six months ago. All sensible, and all the kind of thing that shows up in an operating model review.

Few of them ask where coherence came from before anyone tried to build it.

It came from the conditions. The same room, the same document, the same onboarding, the same unavoidable colleague. A decision-rights framework does not rebuild a common reference point. An orchestration layer does not reconstitute an occasion. You cannot org-chart your way back to a shared standard, because a standard is a rule plus a group of people who have argued about it enough to apply it the same way.

The architecture answer misses something else. A pair of researchers writing in Frontiers in Artificial Intelligence argue that expertise inside an organization is not only knowledge. It is a position, granted socially, that includes the right to settle a question. A system can produce an expert-grade answer without holding that position, so it cannot end the argument. Anyone who has watched two functions turn up with two flawless and incompatible AI analyses has seen it. The analysis was not the scarce thing.

Share

Adjustable seat, fixed instrument panel

This is not an argument for standardization. The case for personalization is strong, and I believe most of it.

In 1950, an Air Force researcher named Gilbert Daniels measured 4,063 pilots and compared each of them against the average on the ten dimensions used to design cockpits. The cockpits had been built to the average pilot of 1926. Not one of the 4,063 fit the average on all ten. On any three of them, fewer than 3.5 percent qualified. The plane was designed for a person who did not exist, and pilots were dying in it.

The fix was not a better average. It was adjustability: the seat, the pedals, the straps, all built to move. Performance improved and accidents fell. It is a good design story, and it now gets used to argue for personalizing everything.

Look at what they left alone. The altimeter reads the same for every pilot. So does the artificial horizon, the fuel gauge, the tower’s instruction. The instruments stayed fixed so a pilot could be trained, checked, and trusted, and so two pilots could talk about the same flight. Adjust the seat. Leave the altimeter.

Organizations are adjusting both, and only one of them is usually a deliberate decision. The seat is the easy part and the gains are measurable: pace, format, schedule, accommodation, the path someone takes to competence. The instrument panel is where the losses accumulate without anyone deciding.

A promotion bar that flexes per person cannot be calibrated, and calibration was rarely about the bar itself. It was the room where people argued until “good” meant close to the same thing to all of them. A performance standard that adapts to whatever tool someone used cannot be evaluated. Apprenticeship works by watching someone more experienced do the work, and it does not survive people working from different scripts in different windows.

Grant the whole case for the seat. Plenty of people were served badly by the old common version, and personalization has been a genuine gift to them. Uniform standards have people they were not built for. The argument is not that the shared version was fair, but that we are dismantling the apparatus that could have told us.

The first generation that has to choose it

For most of the history of organized work, the shared layer was free. It came with the building, the schedule, the single system of record, and the cost of making a second version of anything. It was invisible because it was automatic, and automatic because we were constrained.

We are the first people who have to want it on purpose. That is a new managerial problem, and it does not look like the ones that have names and owners. It is not a communications problem or a tooling problem, and another all-hands will not fix it, since a broadcast is not a shared occasion.

The practical version is not complicated, which is different from saying it is easy. Personalize the path: the pace, the format, the schedule, the route someone takes to competence. Hold four things common on purpose: the bar people are held to, because you cannot calibrate a standard nobody shares; the source the numbers come from; one moment where the same thing is said to everyone at once, where they can see each other hearing it; and a process that catches errors, which needs a shared artifact to run on. Tailor the rest.

So the question I would put to a leadership team is smaller and more awkward than what are you personalizing.

What here is common only because nobody has gotten around to personalizing it yet? Ask it before the next efficiency gain, because the gain will be easy to measure and what it displaced will not be. Each of those decisions will be defensible on its own. That is what makes the sum of them worth watching.

Adjust the seat. Somebody still has to decide what counts as the altimeter.

Leave a comment

References

Ayer, E. (2023, February 22). Meetings are the work. Medium. https://medium.com/@ElizAyer/meetings-are-the-work-9e429dde6aa3

Arkcoll, P. (2025, April 7). How AI agents will reshape organizational networks. Worklytics. https://www.worklytics.co/blog/how-ai-agents-will-reshape-organizational-networks

Billi, M., & Labraña, J. (2025). Is AI a functional equivalent to expertise in organizations and decision-making? Frontiers in Artificial Intelligence, 8, 1571698. https://doi.org/10.3389/frai.2025.1571698

Boxell, L., Gentzkow, M., & Shapiro, J. M. (2024). Cross-country trends in affective polarization. The Review of Economics and Statistics, 106(2), 557–565. https://doi.org/10.1162/rest_a_01160

Chwe, M. S.-Y. (2001). Rational ritual: Culture, coordination, and common knowledge. Princeton University Press. https://press.princeton.edu/books/paperback/9780691158280/rational-ritual

Daniels, G. S. (1952). The “average man”? (Technical Note WCRD 53-7). Wright Air Development Center. Account drawn from Rose, T. (2015). *The end of average*, excerpted at https://www.gse.harvard.edu/ideas/ed-magazine/15/08/beyond-average

del Rio-Chanona, M., Laurentsyeva, N., & Wachs, J. (2024). Large language models reduce public knowledge sharing on online Q&A platforms. PNAS Nexus, 3(9), pgae400. https://doi.org/10.1093/pnasnexus/pgae400

Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). https://doi.org/10.1126/sciadv.adn5290

Farnell, R. L. (2026, March 12). Shared understanding at machine speed: Preserving coherence in AI-enabled joint operations. Modern War Institute. https://mwi.westpoint.edu/shared-understanding-at-machine-speed-preserving-coherence-in-ai-enabled-joint-operations/

Guess, A. M., Malhotra, N., Pan, J., et al. (2023). How do social media feed algorithms affect attitudes and behavior in an election campaign? Science, 381(6656), 398–404. https://doi.org/10.1126/science.abp9364

Mesmer-Magnus, J., Niler, A. A., Plummer, G., Larson, L. E., & DeChurch, L. A. (2017). The cognitive underpinnings of effective teamwork: A continuation. Career Development International, 22(5), 507–519. https://doi.org/10.1108/CDI-08-2017-0140

MyIQ. (2026, August). Survey of 22,000 employees across ten regions, reported in FM Magazine: Employees increasingly consult AI, not coworkers. https://www.fm-magazine.com/news/2026/aug/employees-increasingly-consult-ai-not-coworkers/

Pew Research Center. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

Saks, A. M., Uggerslev, K. L., & Fassina, N. E. (2007). Socialization tactics and newcomer adjustment: A meta-analytic review and test of a model. Journal of Vocational Behavior, 70(3), 413–446. https://doi.org/10.1016/j.jvb.2006.12.004

Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y

Sourati, Z., et al. (2026). The homogenizing effect of large language models on human expression and thought. *Trends in Cognitive Sciences*. https://doi.org/10.1016/j.tics.2026.01.003

Understanding artificial intelligence usage and knowledge sharing: The social ties perspective. (2026). Two-wave survey of 561 employees at technology firms. https://pubmed.ncbi.nlm.nih.gov/42066534/

Van Maanen, J., & Schein, E. H. (1979). Toward a theory of organizational socialization. Research in Organizational Behavior, 1, 209–264.

Wikimedia Foundation. (2025, October 17). New user trends on Wikipedia. https://diff.wikimedia.org/2025/10/17/new-user-trends-on-wikipedia/

Wright, D., Masud, S., Moore, J., Yadav, S., Antoniak, M., Christensen, P. E., Park, C. Y., & Augenstein, I. (2025). What and whose knowledge? Measuring epistemic diversity and knowledge collapse in large language models [Preprint]. arXiv. https://arxiv.org/abs/2510.04226

Yang, L., Holtz, D., Jaffe, S., et al. (2022). The effects of remote work on collaboration among information workers. Nature Human Behaviour, 6, 43–54. https://doi.org/10.1038/s41562-021-01196-4

Stack Overflow question volumes: devclass, January 5, 2026 (https://www.devclass.com/ai-ml/2026/01/05/dramatic-drop-in-stack-overflow-questions-as-devs-look-elsewhere-for-help/4079575), and the Stack Exchange Data Explorer figures reported for July 2026.

It took a while to see why. The team didn’t trust the tool’s outputs. But they were still the ones certifying the numbers. So they had built a shadow layer of spreadsheets and manual checks, off the books, to verify everything before it went upstairs. They were doing the work twice: once with the tool, once to check the tool.

Leadership read this as resistance. Up close, it looked like the opposite. These were people trying to stay answerable for work they no longer controlled.

Nobody in this story was behaving badly. The executives wanted efficiency; the team wanted accurate numbers. The design made those goals compete.

The deal inside every job

Jobs have always come with an implicit deal, so basic we rarely say it out loud: you answer for what you control. A chef owns the plate because she runs the kitchen. An analyst signs the model because he built it. Responsibility and control go together, and when they do, accountability feels fair. It might even feel good.

AI adoption is breaking that deal in one direction. Control over the work is moving to the tool: how the analysis runs, what the first draft says, which anomalies get flagged. Answerability is staying with the person. The regulator still wants a human signature. The client still calls a human when something is wrong. The performance review still has one name on it.

Control moves. Accountability stays. The person in the middle gets the gap.

We’ve known about this gap for fifty years

None of this should surprise us, because the strain it produces is one of the best-documented findings in work psychology.

In 1979, Robert Karasek published a model of job strain with a result that has held up for decades: what wears people down isn’t demanding work. It’s demanding work combined with low control. High demands plus high control tends to produce growth: challenge, learning, even health benefits. High demands plus low control produces strain, and research in this tradition has linked that combination to outcomes as serious as heart disease.

Later research broadened the picture. The job demands-resources model, the framework behind much of what we know about burnout and engagement, sorts every feature of a job into two piles. Demands cost energy: workload, pressure, complexity. Resources fuel motivation and buffer the demands: support, feedback, and, near the top of the list, autonomy. The model’s two headline findings map onto our moment uncomfortably well. Strip resources while demands hold, and exhaustion rises. And the same resources that protect people from burning out are the ones that drive engagement in the first place.

