Writing

It isn’t impostor syndrome if they really are judging you

October 2026 · an 11-minute read on AI at work, concealment and the penalty behind it

Nearly half of UK employees have passed off AI-generated work as their own. The research says their fear of being found out is accurate, and the name we’ve given it sends the response to the wrong place.

Somebody at your organisation drafted an email this morning with a tool they’d rather their manager didn’t know about. They work somewhere that has told them, in a townhall or a strategy deck, that AI skills are now essential. They agree. They’re using it. They’re also hiding it.

Nearly half of UK employees, 47 per cent, say they have presented AI-generated content as their own. Fewer than two in five use AI at work regularly. More people here have passed off work that wasn’t quite theirs than use the tools often enough for it to be a habit.

Across the forty-seven countries in the same study, half of employees say they’ll be left behind if they don’t use AI. In a separate international survey, nearly half said they’d be uncomfortable admitting to their manager that they’d used it for at least one common workplace task. Pressure to use, pressure not to be seen using. Britain sits below the international average on both adoption and concealment, which doesn’t make the bind different here. It makes it slightly later.

The name that’s attached itself to this, in about eighteen months of business commentary, is AI impostor syndrome. It arrived quickly and it hasn’t been examined much. It’s the wrong name, and the wrongness matters more than a definitional quibble usually would.

What the name is supposed to mean

Clance and Imes named the impostor phenomenon in 1978, describing high-achieving women who believed their success was luck or oversight and feared exposure. The definition has been argued over ever since, but its core has held. Impostor phenomenon requires a gap. Real competence on one side, felt fraudulence on the other, and a fear that outruns the facts.

The disproportion is the diagnosis. Without it, you have someone who is simply worried about something worth worrying about, which is not a psychological phenomenon and doesn’t need a name from psychology.

So the question raised by all this concealment isn’t whether people feel exposed. Plainly they do. It’s whether they’re right.

There’s a hint in the survey data that they might be. When people who said they’d be uncomfortable disclosing were asked why, the reasons they gave were social rather than procedural. Feeling that using AI is cheating. Fear of being seen as less competent. Fear of being seen as lazy. The reason cited least often, by about a fifth, was that AI use is discouraged by company policy. Whatever this is, it isn’t fear of breaking a rule. It’s fear of what colleagues will make of them.

What colleagues make of them

In a pre-registered experiment, more than a thousand engineers evaluated identical Python code. Some were told it had been written with AI assistance. The AI-labelled work was rated around nine per cent less competent, for work that was the same work. The penalty was roughly twice as large when the code was attributed to a woman as when it was attributed to a man, and larger for older engineers. Male non-adopters were the harshest judges of women who used AI.

That gendered pattern hasn’t replicated, and it’s worth saying so. A second set of experiments, published in PNAS, varied the target’s gender, age and occupation across nearly four hundred versions of the same description and found that none of them changed the size of the penalty. The stigma looked general rather than aimed. Two studies, two answers, and the gender finding is the more fragile of them.

What that PNAS work does establish is the penalty itself, outside engineering and at scale. Across four preregistered experiments with more than four thousand participants, people who were said to have used AI were rated lazier, less competent, less diligent and less independent than people who received the same help from a colleague, or no help at all. People deciding who to hire acted on it. And a third study, of professional communication, finds a double standard sitting underneath: people rate their own AI use leniently whether the assistance was light or heavy, but judge other people’s by how much they used.

Put those beside the concealment figures and something changes. The employee hiding their use of Copilot is estimating a social cost. The estimate is roughly correct. There is no gap between how competent they are and how competent they’re being taken to be, or rather, the gap has been closed from the outside, by colleagues who really do discount the work.

That isn’t impostor phenomenon. It doesn’t meet the definition. It’s an accurate reading of a room.

I should be careful here, because this is the joint the whole argument turns on and it isn’t something anyone has demonstrated. No study has measured impostor phenomenon and the competence penalty in the same people. What exists is two literatures that don’t cite each other, and the inference between them is mine to defend rather than the evidence’s to supply. It’s a strong inference. It isn’t a finding.

Why the wrong name does damage

Names route responses. Call something impostor syndrome and you’ve filed it under individual cognition, and the whole apparatus that follows is individual: reframe the self-talk, list the evidence of your competence, notice the distortion, work on the belief.

Applied to someone who has correctly perceived that their colleagues discount AI-assisted work, that apparatus tells them their accurate reading is a thinking error. It’s hard to think of a more efficient way to make a person feel less sane and no less exposed.

There’s a second problem with sending people down that road, which is that the road isn’t in good repair. The most-cited review of impostor syndrome treatment found remarkably little evidence for any of it. The strongest trial since compared coaching against training in about a hundred young employees and found coaching worked, with fear of negative evaluation carrying the effect. One good small study is one good small study. It isn’t a treatment literature.

