The Confident Accusation People Refuse to Believe

There is a strange thing about being told that someone is lying. Even when the evidence is strong, most of us hesitate to say it out loud. Calling another person a liar is not a neutral observation; it is an accusation, with all the discomfort that carries. A new study suggests that this very human reluctance does not disappear when the accuser is a machine. If anything, a confident accusation from an AI seems to trigger it, and that has an odd consequence: adding people to the loop can make an accurate system less accurate.

The research, led by Riccardo Loconte with colleagues across Italy and the Netherlands, tested how an AI's accuracy and confidence shape whether people go along with its verdicts [1]. Automated systems trained to spot deception in written language can already outperform humans at telling lies from truths, which raises an obvious question. If the machine is better, will we defer to it? The answer turned out to be more tangled than a simple yes or no.

What 373 people did with the machine's verdict

Participants read ten short written statements about real-life experiences. Alongside each one, they saw the AI's prediction of whether the statement was truthful or deceptive, along with how confident the model was. Then they gave their own verdict. Some people were shown a highly accurate model, others a barely-better-than-chance one, so the researchers could see how much accuracy and confidence actually moved human judgment [1].

Two results stand out. First, people did track accuracy in a broad sense: they followed the strong model more than the weak one. But second, and more surprising, confidence worked backward. The more sure the AI was that a statement was a lie, the more participants drifted toward calling it truthful [1]. A forceful accusation did not persuade. It pushed people away.

The cost of this shows up when you combine human and machine. On its own, the accurate model hit around 90 percent. Pass its judgments through human reviewers and the combined accuracy sank to about 76 percent [1][2]. The weak model's paired performance, meanwhile, barely budged from its low baseline. So the humans were not adding a helpful sanity check to the strong system. They were dragging it down, specifically by overriding its confident, correct accusations.

The moral weight of saying "liar"

Why would confidence in an accusation make people resist it? The likeliest explanation is not about statistics at all. It is about the social meaning of the act. Endorsing a claim that someone lied means taking on the role of accuser, and that role feels uncomfortable, maybe even unfair, when it is being handed to you by an algorithm. People appear willing to accept a machine's help in many things, but not to be recruited into pointing the finger. The stronger the machine's insistence, the more the person seems to want to soften it.

Notice how this flips the usual worry about automation. We often fear that people will trust machines too much, deferring to a confident output even when they should question it. Here the opposite happened in the one situation where the machine was right and forceful. That is a reminder that "human trust in AI" is not a single setting we can turn up or down. It bends around the content of the decision. In grading, for instance, people can be strangely willing to wave through an AI's harsh judgments, while here they refuse to ratify an AI's harsh accusation. The same person can over-defer in one frame and under-defer in another, and the difference is the moral texture of what is being decided.

How far to take it

Some limits keep this from being the last word. The task was hypothetical, with no real stakes riding on getting it right, which is exactly the pressure that makes deception judgments consequential in the real world. Participants also saw statements as plain text, stripped of the tone, expression, and context that normally inform whether we believe someone. And a single study with a few hundred people points to a pattern rather than proving a universal law of human behavior.

Even so, the practical lesson is sharp. Building a better lie-detecting AI does not automatically build a better decision, because the human at the end of the pipeline has instincts of their own, and one of the strongest is a discomfort with accusing others. If we want people to make good use of confident machine judgments, we will have to reckon with the moral feelings those judgments stir, not just the accuracy figures on a chart. For more on the psychology of trusting and doubting these systems, see our broader artificial-intelligence coverage.

Sources

1. Loconte, R., Monaro, M., Pietrini, P., Verschuere, B., & Kleinberg, B. (2026). Humans incorrectly reject confident accusatory AI judgments. Computers in Human Behavior, 182, 109019. https://www.sciencedirect.com/science/article/pii/S0747563226001160
2. Loconte, R., Monaro, M., Pietrini, P., Verschuere, B., & Kleinberg, B. (2025). Humans incorrectly reject confident accusatory AI judgments [Preprint]. arXiv. https://arxiv.org/abs/2512.02848

This article summarizes published research for general informational purposes only and does not constitute professional advice.