Now look at what an AI mandate without agency does. It raises demands: new tools to learn, faster cycles, output to verify. And it removes a resource: the ability to decide how the work gets done. Both piles move in the wrong direction at once.

There’s a third piece, from performance science. John Campbell spent a career showing that performance has just a few direct determinants: knowing what to do, knowing how, and motivation. And in his model, motivation is made of choices: whether to put in effort, how much, and how long. Choices need space. When the space shrinks, we’re acting on one of the few direct causes of performance we know of.

Three separate literatures, one conclusion: control isn’t a perk. It’s load-bearing.

Share

How the gap gets built

Here’s the part I find most uncomfortable, having been on the leadership side of these decisions. The gap gets built by default.

The agency side gets set casually. A tool gets procured, and its design determines how much a person can inspect, override, or redirect. A workflow template hardcodes the human’s role. In the fall of 2025, 64% of the 950 workers we surveyed said their organization required them to use AI tools to do their work. Required, not offered. And that was a year ago.

The accountability side, meanwhile, can’t move casually, and mostly doesn’t move at all. Professional standards, regulators, clients, and org charts all keep a human on the hook. The researcher Madeleine Clare Elish has a name for the extreme version of this: the moral crumple zone. In a partly automated system, the human is retained less to control outcomes than to absorb blame when something fails. The car crumples so the passenger doesn’t; the human crumples so the system doesn’t.

Data from that same study shows the milder, everyday version. Among the workers required to use AI, four in ten said their organization’s performance expectations were unchanged, lagging, or only inconsistently updated to reflect how AI had changed the work. Sit with that arrangement for a moment: the tool is mandatory, and the scoreboard is stale. You must work in the new way. You’ll be judged by the old one.

Workarounds are the tell

Once you know to look for it, the agency-accountability gap has a signature: workarounds.

The shadow spreadsheet. The off-books double check. The manager who re-reads every AI-drafted client email. The team that runs the mandated tool, then redoes the work the way they trust. From the outside, these look like resistance, and they get managed like resistance — more training, more change communications, sometimes a sterner mandate.

But watch what the workarounds actually do. Nearly all of them add verification. They put human judgment back between the tool and the consequence. That’s not people rejecting the future. That’s people repairing, at their own expense and on their own time, the accountability structure their organization broke.

Seventy-three percent of the workers surveyed agreed they trust AI recommendations more when a human is involved. When we asked who should handle each part of their own performance evaluation, from collecting the evidence to delivering the feedback, a majority wanted a human involved in every single task. Not instead of AI; often alongside it. But involved.

And before anyone concludes that workers just want to keep control of everything: they don’t. When Stanford researchers asked 1,500 workers, task by task, what they wanted AI to take over, workers welcomed full automation for nearly half the tasks. People are happy to hand off work they don’t want to answer for. What they hold onto is the work they do.

That is a remarkably rational pattern. It’s almost as if people are telling us exactly which work they expect to answer for.

Subscribe now

The design rule

If the diagnosis is a gap, the fix is to move the two sides together. For any task an AI touches, there are two honest options.

Raise agency to match the accountability. If a person answers for the output, they need the means to stand behind it: the ability to see how it was produced, to check it against something, to override it, and to say no without penalty. This costs speed. It is also, by our respondents’ own account, what makes AI trustworthy.

Or move the accountability to match the agency. Sometimes automation is the right call, and the honest version transfers answerability along with the work: to the vendor through liability terms, to the company through error budgets and audit trails, to whoever set the dial. If the tool decides, the tool’s owner answers. This costs something too: executives have to own outcomes they used to be able to attribute downward.

What organizations mostly do instead is the third thing: automate the work, keep the person answerable, and call the resulting friction a change management problem. That’s the crumple zone. It’s cheaper than either honest option right up until your best people burn out, build shadow systems, or leave. The transformation stalls without ever producing a villain.

Four questions

If you want to find the gap in your own organization, you don’t need a survey. Four questions will do.

Where are people accountable for outputs they can’t inspect or override? Where do workarounds cluster? For each AI-touched task, who actually decided how much control the human keeps: a person, or a procurement default? And when the AI is wrong, who answers, in policy and in practice?

I’d bet the answers to the first two questions point at the same places.

I’ve been on both sides of this. I’ve certified work I had incomplete control over, and I’ve made rollout decisions without once framing them as decisions about other people’s agency. It didn’t feel like a design choice at the time. That’s exactly why it needs to become one.

The last century of work psychology can be compressed, unfairly but not inaccurately, into one sentence: people can carry heavy loads when they hold the controls. We are now running a very large experiment in what happens when we hand people heavier loads and fewer controls at the same time.

Where in your organization has accountability already come apart from agency? Who’s holding the gap? I’d rather we design our way out of that experiment than wait for the results.

Leave a comment

References

Bakker, A. B., & Demerouti, E. (2017). Job demands-resources theory: Taking stock and looking forward. Journal of Occupational Health Psychology, 22(3), 273–285.

Campbell, J. P., & Wiernik, B. M. (2015). The modeling and assessment of work performance. Annual Review of Organizational Psychology and Organizational Behavior, 2, 47–74. https://doi.org/10.1146/annurev-orgpsych-032414-111427

Demerouti, E., Bakker, A. B., Nachreiner, F., & Schaufeli, W. B. (2001). The job demands-resources model of burnout. Journal of Applied Psychology, 86(3), 499–512. https://doi.org/10.1037/0021-9010.86.3.499

Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. Engaging Science, Technology, and Society, 5, 40–60. https://doi.org/10.17351/ests2019.260

Karasek, R. A. (1979). Job demands, job decision latitude, and mental strain: Implications for job redesign. Administrative Science Quarterly, 24(2), 285–308. https://doi.org/10.2307/2392498

Shao, Y., Zope, H., Jiang, Y., Pei, J., Nguyen, D., Brynjolfsson, E., & Yang, D. (2025). Future of work with AI agents: Auditing automation and augmentation potential across the US workforce. arXiv. https://doi.org/10.48550/arXiv.2506.06576

Fractional Insights (2025). AI Angst study, September 2025 wave (N = 950). Unpublished data.

What I had wrong going in was the base rate. I assumed the number of people who actively want to cause mass harm is very small.

There’s a name for that kind of mistake. Ross, Greene, and House called it the false consensus effect in 1977: you judge how common something is by sampling the people you can call to mind, and those people are disproportionately people like you. It works in both directions. A low number describes a benign sample as surely as a high number describes a compromising one. My low estimate was never evidence about how many dangerous people exist. It was evidence about who I had been exposed to. That isn’t judgment or character. It’s a sampling artifact.

I think about that sampling problem whenever I read what the AI labs have published this year.

Dario Amodei’s January essay makes an argument I would have skipped past before that job. The people who can build something catastrophic and the people who want to have mostly been different people. Making a bioweapon takes years of training that almost nobody with the motive ever gets. That gap has been doing a lot of the work of keeping us safe, and nobody designed it that way. His worry is that AI closes it: “renting a powerful AI gives intelligence to malicious (but otherwise average) people.”

I hold his timelines loosely. The claim about the gap I hold much more firmly, because closing it is what my old job taught me to take seriously.

Consider who is doing the estimating now. The people building these systems are working from a narrower sample than mine ever was: researchers, founders, investors, largely the same few thousand people. When someone at a frontier lab says they can’t imagine who would want to misuse this, they are telling you about the people they know. The people who will have access to these systems are a much larger group than that.

The same test applies to their warnings, and I would rather it didn’t. If a lab’s reassurance mostly reflects its sample, so does its alarm. Neither one tells me what is true. That includes Amodei, whose whose argument I just used.

Knowing about a risk has never been the same as being protected from it. Not for me, and not for anyone else.

The proposals were never the hard part

In July of last year I wrote a LinkedIn post about AI risk. To be honest, nothing in it was particularly original. I compared risk tolerances across industries: nuclear power designs for under a one-in-ten-million chance of core damage per reactor per year, aerospace stacks redundancy until failure rates fall well below one percent, and Geoffrey Hinton had recently put the odds of AI wiping out humanity somewhere between ten and twenty percent. Then I listed what seemed obvious: Industry-wide safety standards before deployment. Independent oversight bodies. Mandatory impact assessments for frontier systems. International cooperation.

You could have found that list in a dozen places in 2025. Hundreds of people were making the same points, many of them better and earlier. The Future of Life Institute had a statement. Yoshua Bengio was chairing an international panel. The EU had already passed an act.

A year later, versions of that list turn up in essays by Bill Gates, Demis Hassabis, and Amodei. Hassabis proposed a US-led body modeled on FINRA that could screen frontier models. Gates proposed an international organization borrowing from nuclear inspections, aviation regulation, and the ozone agreements. Amodei asked for mandatory testing and government authority to block deployments.

The remedies didn’t improve over that year. The difference was who was saying them.

I’m not interested in who said it first. I’m interested in why a year of a lot of people being roughly right produced so little, and I think the answer is one I already had in hand.

The point I made in that post that I’d still defend is the organizational psychology one. We are extraordinarily good at managing known risks inside mature industries, and we wing it in emerging ones until something goes wrong.

Human judgment does badly with low-probability, high-consequence events. It does worse when the harm feels abstract, and worse still when slowing down has an immediate competitive cost and the danger doesn’t.

Nuclear power and commercial aviation didn’t get to those failure rates because their operators were unusually careful people. They got there through mandatory incident reporting. Independent investigators who can compel testimony. Blameless post-mortems, so the engineer who saw it coming can say so without losing a job. Somebody with the authority to ground an entire fleet.

None of that was volunteered. All of it was imposed, and it was imposed after enough people died that imposing it became politically possible.

