We say we don't trust AI. Our behaviour says otherwise.

You may have said it yourself: "AI is useful, but I always check its work." Most leaders believe that about themselves, and most organisations have a policy that says a human reviews what the machine produces. The research suggests both beliefs deserve a closer look.

The picture that emerges is not that people trust AI too much or too little. My reading of the research is that we judge it inconsistently: by who is using it, by how it is labelled, and even by how long it takes to answer.

We say we distrust it. Then we defer to it.

The clearest evidence comes from a classroom experiment. In 2026 researchers asked more than 1,300 teachers in Greece to review student work in their own subject. Each piece came with a checklist and a grade that had been set deliberately wrong. Some teachers were told the grade came from an AI system, others from a human colleague.

When the grade was too harsh, teachers were less likely to correct it if it came from AI. The gap between their final grade and the fair grade was 22% larger for AI errors than for the same errors from a colleague. When the grade was too generous, they corrected it whatever the source.

The striking part is what the teachers said. In a survey, they rated the AI as less fair, less competent and less accountable than a colleague, and most said they did not want to use it. Yet in the grading task itself, they treated a harsh AI grade as competent and let it stand. Their distrust did not make them more careful. The deference was strongest among younger, more educated and more tech-confident teachers. As one of the authors, Rigissa Megalokonomou, put it: "Putting a human in the loop is not enough on its own."

This fits a longer debate. Earlier studies found that people avoid algorithms once they have seen them make a mistake. Others found the reverse: in six studies published in 2019, people changed their forecasts more when the same advice was labelled as coming from an algorithm rather than a person. My reading is that the direction of trust varies, but in neither case does it rest on careful checking.

Admit it, and lose trust

The second finding is the one most likely to affect your team. In 13 experiments with more than 5,000 people, published in 2025, researchers tested what happens when someone discloses that they used AI. Every time, trust in that person fell.

Students trusted a professor 16% less when they learned AI had helped with grading. Investors trusted a firm 18% less when its adverts disclosed AI use. Clients trusted a graphic designer 20% less. Job applicants who said they had used AI on a cover letter were also trusted less. Softer wording, such as "only for proofreading" or "reviewed by a human", did not stop the drop, and it happened even among people who used AI often themselves.

There was one way to make it worse. When someone else revealed that a person had used AI, trust fell further than when they admitted it themselves. The researchers called that "the worst possible outcome".

One rule for me, another for you

We also judge our own use differently. In two surveys of about 1,240 adults in Britain and the United States, people expected others to use AI tools for their messages far more than they would themselves. In the American sample, people put the chance that others would use them at about 71 out of 100, against 40 for themselves. Shown both together, British respondents rated open use as more acceptable than secret use; rated separately, the difference all but disappeared. The data were collected in 2021 and 2022, before the latest tools spread, so the gap may have changed.

The message matters as much as the messenger

The source of advice matters, but less than we might expect. In a 2026 study, 1,502 Americans rated the same health advice attributed to a human nurse, a specialist AI nurse or ChatGPT. The human nurse scored slightly higher, at 3.5 out of 5 against about 3.2 for both AI sources. The advice itself mattered at least as much as who gave it: advice that matched common expectations was rated more credible than accurate advice that went against common belief.

As the lead researcher put it, people "didn't dismiss the AI sources as untrustworthy". My reading is that people judge advice largely on whether it feels right, and that is a risk whoever gives it.

A pause makes it look wiser

Even timing changes our judgement. In a 2026 study, 240 people used a chatbot for everyday work tasks, with answers that arrived after 2, 9 or 20 seconds. The same answers were rated more thoughtful and more useful when they took longer. Many people read the pause as the AI "thinking". The researchers warned that people may place undue trust in slower systems.

What this is costing you

None of these studies was run in a company, so applying them to leadership is my own step. The patterns will be familiar, though:

• a shortlist of candidates screened by AI and signed off with a glance;
• a performance summary drafted by AI and accepted because it looks thorough;
• team members quietly using AI and saying nothing, because they suspect it will count against them;
• a "human in the loop" policy that nobody has ever tested.

The first two carry a particular risk. The teachers' study suggests that AI errors that are harsh on a person are the ones most likely to go uncorrected. The second two mean you may not know how much of your team's work AI already shapes.

Questions worth asking

A few honest questions show where you stand.

• Do you check AI work as carefully as you check a colleague's?
• If a colleague told you they had used AI on a report, would you trust the report more, less or the same?
• Has your organisation agreed when AI use should be disclosed, or is everyone deciding for themselves?
• When did anyone last test whether your human review actually catches AI mistakes?

What you can do this week

Everything in this section is a suggestion of mine, drawn from the research rather than tested by it. Each one takes minutes and costs nothing.

Suggestion

If you lead a team

• Check AI as you would a junior colleague. Ask for the reasoning, not just the answer, and look hardest at judgements that go against a person.
• Test your review step. Take one AI-assisted output this week and check it line by line yourself. Note what the normal review would have missed.
• Set a team rule on disclosure. Agree when AI use should be mentioned, so nobody has to choose between honesty and credibility.
• Judge the work, not the tool. When someone tells you they used AI, respond to the quality of what they produced.
• Ignore speed and polish. A slow, confident or well-formatted answer is not a checked answer.

Suggestion

For yourself

• Keep a five-day tally. Each time you accept an AI answer, ask whether you would have accepted it from a junior colleague without checking.
• Disclose on your own terms. If you use AI on something that matters, say so yourself and say what you checked. Being found out is worse.
• Notice your double standard. When you judge a colleague for using AI, ask whether you would judge yourself the same way.

Suggestion

At home

• Ask for it in their own words. When a child uses AI for homework, ask them to explain the answer in their own words before accepting it.
• Agree the family rule. Talk about when using AI is fine and when it should be mentioned, so honesty is not a risk.

Sources

• Goulas, S., Megalokonomou, R., & Sotirakopoulos, P. (2026). Why do experts miss AI's errors? Evidence from a randomized labeling experiment. PNAS Nexus. doi:10.1093/pnasnexus/pgag146
• Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405. doi:10.1016/j.obhdp.2025.104405
• Purcell, Z. A., Dong, M., Nussberger, A.-M., Köbis, N., & Jakesch, M. (2026). People have different expectations for their own versus others' use of AI-mediated communication tools. British Journal of Psychology, 117(2), 548 to 566. doi:10.1111/bjop.12727
• Landrum, A. R., Verma, N., & Kehrberg, A. (2026). Trusting generative AI for health advice: Preregistered survey experiment. Journal of Medical Internet Research. doi:10.2196/97882
• Tan, F. F.-Y., & Nov, O. (2026). Study of chatbot response delay and perceived thoughtfulness, presented at the ACM CHI Conference on Human Factors in Computing Systems (CHI '26), as reported by Neuroscience News, 24 April 2026.
• Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90 to 103.
• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114 to 126.