The alternative account is that impostor feelings are elicited by environments rather than housed in people: that the same person will feel like a fraud in one workplace and not in another, and that what changed was the room. On that view, this particular version has an obvious environmental source. Nobody has set a norm. Organisations have told people to use AI and left them to work out for themselves whether saying so is admissible, which means the answer gets worked out silently, by watching what colleagues do, and everyone is watching people who are hiding.

The penalty isn’t unconditional either, and the condition is the giveaway. In the same experiments it vanished when AI was described as useful for the task in hand: candidates who used AI were then rated as better fits, not worse. What gets judged isn’t AI use as such. It’s AI use where nobody has said whether it’s allowed.

The obvious organisational fix doesn’t quite work, though, and it’s worth conceding before anyone offers it. When the researchers described AI use as common rather than rare, the penalty didn’t shift. Telling people that everyone does it isn’t enough on its own. What did shift it was whether the person doing the judging used AI themselves, which is a slower thing to change, and not something a policy can announce.

There isn’t a settled name for any of this, which is itself part of the diagnosis. Nobody has decided what the thing is, so the labels on offer have filled the gap by guessing, and every one of them guesses the same way. Impostor syndrome, AI guilt, AI shame, AI anxiety. All of them locate the problem inside the person who used the tool, and none of them mention the colleague marking the work down. That isn’t a naming error to be fixed with a better coinage. It’s the assumption doing the work before anyone has examined it.

The same problem, further upstream

Students are further along all of this than employees, and the numbers are worse.

Four in five use AI regularly in their studies. Two-thirds haven’t been transparent about it, presenting AI-generated content as their own and hiding their use of the tools. Over three-quarters have felt they couldn’t complete their work without it. Four in five say they put less effort into their studies knowing they can rely on it. A quarter to a third report reduced critical thinking, and less communication with instructors and peers.

Those figures come from the same global study that produced the workplace concealment data, so the comparison is like-for-like. Students conceal more than employees do, which is what you’d expect where the sanctions are formal and the rules are unclear.

Something here does look different from the workplace version, though, and it’s worth naming because it’s the strongest objection to everything above. If a student genuinely can’t complete the work without AI, and feels they can’t, then the feeling isn’t disproportionate either. But it isn’t a misread room. It’s an accurate report of not having built something. That would be a third thing, distinct both from impostor phenomenon and from the disclosure problem, and it would deserve a different response again.

Whether that’s what’s happening, nobody knows. Nobody has measured competence and impostor feeling in the same students. The deskilling claim circulates with much more confidence than its evidence supports: the strongest version I could find was an academic describing a large body of data on skill attrition without naming any of it.

What has been measured is narrower and points somewhere else. Four studies now put impostor measures and AI use in the same sample. A five-country survey of nearly eighteen hundred nursing students found that favourable attitudes towards AI went with lower impostor scores, and that AI training and confidence went with lower scores too, but that ChatGPT use specifically was associated with higher ones. Two small mediation studies of undergraduates, both around two hundred participants, found AI dependence associated with impostor feelings, with the association running partly through academic self-efficacy in one and through fear of negative evaluation in the other.

These are small, cross-sectional, convenience-sampled studies, three of them from single national samples outside the West, and one of them in a journal whose own publication dates are internally inconsistent. No single one of them would support a claim. What makes them worth mentioning is where they converge. The mediator that keeps appearing is fear of being judged, not loss of skill.

That’s the same mediator that carried the effect in the coaching trial. It’s the same mechanism underneath the workplace concealment findings. Three literatures that don’t know about each other keep arriving at social evaluation.

Which means the tempting story, that AI hollows out competence and the impostor feeling is the accurate residue, is not the story the research currently tells. It might still be true. It just isn’t what anyone has found.

What might be doing it

If the driver is social evaluation rather than skill loss, the question becomes what AI use does to a person’s standing among colleagues, and there the survey data describes a pattern without explaining it.

Half of employees say they use AI rather than collaborating with peers or supervisors to get work done. One in five say AI use has reduced their communication, interaction and collaboration at work. Among AI users, four in five turn to AI instead of asking a colleague at least occasionally, and the proportion doing so frequently is more than twice as high among the youngest workers as among the oldest.

Competence has generally been built by asking someone who knows, which involves being seen not knowing, in front of them. That exposure is the cost and it’s also the mechanism: the relationship forms in the asking. AI removes the exposure and the relationship together. It never says come and sit with me, I’ll show you.

Whether that isolation produces impostor feelings or merely accompanies them is not established. One conference paper reports experiments finding that AI augmentation at work induces impostor thoughts, which in turn reduce knowledge sharing and helping behaviour towards colleagues, which would make it a loop rather than a line. The data behind it isn’t public. I’d hold it loosely.