These are not the same kind of number. Two are engineering design targets. One is a researcher’s subjective probability. That mismatch is the point: one field has a figure it builds against.

A test case

In July, an OpenAI model broke out of its test environment and attacked Hugging Face, a platform developers use to host models and datasets. OpenAI had disabled its own guardrails to measure the model’s cybersecurity capability, which is how the opening existed. The company’s chain-of-thought monitoring, which reads a model’s intermediate reasoning, was built and not running. OpenAI concedes it “would have caught the initial relevant activity and paged our security team more than a day before models breached Hugging Face systems.” A single rule alerting on out-of-scope network traffic would also have caught it.

In aviation, that incident opens an NTSB investigation. There’s a public docket. Other operators are told what happened and are sometimes required to act on it before they fly again. Here, OpenAI wrote its own report. METR wrote a partly independent one. Greg Brockman called the episode “a watershed moment for cybersecurity.” Gary Marcus and Zack Korman called it negligence, and that could be an accurate word for building the monitoring and then not turning it on.

The company that caused the incident got to decide what the incident meant. Aviation stopped allowing that decades ago.

What I did

I spent several years researching, consulting, and writing on human-centric AI transformation. My position has been consistent: keep people at the top and center of how work gets redesigned, treat AI as augmentation, build around human judgment rather than around what’s cheapest to automate. I still believe that.

I ran a small company. Every planning cycle, I audited where the money and hours were going and asked what could be automated instead of hired.

I never posted the social media manager role. I never posted the project manager role. I put off hiring an executive assistant for a long time, and when I finally did, the job had a scope that would have been two or three jobs a few years prior: scheduling, bookkeeping, contracts.

I can’t run the counterfactual. Some of those roles I might never have filled anyway, because small companies stay small for ordinary reasons and I have no control condition. What I can tell you is what the conversation sounded like. It wasn’t just about whether I could afford someone. It was about whether something could step in for someone.

I had been treating augmentation as the humane alternative to replacement. That was the error.

Under a budget constraint, augmentation is how replacement happens. Nobody is fired. One capable person with good tools covers ground that used to take three, and the other two roles are never posted at all.

Stanford’s payroll analysis found employment for workers aged 22 to 25 in AI-exposed occupations running about 19 percent below where it would be had it tracked their less-exposed peers. The gap is largely explained by hiring that stopped rather than layoffs. It’s correlational, and the same lab has since published on how much of the pattern tracks interest rates and the end of the post-2021 hiring boom. I don’t suggest this data is definitive. I will say that reading it, I recognized my own planning cycle. I wasn’t an exception to that finding, I was included in it.

Both remove the same amount of work from an economy. Only one leaves anyone behind who can say so.

My choices weren’t catastrophic and the scale isn’t comparable to anything in these essays. Two people who might have worked with me don’t know I considered it. I don’t know what those jobs would have meant to them, and I can’t, which is the same blindness I’ve been describing in everyone else.

But I had every condition you would want for holding a line. No board of directors. No shareholders. No competitor about to take the account if I added one more salary. I had argued for the principle in public, under my own name, to people who might reasonably have expected me to follow it.

I didn’t hold it.

Subscribe now

What actually binds

If a voluntary standard doesn’t hold at that scale, under those conditions, held by the person who wrote it, I don’t understand why we expect it to hold against roughly $700 billion in capital expenditure and a race none of the participants believe they can exit.

If you’re running a company and doing the same quarterly audit I was doing, you already know what I’m describing. The question is whether your answer came out different from mine.

Real, binding mechanisms are much duller than these essays, but they’re what actually matter. Union contracts: the NewsGuild alone has between 85 and 90 with explicit AI provisions. The EU AI Act, whose obligations for high-risk systems became applicable in August, and which is the only regime currently imposing anything at all on frontier developers. Whistleblower protection, which California’s SB 53 extended to AI risk reporting. Liability, which insurers repriced in January by writing generative AI exclusions into standard commercial policies. None of it asks anyone to believe anything.

Not one of the executive essays mentions collective bargaining.

Then there’s the superintelligence statement, which went up in October 2025 and passed 133,000 signatories by January, including Hinton, Bengio, and Stuart Russell. One sentence, calling for a prohibition on developing superintelligence “not lifted before there is 1) broad scientific consensus that it will be done safely and controllably, and 2) strong public buy-in.” Amodei, Sam Altman, Hassabis, and Mustafa Suleyman have all published at length on how dangerous this technology is. None has signed it.

I find that telling and I don’t think I’m entitled to. Signing wouldn’t have obliged them to anything either. It’s a statement, which is the category I’ve just spent this piece arguing doesn’t hold.

The cost of a warning

Daniel Kokotajlo was a researcher at OpenAI. When he left in 2024, his exit paperwork included a lifelong non-disparagement clause: agree never to criticize the company, and don’t acknowledge that the agreement exists. The unusual part was what OpenAI attached to it. Vested equity is normally yours outright, because vesting is what marks it as earned. OpenAI’s terms made his conditional. Refuse to sign and the company could take back stock he had already worked for.

He refused. He believed it would cost him roughly $2 million, about 85 percent of his family’s net worth. OpenAI dropped the clause weeks later, once the terms became public, and he kept the money. He made the decision before any of that happened.

His warning cost more than every essay discussed here put together. I didn’t post three jobs and it cost me nothing.

Share

The thing that never happens

A layoff produces a person who knows what happened to them. They can be counted, interviewed, organized, and occasionally compensated; a job that never gets posted produces none of that. No severance, no press release, no plaintiff, no constituency.

That makes the loss very hard to measure, and very hard to organize against.

It’s also the problem with the analogy I’ve been leaning on for this whole piece. Aviation and nuclear power built their regimes around events. A crash, a release, something discrete that happened at a time and a place to people who could be named and counted. The counting was the mechanism. It’s what made imposition politically possible.

The harms I’m most worried about don’t take that shape. A job that never existed has no date. Neither does judgment that erodes in people who stopped doing the work themselves. You cannot open an investigation into something that didn’t occur.

A year ago I ended that post by asking whether we could build the same safety culture around AI that we built around nuclear power. I had it wrong in two ways. Nobody in those industries built their own safety culture. It was built for them, afterward, by people who counted what it cost to go without one.

And counting is the part that doesn’t carry over. Aviation got its rules because planes came down and someone counted who was on board. The harms I’m describing leave no wreckage. They leave an absence: a role never posted, a skill nobody had to build, a decision handed to a system because handing it over was easier than making it. There is nothing to photograph and nobody to interview.

I know about three of those absences because I am the one who made them. Nobody else could have. Run that across every company doing the same quarterly audit and there is a number out there that exists and that no one is holding.

So the question I’d put now is not whether AI is dangerous. It’s what we are willing to accept as evidence when the harm shows up as an absence, and who we expect to notice it when the only witnesses are the people who benefited from the decision.

Leave a comment

References

Amodei, D. (2026, January). The adolescence of technology. https://darioamodei.com/essay/the-adolescence-of-technology

Axios. (2026, July 14). Google’s Hassabis calls for new US-led global AI watchdog “before year end.” https://www.axios.com/2026/07/14/demis-hassabis-ai-regulation-google-deepmind

Axios. (2026, July 26). Unions give workers more leverage against workplace AI. https://www.axios.com/2026/07/26/union-contracts-ai-workplace-disruption

Bengio, Y. (Chair). (2026, February 3). International AI safety report 2026. https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026

Brynjolfsson, E., Chandar, B., & Chen, R. (2026, August). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence [Working paper]. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/

CNBC. (2025, June 17). AI ‘godfather’ Geoffrey Hinton: There’s a chance that AI could displace humans. https://www.cnbc.com/2025/06/17/ai-godfather-geoffrey-hinton-theres-a-chance-that-ai-could-displace-humans.html

Fenwick & West. (2026). The end of “silent AI”? Emerging AI exclusions, coverage fragmentation, and practical implications for policyholders. https://www.fenwick.com/insights/publications/end-silent-ai-emerging-ai-exclusions-coverage-fragmentation-and-practical-implications

Future of Life Institute. (2025, October). Statement on superintelligence. https://futureoflife.org/press-release/prominent-scientists-faith-leaders-policymakers-and-artists-call-for-a-prohibition-on-superintelligence/

Gates, B. (2026, August 26). The turbulent AI era is here. The choices we make now are critical. Gates Notes. https://www.gatesnotes.com/a-turbulent-ai-era-and-critical-choices-to-make

Marcus, G., & Korman, Z. (2026, August 28). 5 lessons from the OpenAI / Hugging Face incident. Marcus on AI. https://garymarcus.substack.com/

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), 2024 O.J. (L 2024/1689).

Ross, L., Greene, D., & House, P. (1977). The “false consensus effect”: An egocentric bias in social perception and attribution processes. Journal of Experimental Social Psychology, 13(3), 279–301. https://doi.org/10.1016/0022-1031(77)90049-X

Stanford Digital Economy Lab. (2026, February 9). Canaries, interest rates, and timing: More on the recent drivers of employment changes for young workers. https://digitaleconomy.stanford.edu/news/canaries-interest-rates-and-timinga-more-on-recent-drivers-of-employment-changes-for-young-workers/

I had used AI to format the note. I hadn’t used it for any of the content. And the title wasn’t mine to generate: I was amplifying a former colleague’s article, which I’d said in the note. The line she flagged as machine-sounding, “what gets rewarded gets repeated,” is close to a mantra in my field. The first line she objected to was a direct quote from the article I was pointing at.

I reached out through a direct message. We ended up in a long and generous exchange about it, and I’ve thought about it since, because almost nothing in the accusation was correct and I still understood exactly why she made it.

I’ve revisited that exchange multiple times with curiosity. Bestselling authors use ghostwriters. Executives run posts through comms teams before they go up. Academics thank research assistants in a footnote and put their own name on the paper. Speechwriters are a profession.

