Does using AI at work make colleagues think less of you?

Most advice about AI at work assumes the only question is whether the tool helps you. There is a second question that gets less attention: what do the people around you conclude when they learn you used it?

Two sets of experiments published in 2025 give a consistent and slightly awkward answer. People who use AI are rated as lazier and less competent, people who admit to using it are trusted less, and people who are found out are trusted least of all. The evidence is experimental and mostly based on hypothetical scenarios, which limits how far it can be pushed. It is still the best data available on a problem most developers have felt in a code review.

The social penalty study

Researchers at Duke University’s Fuqua School of Business ran four preregistered experiments (N = 4,439), published in the Proceedings of the National Academy of Sciences. The abstract states the result plainly: people who use AI at work anticipate and receive negative evaluations regarding their competence and motivation.

Both halves of that sentence matter. “Anticipate” means employees expect to be judged: the authors propose that people believe they will be evaluated negatively by others for using AI tools and that this belief is justified. “Receive” means the expectation is accurate. When other participants evaluated a described employee who used an AI tool, they rated that person lower on competence and motivation. Duke’s summary of the work is blunt: employees judge colleagues who use AI as less competent and lazier.

The paper also tested whether the judgement has consequences. It does: the authors report evidence that these social evaluations affect assessments of job candidates. Duke’s account adds that recruiting managers may penalise candidates who rely on AI to complete tasks.

The authors’ own summary of the bind: although AI can enhance productivity, its use carries social costs.

Who applies the penalty

The most useful finding for anyone deciding how to behave is about the evaluator, not the user. According to Duke’s account of the study, the penalty disappears when managers use AI themselves. People who rely on the tools do not read a colleague’s use of them as laziness.

That suggests the penalty is about norms, not about the work. In a team where everyone uses an assistant, using one signals nothing. In a team where the lead does not, it can signal a shortcut.

Disclosure makes it worse, not better

The obvious response is to be open about it. A separate research programme at the University of Arizona tested that directly, and the result runs against intuition. Across diverse tasks and roles, thirteen experiments consistently demonstrate that actors who disclose their AI usage are trusted less than those who do not. The university reports that the 13 experiments involved more than 5,000 participants.

The size of the drop varied by setting. In the examples the university highlights, trust from students dropped 16% when they learned a professor used AI for grading, and clients placed 20% less trust in graphic designers after AI disclosure.

Familiarity did not remove it. One of the authors noted that even with people who were very familiar with AI and used it frequently, the erosion of trust was there. Softer wording did not help either: saying AI was used only for proofreading, or that a human had reviewed the output, still reduced trust in every version tested.

Being caught is the worst outcome

This is the finding that stops the research from being an argument for secrecy. Lead author Oliver Schilke put it directly: trust drops even further if somebody else exposes you after using an AI detector or finding out about it some other way. In his words, a third party revealing your AI use is the worst possible outcome as far as trust is concerned.

So the ranking, from the experiments, runs like this. Not using AI, or using it undetected, costs nothing in trust. Disclosing costs something. Being exposed costs the most. Since you cannot guarantee remaining undetected, and AI-generated work is often recognisable, concealment is a gamble with the largest downside.

Why the penalty is not irrational

It is tempting to treat all this as prejudice that will fade. Some of it probably is. But colleagues also have practical reasons to look harder at AI-assisted work. In one 2026 survey of 2,078 US workers, run by a business publication and not peer reviewed, 45% of workers said they have had to fix or redo work from a colleague because it relied too heavily on AI.

If nearly half of your colleagues have cleaned up after someone’s unreviewed output, scepticism towards “I used AI for this” is partly earned. The penalty attaches to the inference that you did not check the work, and that inference is sometimes correct.

What is not known

Whether this happens with real colleagues. Nearly all of these experiments asked participants to judge a described person or a short scenario. Nobody has tracked actual teams to see whether known AI users get worse reviews, fewer promotions or less trust over time.

Whether the penalty is fading. The studies were published in 2025, while norms were moving quickly. The Arizona authors themselves expect the cost of disclosure may fade as use becomes ordinary. There is no repeat measurement yet.

Whether it differs by task. Using AI to draft boilerplate and using it to write a performance review are unlikely to be judged alike, and the experiments do not map that terrain in detail.

Whether demonstrating review removes it. The studies tested softer wording. They did not test showing your checking, for example tests written, sources verified, or a diff you can explain line by line.

What to do with this

The experiments do not license a tidy rule, but they point in a direction.

Do not hide AI use where a policy or a professional duty requires disclosure. Exposure is the worst outcome in the data, and in some settings concealment has consequences well beyond trust.

Let the work carry the evidence of your judgement. The penalty is an inference about effort and competence. The direct counter is work you can defend: you can explain why the code does what it does, you caught the error the model made, you know which parts you changed. That has not been tested in a trial, but it addresses the mechanism the studies identified.

If you lead a team, the norm is yours to set. The penalty vanished when evaluators used AI themselves, and the Arizona researchers recommend cultivating an environment where AI use is seen as normal and legitimate. A lead who is open about their own use, and specific about what review they expect, removes most of the reason for anyone to hide it.

Common questions

Do people judge you for using AI at work?

In experiments, yes. Four preregistered studies with 4,439 participants found that people who use AI at work anticipate and receive negative evaluations of their competence and motivation, and the effect carried through to assessments of job candidates.

Should I tell my boss I used AI?

Disclosure lowered trust in all 13 experiments in a University of Arizona study, but being exposed by someone else lowered it further, which the lead author called the worst possible outcome. Where a policy requires disclosure, follow it; the research does not support concealment as a safe strategy.

Does the stigma apply if my manager uses AI too?

Apparently not. Duke’s summary of the PNAS study reports that the penalty disappears when managers use AI themselves, which suggests it reflects team norms more than the quality of the work.

Will the stigma around AI use go away?

Nobody has measured the trend. The researchers expect the cost of disclosure may fade as AI use becomes widespread, but the studies were single snapshots published in 2025.

The bottom line

Using AI at work carries a reputational cost in experiments: lower ratings for effort and competence, lower trust if you disclose, and the lowest trust if you are caught. The cost shrinks or disappears when the person judging you uses AI too. None of this has been measured in real teams, so treat it as a plausible risk, not a law. The defensible position is to follow your disclosure rules, make your own review of the output visible, and if you run the team, make AI use normal enough that nobody needs to hide it.