People tend to distrust AI lie-detection systems the more certain those systems are about accusing someone of deception—and this distrust actively undermines the AI’s accuracy. This new study was published in Computers in Human Behavior.
Artificial intelligence (AI) has proven surprisingly capable at detecting deception. By analyzing patterns in written language, AI models can classify statements as truthful or deceptive with a level of accuracy that far surpasses what humans can typically achieve on their own. In fact, research has demonstrated that laypeople and trained professionals, such as police officers, perform only slightly better than chance when trying to spot a lie.
Despite this, experts argue that AI should not be left to make such high-stakes judgments alone, particularly in legal or forensic contexts where the consequences for the accused can be severe. As a result, growing attention has focused on “hybrid” systems in which humans review and, when necessary, override AI predictions.
A new study examined how people respond when placed in this supervisory role. The researchers were particularly interested in whether trust in an AI lie detector would be influenced by two key characteristics: the system’s overall accuracy and the level of confidence it expressed in each individual judgment.
Led by Riccardo Loconte of the IMT School for Advanced Studies Lucca in Italy, the team recruited 373 English-speaking participants (52% female, average age 39). Participants were randomly assigned to one of two conditions: they were informed the AI they were working with had either a relatively low accuracy of 54%, or a high accuracy of 89%.
Each participant then read 10 short written statements about real-life experiences—such as being hospitalized or causing a car accident—and were shown the AI’s prediction for each one. This was displayed on a sliding scale indicating both the direction of the judgment (truthful or deceptive) and how confident the AI was. Participants then gave their own verdict on the same scale.
The results confirmed that people were more inclined to follow the high-accuracy AI, but revealed a striking and counterintuitive pattern. The more confident the AI was in labeling a statement as a lie, the more strongly participants pushed back against that verdict, shifting their own judgment toward seeing the statement as truthful instead. The effect ran in the opposite direction too: when the AI confidently labeled a statement as truthful, participants became more suspicious and shifted their judgment toward deception.
As the authors put it, participants tended to deviate from AI predictions when these were made with high confidence, “especially if the model predicted deception” and if the prediction came from the low-accuracy model.
Crucially, this skepticism came at a real cost to overall accuracy. When working alongside the high-accuracy AI, participants achieved a detection rate of 76%—significantly better than chance, but well below the AI’s standalone performance of 90%. Those paired with the low-accuracy model performed similarly to the model itself (57% accuracy vs 54%), meaning human input provided no meaningful improvement there either. In both cases, human involvement failed to add value and, in the high-accuracy scenario, actively degraded performance.
The researchers suggested a few explanations for why people resist confident AI accusations of lying: “Participants [tended] to overestimate their own deception detection capabilities. Moreover, in line with the truth-default theory [whereby people generally assume others are telling the truth unless given strong reason to suspect otherwise], such aversion could also reflect users’ cautiousness in deception accusations, given the social costs of falsely accusing someone of lying.”
The study does have notable limitations. For example, the task was entirely hypothetical and participants were not offered any real-world incentives to make accurate judgments. Furthermore, all statements were presented as text only, stripped of the conversational context that would normally inform a deception judgment.
The study, “Humans incorrectly reject confident accusatory AI judgments,” was authored by Riccardo Loconte, Merylin Monaro, Pietro Pietrini, Bruno Verschuere, and Bennett Kleinberg.