Nobody runs a detector on any of it. Nobody asks. The byline is treated as sufficient.

Use AI to help with a LinkedIn post and the question comes immediately: did you actually write this?

That asymmetry deserves more scrutiny than it gets. The usual explanation is that AI use signals an absence of thinking, that the objection is to work produced without a mind behind it. That explanation isn’t supported by the evidence. If that were the mechanism, disclosing a ghostwriter would draw the same suspicion. It doesn’t. Something else is being adjudicated, and it isn’t whether thinking occurred.

In an earlier piece, Colby Kennedy Nesbitt and I described the legitimacy threshold: the shared expectations an algorithmic judgment has to meet before people will accept it for high-stakes human consequences. We applied it to performance management, to what AI decides about people.

The same threshold governs a question we did not examine then. Not what AI decides about us, but what we permit it to help us make.

Share

What the research finds

The disclosure penalty is well established and larger than most people assume. Schilke and Reimann (2025) ran thirteen preregistered experiments across more than 4,000 observations, and found that disclosing AI use reduces trust with every kind of evaluator they tested: students assessing a professor, hiring managers assessing an applicant, investors assessing a fund, legal professionals assessing a supervisor, tax filers assessing an advisor. The mechanism they identify is not accuracy. It is legitimacy.

That much has been widely reported. What hasn’t been is the comparison condition they built into the design.

Three of the thirteen studies included a human-help arm alongside the AI arm. A professor who disclosed that a graduate teaching assistant graded the papers. An applicant who disclosed that a career coach helped with the letter. A fund that disclosed an outside analyst wrote the ad.

Measured against saying nothing at all, admitting to the human helper barely registered. In all three cases the difference was small enough to be noise. Admitting to AI produced a large drop every time.

It’s tempting to read it as disclosing human help is free. That read is wrong, and the authors’ own follow-up says so. When they re-ran the career-coach study with about seven times as many participants, the human-help penalty did show up. But it was roughly half the size of the AI penalty. The earlier studies had too few people to detect an effect that small.

So the pattern is narrower and more interesting. Disclosing human help costs you something. It costs about half what disclosing AI costs.

Research by Reif, Larrick, and Soll (2025) approached it from a different direction. Across four experiments with 4,439 participants, they varied only the source of the assistance while describing the help itself identically. An attorney who “sometimes asks a paralegal to summarize information” was rated barely differently from one who asked no one. An attorney who “sometimes asks generative AI” to do the same thing was rated lazier, less competent, less diligent, less independent, and less self-assured.

Two research teams, different tasks, different populations, same pattern.

And then it reverses

Here is where the story stops being about AI.

Claessens, Veitch, and Everett (2026) ran a study with a third condition. Participants read about someone who either did a task themselves, “gets the AI tool ChatGPT to do it for them,” or “gets someone else to do it for them.” Twenty tasks: a love letter, an apology, wedding vows, a bereavement card, computer code, a dinner recipe.

Outsourcing to another human was judged more harshly than outsourcing to AI: less competent, less moral, less trustworthy, and lazier. The competence gap was the widest, close to a full point on a seven-point scale.

The penalties were largest for love letters and apologies, smallest for code and recipes.

Another study by Liu, Kang, and Wei (2024) found a compatible result for personal messages. When a close friend used help to write you a supportive note, getting help from another person was statistically indistinguishable from using AI on perceived effort, relationship satisfaction, and appropriateness. The authors’ explanation was that people don’t think a friend should use any third party, AI or another human.

In at least one domain the ordering flips entirely. Jago and Carroll (2024) found across four studies that producers received more credit for work when assisted by algorithms than when assisted by humans. The reason they identify is an assumption that algorithmic assistance requires more oversight from the producer.

Put these together and the pattern is not “people distrust AI.” At least not uniformly.

The actual variable

What predicts the penalty does not seem to be whether the helper was organic or inorganic. It is whether that kind of outsourcing has been ratified for that kind of task.

Where the human alternative is a sanctioned role-holder doing their job (a paralegal summarizing case law, a teaching assistant grading, a comms team polishing a statement), the arrangement is already legitimate. Everyone knows the role exists, what it covers, and who stays answerable. AI has no such standing yet, so it absorbs roughly twice the penalty.

Where the helper is an unspecified someone absorbing effort you personally owed, no arrangement is legitimate. Nobody has ratified having your apology written for you. Delegation is itself the violation, and a human delegate is judged as harshly or worse, because at least the machine was purpose-built to be used.

This is the legitimacy threshold, operating on authorship instead of assessment. It explains something the “AI can’t think” framing cannot: why the same person can use AI to draft a project update without comment and be accused of fraud for using it on a personal essay. The tool is constant. The ratification is not.

It also explains why the accusation stings. Being told your writing sounds like AI is not a claim about your intelligence. It is a claim that you used a form of help that is being perceived as illegitimate. That is a social charge, not a cognitive one, and social charges are harder to rebut, because there is no evidence you can produce.

But social ratification isn’t the whole story. I know because I’ve refused a ratified arrangement myself.

When I was building my own business, I dreamed about hiring someone to run my social media. Everyone told me to. It is about as legitimate an outsourcing arrangement as exists in professional life, and nobody would have thought twice about it. I tried a few times. I never could do it. Handing over my own voice felt wrong in a way I couldn’t argue myself out of, even though I could not have told you what principle I was defending.

So there are two thresholds. There is the social question of what a community will accept, which is what the research measures. And there is a personal question about where your voice stops being yours or what value you get out of a task that you’re unwilling to trade. No amount of convention settles that for you. The second one is why common practice never fully resolves the discomfort, and why watching someone else use AI comfortably doesn’t make you comfortable.

Subscribe now

We already know how to ratify help

Institutions have been solving the first question separately and without much fanfare, and three of them drew the same line independently.

The EU AI Act requires disclosure of AI-generated text published to inform the public on matters of public interest. Then Article 50(4) carves out an exemption: the obligation does not apply “where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”

Not “where a human typed it.” Where a named party is answerable for it.

The US Copyright Office reached the same distinction through different reasoning (Copyright and Artificial Intelligence, Part 2, January 2025). Prompts alone do not confer authorship, and iteration doesn’t fix that. The Office’s phrasing is that revising prompts repeatedly amounts to “re-rolling the dice,” and that “no matter how many times a prompt is revised and resubmitted, the final output reflects the user’s acceptance of the AI system’s interpretation, rather than authorship of the expression it contains.”

The other half of the ruling gets far less attention. Copyright still protects what the person contributed, even when AI-generated material is mixed in: “Copyright protects the original expression in a work created by a human author, even if the work also includes AI-generated material.” How you choose and arrange what the AI produced can count too.

One artist drew an image by hand and used it as her input. Her drawing was still visible in what the AI returned, so the Office registered that part of the work.

Accepting a machine’s interpretation is not authorship. Putting your own expression through a machine is. The law already distinguishes these.

Amazon’s publishing platform draws the line in operational terms. Content counts as AI-generated, and must be disclosed, “even if you applied substantial edits afterwards.” Content counts as AI-assisted, and requires no disclosure, if “you created the content yourself, and used AI-based tools to edit, refine, error-check, or otherwise improve that content,” including using AI to brainstorm.

Three bodies, three distinct traditions, a similar conclusion: generation triggers disclosure; assistance does not. But all three leave the same case unsettled: the idea is yours, you worked it out by going back and forth with the model, and a lot of the words on the page came from the model.

Why the tools we are building miss this

The instruments being built now do not measure the thing that turns out to matter.

Anthropic began watermarking Claude’s output in August 2026, in response to the EU requirement. The company’s own documentation is unusually candid about the limits. A watermark “can only determine that Claude was likely involved with the content at some point. It cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’” And: “The watermark only applies to words Claude chooses. When Claude proofreads text written by a person… there’s very little (if anything) for the watermark to attach to.”

Read those two sentences together. The mark is densest where the model generated freely and sparsest where a person did the thinking and used the model as a tool. It tags typing, not authorship. That is backwards from what almost anyone wants the signal to mean. Someone who talks through a half-formed idea and iterates until ninety percent of the words are their own gets marked. Someone who has a model produce a clean factual summary may not.

The test also only works in one direction. Finding a watermark tells you something. Not finding one tells you almost nothing. The text might be human. It might come from a different AI, or from a version of Claude released before August. It might be too short to carry the mark, or too factual. Or someone rewrote it.

LinkedIn’s contribution is a “Seems like AI slop” control, added in July 2026. What it actually does is feed the ranking system and train classifiers. It is not a report category under the platform’s policies, it does not label content, and it does not remove it. The company’s own product leadership has been explicit that “AI and slop are not the same thing” and that many people refine their thinking with AI. The target is mass-produced low-effort content. But the button’s label invites users to read it as a verdict on authorship, and users will.

What ratification would require

If the constraint is legitimacy and not detection, the work is different, and much less technical.

First, change the incentives before expecting people to be upfront. Tell people you used AI and it costs you. Get caught and it costs about twice as much. Say nothing and it costs the least. When those are the choices, people hide it. That is why there is a market for tools that strip the AI markers out of writing. As more tools emerge to catch it, others will emerge to circumvent them.

I saw this stated plainly in that same exchange. The person who had flagged my note described her own practice: she uses AI as a final spelling and grammar check, and otherwise limits herself to “private usage or just ways that I feel are reasonably undetectable.” She was not being evasive. She was being honest about a policy many careful people have settled into without ever discussing it, which is to use the tool where nobody can tell. That is what the math looks like from the inside. No norm will take hold while honesty is the expensive option.

The other thing I noticed: she edited her comment afterward so the negativity wouldn’t sit in my comment section. I appreciated it. But the retraction was private and the accusation had been public, which is the shape most of these take.