The likely objection

Norms settle, so it’s reasonable to think this is a transitional problem: that the competence penalty is what always happens early, that spreadsheets and calculators and spellcheckers all carried a whiff of cheating once, and that in five years nobody will think twice.

There’s something in that, and it comes from the researchers who documented the penalty rather than from its critics. They found the effect concentrated among evaluators who don’t use AI much themselves, and take that as a sign the stigma may fade as direct experience spreads. Nobody has measured the penalty twice, so the trajectory is inferred rather than observed. The direction isn’t really in dispute.

It’s beside the point for anyone currently inside it, though. Transitional periods are where people decide what to disclose, what to volunteer, and who to ask, and those decisions don’t wait for the norm to arrive. Whoever is drafting that email this morning is doing it now.

The part that’s yours

The people who feel watched are the people watching.

That’s the finding the commentary keeps missing. Every employee hiding their AI use is also, on the evidence, discounting a colleague’s work for the same reason, rating their own assistance leniently and other people’s by volume. The discount you’re afraid of receiving is the one you’re administering.

And there’s something in the research that makes this more than an irony. Whether AI use gets punished turns out to depend on whether the person judging uses it. Among evaluators who used AI weekly or more, the laziness penalty couldn’t be detected at all. In a hiring experiment, the frequent users went the other way and preferred the candidate who used AI. The penalty is being held in place by people who don’t use it, and by people who use it quietly, marking others down for being less careful about it than they are.

Whatever this thing turns out to be called, it isn’t a private problem and it isn’t solved privately. But there’s a version of it that’s entirely within reach, and it isn’t about disclosure at all.

The question isn’t whether you used AI on that piece of work.

It’s who you didn’t ask.

Sources

• Acar, O. A., Gai, P. J., Tu, Y. and Hou, J. (2025) ‘Research: the hidden penalty of using AI at work’, Harvard Business Review, August.
• Ahmad, E. and Arshad, T. (2025) ‘Artificial intelligence dependency, academic self-efficacy, and the imposter phenomenon in university students’, Research Journal of Psychology, 3(4), pp. 414–434.
• Arif, M. Q., Munawar, N. and Rabab, E. (2025) ‘From AI dependence to impostor syndrome: the mediating role of fear of negative evaluation’, Journal of Media Horizons, 6(4), pp. 661–669.
• Bravata, D. M. et al. (2020) ‘Prevalence, predictors, and treatment of impostor syndrome: a systematic review’, Journal of General Internal Medicine, 35(4), pp. 1252–1275.
• Clance, P. R. and Imes, S. A. (1978) ‘The impostor phenomenon in high achieving women: dynamics and therapeutic intervention’, Psychotherapy: Theory, Research & Practice, 15(3), pp. 241–247.
• Coman, A. and Cardon, P. (2025) study of AI-assisted managerial communication, International Journal of Business Communication.
• El-Ashry, A. M. et al. (2026) ‘Attitudes toward artificial intelligence and impostor phenomenon among nursing students: a five-country cross-sectional study’, Nurse Education Today, 161:107011.
• Feenstra, S., Begeny, C. T., Ryan, M. K., Rink, F. A., Stoker, J. I. and Jordan, J. (2020) ‘Contextualizing the impostor “syndrome”’, Frontiers in Psychology, 11:575024.
• Gillespie, N., Lockey, S., Ward, T., Macdade, A. and Hassed, G. (2025) Trust, attitudes and use of artificial intelligence: a global study 2025. The University of Melbourne and KPMG.
• University of Melbourne and KPMG (2025) Trust, attitudes and use of artificial intelligence: a global study 2025 — United Kingdom insights. Country snapshot, workplace indicators.
• Reif, J. A., Larrick, R. P. and Soll, J. B. (2025) ‘Evidence of a social evaluation penalty for using AI’, Proceedings of the National Academy of Sciences, 122(19), e2426766122.
• Slack Workforce Lab (2024) The Fall 2024 Workforce Index. Survey of 17,372 desk workers across 15 countries, fielded 2–30 August 2024.
• Zanchetta, M., Junker, S., Wolf, A.-M. and Traut-Mattausch (2020) ‘“Overcoming the fear that haunts your success”: the effectiveness of interventions for reducing the impostor phenomenon’, Frontiers in Psychology, 11:405.
• Zhao, P. (2024) ‘The impostor phenomenon of workplace artificial intelligence augmentation’, Academy of Management Proceedings, 2024(1). Conference abstract.

The first conversation is free. About twenty minutes, no obligation, and nothing is booked or paid for until after we’ve spoken.

Arrange a free first conversation