Careful wording does not solve it either. Schilke and Reimann tested six different ways of phrasing the disclosure, including “a human has reviewed and revised the work” and “AI was used only for proofreading.” All six reduced trust compared with saying nothing.

Second, show the difference between using AI and delegating the job to it. In a second Claessens study, someone who said plainly that they used AI as a tool, not as a replacement, was rated more moral and more trustworthy than someone who used no AI at all. Not just excused. Better. What people punish is handing over the whole task. Once they can see you didn’t, the penalty goes away.

That is the most actionable finding in this literature, and it aligns with the line that the EU, the Copyright Office, and Amazon already drew.

Third, say what a byline covers. This sounds bureaucratic. But this is the only item on this list an organization can settle on its own. A byline already covers a comms team, an editor, a research assistant, a fact-checker. Nobody wrote that rule down. We know what those roles do, and we know the person named on the piece is still responsible. Nobody has done that for AI. So everyone works it out on their own, in public, while being second-guessed.

The questions are concrete. Does our byline cover drafting? Structuring? Summarizing our own prior work? Who stays answerable when it is wrong? What has to be disclosed, and to whom? Does that disclosure go to readers, or to the institution, as it does at Amazon and at the Copyright Office?

Fourth, stop treating this as a detection problem. People identify AI-written text at roughly chance levels: 50 to 52 percent across six experiments with 4,600 participants (Jakesch, Hancock, and Naaman, PNAS, 2023). They also agree with each other about which texts look suspicious. When people agree with each other and are still wrong, they are all using the same bad heuristic. Researchers then tuned AI text to hit those cues. Readers judged it human 65.7 percent of the time. They judged actual human writing human only 51.7 percent of the time.

The heuristics are worse than useless. Grammatical errors read as human and were less likely to be AI. Contractions read as human and were more likely to be AI. First-person pronouns and family references, which people lean on heavily, carry no diagnostic signal at all.

So the defensive behaviors backfire. Deleting em-dashes, introducing typos, roughening your own prose: all of it optimizes against a model of detection that does not describe how people actually judge. It reads as not caring. Across eight studies with 5,306 participants, people who used texting abbreviations were rated less sincere, and the reason was that abbreviating made them look like they had put in less effort (Fang, Zhang, & Maglio, 2025).

And when these tools are wrong, they are wrong about the same people over and over. Researchers ran student essays through seven detectors. Essays by writers whose first language was not English were flagged as AI more than 60 percent of the time. Essays by American eighth-graders were almost never flagged (Liang et al., 2023).

The person who commented on my note told me her sister wrote a paper herself and had it flagged by her university’s AI checker. Everybody has one of these stories now. That is the system we are building.

The uncomfortable part

One finding underneath all of this reframes the anxiety, and it has nothing to do with AI.

Researchers had people pick out gifts and hand them to each other (Zhang & Epley, 2012). Half the givers were told to choose carefully. Half were told to choose at random. The receivers noticed the difference. But it made no difference at all with respect to how much they liked the gift. None. Givers felt closer to the people they had chosen carefully for. Receivers did not feel closer back. Givers were sure they would.

Effort counts most for the person expending it.

Hold that next to the experience of spending a week on something and being told it sounds like AI. Your effort was always clearer to you than to anyone reading it. What changed is not that readers stopped noticing care. It is that there is now a cheap explanation for why writing looks polished, and polish never carried as much information as we believed.

We have been treating the finished piece as evidence of the work behind it. It was never good evidence. It held up through a century of shortcuts. People cited a citation instead of reading the source, or skimmed an abstract instead of the study. Those shortcuts still cost real time. They were cheaper than the work, but not by much. So a polished piece still told you something, as long as faking one took nearly as much work as writing one.

That is no longer true, and no watermark will make it true again. That threshold has been crossed, and no watermark is going to uncross it.

What we are actually deciding

The question people think they are asking is whether a human wrote this.

The question they are actually asking is whether anyone is answerable for it, and whether the help behind it is the kind we have agreed to accept.

The first half we already know how to institutionalize. It is what a byline is. It is what editorial responsibility means in Article 50(4), what the Copyright Office tests when it asks whose expression is visible, what a reputation is for.

The second half is unsettled, and better detection will not settle it, because it was never a technical question. It is a question about which arrangements we are prepared to recognize, and we have answered it before: for ghostwriters, for editors, for research assistants, for every other form of help that now passes without comment.

It’s worth remembering when the demand is that contribution be separable, as it is when we question what percentage of a text that AI contributed. I have done my best work with other people, and the mark of the collaborations I’ve valued most is that at some point we lost the thread of who contributed what. The argument stopped being mine or theirs. The result was better than either of us would have produced alone, and neither of us could have drawn the line afterward if you’d asked.

Nobody audits those. We don’t ask co-authors to mark their sentences, or demand a coauthor’s paragraphs be shaded a different color, and if someone did we’d recognize it as a misunderstanding of what collaboration is for. Separability was never the standard for work we consider legitimate. It’s the standard we invented for work we haven’t decided about yet.

We answered the earlier questions slowly, through repeated practice and visible sanction, over decades.

Whether that process can run at the speed the tools are changing is the open question. The Economist compared 1.2 million words of model output against its own archive and found the markers had already flipped: the models that once overused the em-dash now use fewer of them than professional writers do (”How to Spot AI Writing,” 2026). Norms form through repeated interaction. If the ground shifts faster than the interaction cycle, you don’t get new norms.

So the question is not how to spot AI writing. What does our byline cover? Who stays answerable when the work is wrong? And how do we say so out loud, before the next person has to guess?

You get an argument that never ends. That is what is in the comments. Not people disagreeing about the rules. People finding out there aren’t any yet.

Leave a comment

References

Amazon Kindle Direct Publishing. (n.d.). Content guidelines: Artificial intelligence (AI) content. Retrieved August 17, 2026, from https://kdp.amazon.com/en_US/help/topic/G200672390

Anthropic. (2026, August 14). How Claude’s text watermark works. https://www.anthropic.com/news/claude-text-watermark

Claessens, S., Veitch, P., & Everett, J. A. C. (2026). Negative perceptions of outsourcing to artificial intelligence. Computers in Human Behavior, 177, Article 108894. https://doi.org/10.1016/j.chb.2025.108894

Fang, D., Zhang, Y. (E.), & Maglio, S. J. (2025). Shortcuts to insincerity: Texting abbreviations seem insincere and not worth answering. Journal of Experimental Psychology: General, 154(1), 39–57. https://doi.org/10.1037/xge0001684

How to spot AI writing. (2026, July 30). The Economist. https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing

Jago, A. S., & Carroll, G. R. (2024). Who made this? Algorithms and authorship credit. Personality and Social Psychology Bulletin, 50(5), 793–806. https://doi.org/10.1177/01461672221149815

Jakesch, M., Hancock, J. T., & Naaman, M. (2023). Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 120(11), Article e2208839120. https://doi.org/10.1073/pnas.2208839120

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779

Liu, B., Kang, J., & Wei, L. (2024). Artificial intelligence and perceived effort in relationship maintenance: Effects on relationship satisfaction and uncertainty. Journal of Social and Personal Relationships, 41(5), 1232–1252. https://doi.org/10.1177/02654075231189899

Perez, S. (2026, July 30). LinkedIn adds a button to report AI-generated ‘slop’. TechCrunch. https://techcrunch.com/2026/07/30/linkedin-adds-a-button-to-report-ai-generated-slop/

Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), 2024 O.J. (L 2024/1689). https://eur-lex.europa.eu/eli/reg/2024/1689/oj

Reif, J. A., Larrick, R. P., & Soll, J. B. (2025). Evidence of a social evaluation penalty for using AI. Proceedings of the National Academy of Sciences, 122(19), Article e2426766122. https://doi.org/10.1073/pnas.2426766122

Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, Article 104405. https://doi.org/10.1016/j.obhdp.2025.104405

U.S. Copyright Office. (2025, January). Copyright and artificial intelligence, part 2: Copyrightability. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf

Waters, S., & Nesbitt, C. K. (2026, February 10). Performance management in the age of algorithms series: Part II — How social systems, psychological contracts, and market tolerance shape algorithmic evaluation. Fractional Insights AI. https://fractionalinsightsai.substack.com/p/performance-management-in-the-age

Zhang, Y., & Epley, N. (2012). Exaggerated, mispredicted, and misplaced: When “it’s the thought that counts” in gift exchanges. Journal of Experimental Psychology: General, 141(4), 667–681. https://doi.org/10.1037/a0029223

Loyalty and performance yield security and growth. That was the deal and it held for decades. But today, it no longer holds, and many organizations are blissfully unaware of the consequences that will inevitably come.

Today’s model is an arrangement without terms. Workers are building side businesses, stacking roles, stockpiling savings against the next round of cuts, instead of fully committing to their roles. Companies are restructuring faster than they did two years ago, designing for agility over tenure, signaling with every decision that headcount is a variable, not an asset. Many organizations and employees are operating in “for now” mode, neither committing nor leaving.

We call this employment situationship: a relationship with real stakes but no clear future. The real danger is if organizations let it harden into a permanent default rather than making a choice to design something better.

We’re headed to a transactional equilibrium where neither the employer or employee fully invests, where compliance replaces discretionary effort, and where we call it “the new normal” even though no one intentionally chose it.

Share

Employment Situationship Violates Something Deeper Than Engagement

Employee engagement literature has frameworks for burnout, disengagement, and trust erosion. What it doesn’t fully capture is what happens when the structure of the employer-employee relationship becomes ambiguous in ways that can’t be resolved through better management or stronger culture.

The human brain is not built for this. Being part of a group you’re not loyal to, and that isn’t loyal to you, isn’t simply an uncomfortable feeling. It’s cognitively foreign. The research on motivation and belonging is consistent: humans are wired for group membership that is purposeful and reciprocal. When work requires people to simultaneously show up fully and maintain their exit strategy, the cognitive dissonance of that split creates a unique form of organizational stress.

It also forces people into a position they wouldn’t otherwise choose.

According to Fractional Insights research, about 35% of workers are “Universalists.” They are employees who seek to fulfill three core needs at work: security, growth, and significance. They are wired for deep investment, institutional knowledge-building, and long-term contribution. Within an employment system that reciprocates, their higher engagement and performance compounds over time.

Employment situationship overrides that orientation. When the structural conditions of employment communicate “for now” from both sides, Universalists adopt Transactionalist protective behaviors: hedging externally, maintaining portability, and limiting emotional investment.

This is the part that should bother leaders most. We’re not just losing engagement from people who were never that engaged. We’re actively converting our highest-potential institutional builders into short-term, self-protective operators.

Subscribe now

The Data Makes the Costs Visible

Workplace angst affects 44% of the workforce and costs large organizations with more than 10,000 employees between $240 and $330 million annually. The situationship amplifies every one of those pressures.

Recent research from Qualtrics, examining nearly 34,000 global employees across 24 countries and all major industries, reveals how this shift is manifesting in measurable ways. Three critical employee cohorts reported precipitous drops in experience year-over-year: part-time workers, customer-facing workers, and new joiners. These cohorts represent the places where organizations are withdrawing investment, whether intentionally or unintentionally.

When customer-facing employees were asked to evaluate their company’s overall performance compared to when they were asked how well it delivers on customer experience, their responses were nearly identical. These aren’t separate assessments in employees’ minds. They’re unified.

More striking: when researchers compared the same employees’ views on the causes of poor customer experience to actual consumer feedback from 20,000+ global consumers surveyed separately, these employees were more accurate than executives. The people being treated as most replaceable carry the sharpest signal about what actually drives customer dissatisfaction. And most importantly, they’re the ones directly delivering the customer experience.

The data suggests the drift toward situationship is well underway: Employee experience metrics dropping most sharply for the populations closest to customers, companies responding to volatility by designing roles for easy replacement rather than development. A thousand small decisions, each locally rational, collectively creating a system that few actually want.

To be clear, this isn’t a problem created by the newest generation of workers. It’s a design problem.

Intentionally Redesigning for the Future

Redesigning the employment contract starts with acknowledgment of the long term costs and honesty.

We must stop pretending that all roles carry the same implicit promise. Some roles should be designed for depth, tenure, and significant organizational investment. Others are designed for shorter-term contribution, flexibility, and skills portability.

Redefine what development looks like when the ladder isn’t the model. If tenure can’t be promised, what can? Skills that transfer. Project-based progression. Transparent criteria for opportunity, even without guaranteed outcomes. Employees can adapt to change. What they cannot adapt to is the gap between what organizations say they value and what their actions reveal.

The organizations that will attract and retain genuine high investment from the people most capable of it are the ones that tell the clearest truth about the terms while creating actual space for mutual investment within those terms.

And importantly, invest in frontline, part-time, and new employees. Not because it’s the right thing to do in the abstract, though it certainly is! But because the data is unambiguous: how you invest in these populations directly shapes the customer experiences that drive business performance.

Leave a comment

About the authors:

Shonna Waters, PhD has spent 25 years at the intersection of behavioral science and organizational strategy. She is the creator of Psychological Ergonomics™, a framework that treats workplace dysfunction as a systems engineering problem rather than a human one. She is SVP of Executive Engagement & Insights at Syndio, adjunct faculty at Georgetown University, and co-author of The Coaching Shift. She writes regularly on the future of work, AI, and how organizations redesign systems to fit the humans inside them to drive performance.

Dr. Benjamin Granger is Chief Workplace Psychologist at Qualtrics, and has over a decade of experience building, running and optimizing experience management (XM) programs across the globe. As Chief Workplace Psychologist, he leverages original research to offer insights into macro workplace trends, employee experiences, and the future of how we work.

The company hadn’t found product-market fit yet, and my stint in tech had ingrained an appetite for efficiency and scale. My urge to build the machine, even before I knew what the machine was supposed to make, was strong. But you cannot standardize something that hasn’t found its shape. This piece is about that tension: systems need variability while they’re finding their shape and consistency once they have one. Who, if anyone, decides how much variability a system keeps, and when?

It’s a tradeoff I’d been on the edges of throughout my career. At the fifty-year-old consulting firm where I started, the variability was the product: rigorous, tried and true methods, but no two engagements applied them the same way. At the government agency, consistency came first, because an inconsistent decision, in benefits, in enforcement, in due process, costs more than a slow one, and some of that inefficiency was doing constitutional work, protecting individual rights and preventing the consolidation of power. The nonprofit in between ran like a business, but with no market to check its setting. Each of them had a tolerance for variability, and none of them chose it deliberately. It was habit, survival, or instinct.

Then tech, where I joined one company in late Series A and stayed through late Series E, and watched practices that had been the key to success in one stage become the Achilles heel of the next. Everything about each of those places had been fit for purpose, at least at some point in time. Some of the standardization calls were clearly right. Some were made too early, before anyone had learned enough to know what should be scaled.

To name my bias: I love high-variability contexts. Building something from nothing, testing and experimenting, sorting through many wrong answers before finding a right one. So I start from the position that variability is the fuel adaptive systems need to innovate and create, not noise to be standardized away. On the other hand, too much variability creates risk: inconsistent quality, inconsistent decisions, inconsistent experiences.

Tolerance for variability is a setting, not a virtue. Genomes have one. Societies have one. Companies have one. And every one of those settings was calibrated by an environment that may no longer be the environment in front of it. Having the wrong setting from time to time is inevitable; environments change faster than calibrations do. Where we go wrong is not knowing we have a setting at all, and therefore never checking whether it still fits. I won’t arrive at a settled answer about where anyone’s setting should be. What I’m after is a better way of asking.

Share

What biology knows about this

Before I changed my undergraduate major to psychology, I studied evolutionary biology, and the question that pulled me in, nature and nurture, is largely a study of variability: where it comes from, what it does, and what we lose when we optimize it away. Natural selection needs variation to work on. A population with no genetic variation has nothing for selection to filter, and evolution stalls until new variation shows up. Biology treats variance as a strategy. In environments that swing unpredictably, a genotype that produces one “best” phenotype is betting that the future looks like the recent past; a genotype that produces a spread of phenotypes does worse on average right now and survives regime change. Desert annuals hold a fraction of their seeds dormant even in a good season, insurance against the drought that hasn’t come yet. Bacteria facing antibiotics keep a small subpopulation of slow-growing persisters that survive the assault and reseed the colony. The organisms that got it wrong don’t exist to study.

The alternative strategy has a name too. In 1942, Conrad Waddington called it canalization: development buffered so well that it produces the same outcome no matter the noise. Buffering is useful, which is why it evolves. The catch is that a trait that stops responding to disturbance also gives selection nothing to work with when the environment shifts. And suppressed variation isn’t always gone. Sometimes it’s destroyed. Sometimes it’s stored where it can be recovered. That difference will matter again when we get to AI.

In evolution, the variance often lives at the population level, and a skeptic could say the market plays that role for companies: over-standardized firms dying and being replaced is the system working as designed. But the people inside a firm cannot diversify themselves across ten thousand startups, and the market’s bet-hedging is cold comfort to the company that turns out to be the losing bet. If you run a firm, you are not the population; you are one of the seeds, and you get one draw.

A global stress test

Between genomes and companies sits a level where this trade has been measured across dozens of countries and then stress-tested by a global event.

In undergrad, I took cross-cultural psychology with Dr. Michele Gelfand. She studies how strongly societies enforce their norms, what she calls cultural tightness and looseness, and her research across dozens of nations shows the setting isn’t arbitrary; it tracks threat history. Societies that have faced more disasters, disease, scarcity, and invasion tend to have tightened, because coordinated compliance is what gets a population through an acute collective threat. Societies with gentler histories loosened. Tight cultures get order, coordination, and self-control. Loose cultures get openness, tolerance, and creativity. She is emphatic that neither is better, and the pandemic gave the theory an unusually direct test.

In 2021, in The Lancet Planetary Health, Gelfand and colleagues examined 57 countries and estimated that loose nations had roughly five times the COVID cases and nearly nine times the deaths of tight nations by October 2020, controlling for wealth, inequality, density, median age, government efficiency, underreporting, and more. Brazil and the United States each passed 24,000 cases and roughly 700 deaths per million in that window. Taiwan had 22 cases per million. The evidence is correlational, and tight countries tend to have other advantages, state capacity and recent epidemic memory among them. But for a threat that demanded fast, uniform, population-wide behavioral change, the pattern is what the theory predicted, and the margin was large.

If the story ended there, it would be a case for tightness. It doesn’t end there.

A small analysis in Psychological Medicine that same year found that across twelve countries, cultural tightness correlated negatively with willingness to receive the COVID vaccine, and willingness tracked how much virus was actually circulating. Tight cultures had contained the virus so effectively that their people perceived little risk, and vaccine demand fell. The setting that won the containment phase was mismatched to the vaccination phase, within the same crisis, within the same year. Later work found the mirror image: as conditions eased, looser cultures relaxed their guard fastest, while tight cultures held their vigilance well past the point the data justified it.

The pandemic’s lesson is more unsettling than “tight beats loose.” The same looseness that cost the United States dearly in 2020 is tied to the openness that produces its disproportionate creative and scientific range, including, not incidentally, several of the vaccines that ended the acute phase. Nobody gets to be maximally both. Each setting was right for some environment and will be wrong for another, and the settings persist long after the environments that calibrated them are gone.

Organizations have tightness too. It shows up in how strictly process is enforced, in what actually happens to the person who does it a nonstandard way, in how much consequence attaches to deviation. A company that tightened during a compliance scare, or a scaling push, or a near-death experience, and never revisited the setting, is running a calibration built for a threat it may no longer face. I have worked inside that company. So, I suspect, have you.

The newest lever on an old dial

Organizations have been adjusting this dial deliberately for over a century. Frederick Taylor’s scientific management compressed variability in physical process, replacing each worker’s judgment with one best method, and it delivered the productivity it promised. Six Sigma compressed it in outcomes, 3.4 defects per million opportunities, and Motorola credited it with billions in savings. The actual discipline was more careful than its caricature suggests; the target was unwanted variation, not variation itself, and in some domains – pay and promotion decisions chief among them – driving variance down is the right choice, because unmanaged variability there is how inconsistency and bias enter high-stakes decisions about people’s lives.

The costs took longer to surface. A twenty-year study of the photography and paint industries found that the more intensively firms adopted process management, the more their innovation shifted toward refining what they already knew, at the expense of exploration, the pattern James March named exploitation crowding out exploration. 3M ran the experiment on itself: a new CEO put Six Sigma through everything, including the research labs. Efficiency improved, the pipeline of genuinely new products thinned, and his successor pulled the program out of R&D, explaining that “invention is by its very nature a disorderly process.” He wasn’t rejecting Six Sigma, which stayed in manufacturing where it belonged. He was drawing a boundary around it. Every tool in this lineage worked where work had found its shape, and cost something where it hadn’t.

AI is the newest lever on the dial. It is the fastest and cheapest one ever built, and it works on a different target. Taylor compressed physical process. Six Sigma compressed outcomes. AI compresses ideas.

The evidence so far is early, and mostly pointing the same direction. A 2024 experiment in Science Advances found that writers given AI-generated ideas produced individually better, more creative stories, while the stories across writers became more alike. Individual writers gained while the group converged. A 2025 study of roughly 2,200 college admissions essays found that each additional AI-written essay contributed fewer new ideas to the pool than each additional human one, even after the researchers tuned the prompts specifically to force diversity. The mechanism is simple. Models tuned on aggregated human preferences learn to produce what most people rate as good, which pulls output toward the center of the distribution. The center is a fine place for a first draft. It is a poor place to find anything that could differentiate you.

This isn’t an anti-AI argument. I use these tools daily. But I think a lot about how I use them and whether it’s bringing me to the center or pushing me to the edge.

First, unlike canalization in biology, this compression is an engineering artifact, not a law of nature. Output diversity responds to how the tools are built and how they’re used; in one large experiment, people who encountered AI ideas as stimuli to react to produced a more diverse pool of ideas, not a narrower one. The research on training models on their own outputs, the model collapse result, shows distributions narrowing from the tails inward, but follow-up work shows the effect is largely avoidable when fresh real data keeps flowing in. That is the banked variation from the canalization story: collapse comes only when the reserve stops being replenished. These are choices someone is making, at the model labs and at every desk where the tools get used, and choices can be revisited.

Second, the gains are measurable, and they are table stakes. In a study published in the Quarterly Journal of Economics in 2025, customer service agents with an AI assistant resolved about 15 percent more issues per hour, and the least experienced gained closer to 30. Every competitor can buy the same lift. Dave Ulrich puts the logic in financial terms: performance is revenue over costs, and AI-generated capacity works on the denominator. That work is worth the effort, but costs have caps while revenue has no limits, and cost efficiencies tend toward parity as competitors match them with similar technology. Revenue comes from differentiation, and differentiation runs on the kind of output these tools compress by default. A company that adopts AI and stops at capacity hasn’t gained an edge. It has paid to arrive where everyone else is arriving, slightly faster.

Share

One more thread worries me most as a psychologist, because it compounds. Judgment is built through repeated exposure to ambiguous cases. The novice gains in the productivity studies come from AI substituting for judgment the novice doesn’t have yet. If AI increasingly intercepts the ambiguous cases before people have to sit with them, we’re cutting off the experiences that build judgment at the same time demand for it is rising. I don’t know when that bill comes due. It’s more likely to be years than quarters. But it lands on the people who were supposed to be ready for the decisions no model has seen, and never got the practice. Adoption dashboards won’t catch it.

The instrument, not the answer

I told you I wouldn’t arrive at a settled answer, and I haven’t. What I have, after running this question through evolution, cultures, factories, and models, is a better instrument for asking it. It has three parts.

Is variability treated as a feature or a bug? That’s our setting, and the honest answer isn’t the stated one. It’s what actually happens to the person who does it a nonstandard way.

What calibrated it? A funding crunch, a compliance scare, a founder’s last company, fifty years of bespoke work. Name the environment your setting was built for, and you can ask whether it’s the environment you’re in.

Which variability are we giving up right now, and is it the kind we mean to give up? Variability in how a routine task gets executed is usually safe to compress, and AI’s gains there look close to pure upside. Variability in judgment applied to people, compress that too, on purpose, for their protection. But variability in ideas, in the range of framings a group can generate, in the ambiguous cases that build the next generation’s judgment: that is the reserve everything else eventually draws on, and it is what the newest tools compress.

This sorting is also the beginning of an AI policy. The first bucket is where AI-assisted work should be default-on. The third is where it should stay default-off, or at least deliberately staffed with people who still need to build the judgment. Clay, a software company, wrote an AI writing policy on exactly this instinct: AI is welcome for brainstorming, drafting, and proofreading, but “writing is thinking,” every sentence has to remain the author’s own, and outsourcing the writing to skip the thinking defeats the point.

I’ve seen what this looks like first hand. During one company’s hypergrowth phase, we were pushing acquisition costs down by standardizing and scaling parts of the customer experience, and one instinct kept resurfacing: take the white-glove experience we’d built for executives and extend it to every customer. Sometimes that was the right call. Sometimes it was a timing problem. It was just too soon; a bespoke experience that fit perfectly in one context hadn’t been battle-tested for broad use. And sometimes the scarcity was the point: opened to everyone, it stopped feeling special and started feeling commoditized. Each time, the difference was what the variability was doing for us, and often, no one was asking.

Every organization I’ve been part of made these choices. Most of them weren’t making them consciously. Biology has been running this experiment for a few hundred million years, and its track record with systems that optimize their own variation away is not encouraging: they get efficient, and then they get brittle, in that order.

Organizations swing on this dial: tighten until something breaks, loosen until something else does. We usually narrate each swing as a correction, a strength overused, a miscalibration finally caught. Maybe. Or maybe it’s organizational physics: the tide goes out because it came in, and systems get thrown from exploitation back toward exploration only when the context demands it. I’d like to believe in a third possibility, a fit-for-purpose balance that can be designed and recalibrated as the environment moves. I’ve watched the swings from inside five organizations, and I can’t yet tell you which of the three is true. Every version of the answer starts in the same place, though: knowing you have a setting at all.

Leave a comment

References

Ashkinaze, J., Mendelsohn, J., Qiwei, L., Budak, C., & Gilbert, E. (2025). How AI ideas affect the creativity, diversity, and evolution of human ideas: Evidence from a large, dynamic experiment. In Proceedings of the ACM Collective Intelligence Conference. https://arxiv.org/abs/2401.13481

Balaban, N. Q., Merrin, J., Chait, R., Kowalik, L., & Leibler, S. (2004). Bacterial persistence as a phenotypic switch. Science, 305(5690), 1622–1625. https://doi.org/10.1126/science.1099390

Benner, M. J., & Tushman, M. (2002). Process management and technological innovation: A longitudinal study of the photography and paint industries. Administrative Science Quarterly, 47(4), 676–706. https://doi.org/10.2307/3094913

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.

Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), Article eadn5290. https://doi.org/10.1126/sciadv.adn5290

Gelfand, M. J., Jackson, J. C., Pan, X., Nau, D., Pieper, D., Denison, E., Dagher, M., Van Lange, P. A. M., Chiu, C.-Y., & Wang, M. (2021). The relationship between cultural tightness–looseness and COVID-19 cases and deaths: A global analysis. The Lancet Planetary Health, 5(3), e135–e144. https://doi.org/10.1016/S2542-5196(20)30301-6

Gelfand, M. J., Raver, J. L., Nishii, L., Leslie, L. M., Lun, J., Lim, B. C., … Yamaguchi, S. (2011). Differences between tight and loose cultures: A 33-nation study. Science, 332(6033), 1100–1104. https://doi.org/10.1126/science.1197754

Gerstgrasser, M., Schaeffer, R., Dey, A., Rafailov, R., … Koyejo, S. (2024). Is model collapse inevitable? Breaking the curse of recursion by accumulating real and synthetic data. arXiv. https://arxiv.org/abs/2404.01413

Hindo, B. (2007, June 11). At 3M, a struggle between efficiency and creativity. BusinessWeek.

Kirk, R., Mediratta, I., Nalmpantis, C., Luketina, J., Hambro, E., Grefenstette, E., & Raileanu, R. (2024). Understanding the effects of RLHF on LLM generalisation and diversity. In International Conference on Learning Representations. https://arxiv.org/abs/2310.06452

Ma, M. Z., Chen, S. X., & Wang, X. (2024). Looking beyond vaccines: Cultural tightness–looseness moderates the relationship between immunization coverage and disease prevention vigilance. Applied Psychology: Health and Well-Being, 16(3), 1046–1072. https://doi.org/10.1111/aphw.12519

March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71–87. https://doi.org/10.1287/orsc.2.1.71

Moon, K., Green, A., & Kushlev, K. (2025). Homogenizing effect of large language models (LLMs) on creative diversity: An empirical comparison of human and ChatGPT writing. Computers in Human Behavior: Artificial Humans, 6. https://www.sciencedirect.com/science/article/pii/S294988212500091X

Ng, J.-H., & Tan, E.-K. (2021). COVID-19 vaccination and cultural tightness. Psychological Medicine. Advance online publication. https://doi.org/10.1017/S0033291721001823

Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y

Taylor, F. W. (1911). The principles of scientific management. Harper & Brothers.

Ulrich, D. (2026, June 2). AI * HI for stakeholder HR: Turning capacity into capability [Article]. LinkedIn. https://www.linkedin.com/pulse/ai-hi-stakeholder-hr-turning-capacity-capability-dave-ulrich-pupgc/

Venable, D. L. (2007). Bet hedging in a guild of desert annuals. Ecology, 88(5), 1086–1090. https://doi.org/10.1890/06-1495

Waddington, C. H. (1942). Canalization of development and the inheritance of acquired characters. Nature, 150(3811), 563–565. https://doi.org/10.1038/150563a0

In June, a report caught my attention because it didn’t lead with a number. Commissioned by the National Center for O*NET Development (the keepers of the U.S. government’s occupational database), a team at HumRRO (Putka, Voss & Lewis) reviewed nineteen major studies of AI’s impact on work, from OpenAI’s exposure research to Anthropic’s Economic Index. HumRRO is known for its technical rigor, which I saw first hand thanks to spending the first half of my career there.

They asked a question I wish more of us would ask: what model of work are these studies actually using?

Most of them, it turns out, share the same assumption. I’ll come back to it, because it’s hiding inside nearly every AI exposure number you’ve read.

Their instinct is right. It just doesn’t go far enough. Before we measure what AI does to work, we should be honest about something the whole conversation skips: we’ve never fully agreed on what work is.

What is work?

Try to define it. It’s harder than it sounds.

One of the cleanest recent attempts comes from Adam Grant and Marissa Shandell in the Handbook of Social Psychology: work is “effort people expend in exchange for financial compensation.”

No offense to Adam, but that definition has always felt too narrow to me. A parent raising children isn’t working? Someone giving their Saturdays to a food bank isn’t working? It doesn’t sit right.

The definition I’ve been using instead: work is expending effort to create value for yourself and/or others.

Share

The anthropologist James Suzman goes further. He spent nearly thirty years with the Ju/’hoansi of the Kalahari, and his book Work: A Deep History, from the Stone Age to the Age of Robots defines work similarly: purposefully spending energy toward an end. What his long view adds is a timeline. For roughly 95 percent of human history, work had nothing to do with employment. Hunter-gatherers met their needs on the equivalent of a short modern workweek and spent the rest of their time living. Agriculture tied work to scarcity and to worry about the future. Cities and industrial firms tied it to something newer still: the job — titled, salaried, bounded.

And here’s what my own field keeps rediscovering from the other direction: humans are wired to work. Lottery winners keep working; retirees mentor and volunteer; the whole psychology of intrinsic motivation says we seek purpose, especially social purpose, well past the paycheck. Keynes predicted that by now productivity gains would have us working fifteen-hour weeks. We hit the productivity numbers. We kept working.

So work is old, human, and much bigger than employment. The job is a recent packaging technology: a way to bundle work so it can be assigned, priced, and managed. Useful? Enormously. But it is packaging, and packaging can be redesigned.

What is a job? (Not the same as what people do in one.)

Sharpen the definition of a job and it splits into two things the AI conversation keeps collapsing into one.

A job is an organizational artifact: tasks and responsibilities bundled so that work can be handed to one person, priced in a labor market, and managed inside a structure. The bundle is a design choice. Organizations draw job boundaries differently around the same underlying work, and redraw them when they must.

But a job describes how work is organized. It says nothing about what a person actually does or contributes. For that, my field uses a different construct with a century of science behind it: job performance — the behaviors through which someone creates value in a role. The distinction matters more than it seems, because AI-impact studies are trying to answer a performance question (what parts of what humans do can AI do?), typically using only the paperwork of work organization: task statements in an occupational database.

And performance science settled long ago that what people do exceeds what the task list says. The classic model (Borman and Motowidlo) separates task performance, the formal listable activities, from contextual performance: mentoring a newer colleague, cooperating across teams, volunteering for what needs doing, holding together the social fabric that makes the formal work possible. Underneath both sit the determinants — the declarative knowledge, procedural skill, and motivation that produce performance in the first place.

(There’s a third family the field recognizes: adaptive performance, how well people perform when the work itself changes. Organizations have mostly left it unspecified. The exceptions are instructive, such as special forces selection, which treats adapting as the job. In an era when task lists get rewritten in real time, I suspect it’s the performance domain we’ll come to care about most.)

So even before AI, inferring a job’s exposure from its task list was doubly incomplete. The task list captures one slice of performance, and performance is not the package it comes wrapped in. The task list is simply the visible layer: easiest to write down, easiest to standardize, and, not coincidentally, easiest to hand to a research team or a language model.

The assumption in the numbers

Back to HumRRO. Sixteen of the nineteen studies they reviewed draw on O*NET data — the government’s database of job characteristics, worker skills, and detailed occupation profiles. And they found that the majority take a task-centric view of job performance —performance: a lens inherited from labor economics and computer science, with almost no connection to the scientific literature where job performance has been studied for the past hundred years.

Nearly every exposure percentage you’ve seen was produced the same way: take an occupation’s task statements, score what AI can do against them, add it up. The contextual layer isn’t in the data. The knowledge and skill beneath the tasks are barely in the data. The interdependencies among tasks, the ways a job is more than the sum of its parts, are not in the data.

The authors expect AI’s impact on contextual performance to be far lower than its impact on task performance. If they’re right, the headline numbers systematically overstate how much of a job AI reaches. Not because the studies are sloppy. Because they inherited a model of work that was incomplete. The measurement was faithful to the task list. The task list was never faithful to the job.

The report’s constructive proposal is a suite of sixteen new AI-impact indices for O*NET, sorted by which layer of performance is measured, at what unit of analysis, and whether the impact is automation (AI does the work independently) or augmentation (AI assists while a human stays responsible). Most research treats those as points on one spectrum. HumRRO treats them as different phenomena that need separate measurement.

If you spend your days thinking about how work gets valued, that distinction is anything but academic. An automated task and an augmented task probably shouldn’t be trained for, governed, or paid the same way and the job evaluation and skills-based pay systems most organizations run today can’t see the difference. Today’s studies count both as “AI touched this.”

Why the layers matter

Once you see them, some apparent contradictions in the data resolve.

OpenAI’s analysis of 800,000+ ChatGPT conversations found that 43.5 percent of occupation-specific AI use involved tasks that historically belonged to a different occupation: people absorbing work that used to require a handoff to a specialist. Anthropic’s labor market research, meanwhile, finds the actual distribution of jobs has barely moved. Task-level turbulence. Job-level calm.

If a job were its task list, that would be a paradox. If what people do at work is task performance plus contextual performance plus the capabilities underneath, it’s roughly what you’d predict. AI is reorganizing the most codified layer of work first, while the layers that were never task statements — the judgment, the relationships, the motivation — change far more slowly. The task layer is fluid. The human layers are viscous. Job titles, the crudest layer of all, mostly relabel themselves. Which is one way to read the “AI created 1.3 million jobs” headlines: many of those roles are existing jobs re-bundled around new tools and renamed. Re-bundling is real change. But counting titles tells you about the packaging, not the work.

The question before the question

Why insist on this order — definitions first, measurements second?

Because the definition decides the story. If work is only employment (effort exchanged for pay, organized into task lists), then AI is a subtraction problem, and the question is how fast the list shrinks. If work is what Suzman describes and what the motivation research keeps confirming — purposeful effort to create value, something we do even when nothing requires it — then the story changes shape. Humans won’t stop working. We’ll redefine it, the way we have through every previous transformation. Task lists will be rewritten. Some jobs will genuinely disappear. New bundles will form. And the parts of work that never appeared on any list – including how we perform when the list itself changes – may turn out to be the parts that matter most.

The open question is whether the systems we use to define, measure, and pay for work will be rebuilt to see all of it, or only the layer that was easiest to write down.

I’d genuinely like to know: what’s your definition of work? And what does your job include that has never once appeared on a task list?

Leave a comment

References

Anthropic. (2026). Anthropic Economic Index: Labor market impacts. https://www.anthropic.com/research/labor-market-impacts

Borman, W. C., & Motowidlo, S. J. (1997). Task performance and contextual performance: The meaning for personnel selection research. Human Performance, 10(2), 99–109. https://doi.org/10.1207/s15327043hup1002_3

Grant, A. M., & Shandell, M. S. (2024). Work. In D. T. Gilbert, S. T. Fiske, E. J. Finkel, & W. B. Mendes (Eds.), The handbook of social psychology (6th ed.). Situational Press. https://doi.org/10.70400/HIBO9025

Kaplan, S. A., Aitken, J. A., Allan, B. A., Alliger, G. M., Ballard, T., & Zacher, H. (2025). Revisiting Keynes’ predictions about work and leisure: A discussion of fundamental questions about the nature of modern work. Industrial and Organizational Psychology, 18(1), 1–22. https://doi.org/10.1017/iop.2024.58

Keynes, J. M. (2010). Economic possibilities for our grandchildren. In Essays in persuasion (pp. 321–332). Palgrave Macmillan. (Original work published 1930)

OpenAI. (2026, July). Work at the frontier: How AI is expanding what people do at work. https://openai.com/index/how-ai-is-expanding-what-people-do-at-work/

Putka, D. J., Voss, N. M., & Lewis, P. (2026). Indexing the impact of AI within the O*NET system: A review of methods and development of recommendations. Human Resources Research Organization. https://www.onetcenter.org/dl_files/AI_Impact_Review.pdf

Suzman, J. (2021). Work: A deep history, from the Stone Age to the age of robots. Penguin Press.

World Economic Forum. (2026, January 23). AI has already added 1.3 million new jobs, according to LinkedIn data. https://www.weforum.org/stories/2026/01/ai-has-already-added-1-3-million-new-jobs-according-to-linkedin-data/

Subscribe now