Why people tell machines the truth
It sounds backwards. But on the most sensitive subjects, people are often more honest with an AI than with another person.
Here is a finding that surprises almost everyone who hears it. In a study of health-screening interviews, people disclosed more, and showed less fear of self-disclosure and less impression management, when they believed they were talking to a computer rather than a human. Same virtual interviewer, same questions. The only thing that changed was whether people thought a person was on the other end. The belief that no human was watching was enough to loosen their tongues. In related work, military personnel reported more freely to a virtual interviewer than on the official post-deployment health form they were required to complete.
The reason is not that the machine is clever. It is that the machine does not judge. The single biggest brake on honesty is the fear of what the listener will think of you, and that fear eases when the listener is not a person who might remember your face, repeat what you said, or hold it against you later. One line of research calls this the machine heuristic: people will disclose more to a system precisely because they assume a machine handles their information more impersonally and more securely than a person would.
It is not a stray result. As reported in recent reviews of the field, a meta-analysis of thirty-nine studies found that computer-administered questions drew greater self-disclosure than face-to-face interviews, and a 2024 review of self-disclosure to conversational agents concluded that the weight of evidence points to people revealing more to these systems than to humans or to web surveys. Remove the human gaze, and people say more of what is true.
One finding matters most for anything built on this. People disclose more when an agent asks for information one item at a time, conversationally, than when it presents everything at once like a survey. The drip is gentler, and more honest, than the dump. A conversation that unfolds feels like being listened to. A wall of questions feels like being processed, and people answer a process the way they answer a process.
Two honest caveats, because the research has earned them. The findings are mixed: some careful studies find that disclosure to a machine is merely equal to a human rather than greater, not less, but not the dramatic uplift the headline suggests. And the effect depends on the agent responding well. An interviewer that feels cold, robotic, or inattentive loses it, because what people are responding to is the sense of being heard without being judged, and a poor listener provides neither. Candour is not automatic, even for a machine. It still has to be earned.
But the direction is clear, and it is the whole premise of conversational listening. Take away the fear of being judged, ask gently and one thing at a time, respond as though you are actually listening, and people will tell you what they would never write on a form. People are not dishonest by nature. They are careful in front of other people. Take away the audience, and the truth comes more easily.
References Lucas, G. M., Gratch, J., King, A., & Morency, L.-P. (2014). It’s only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94-100.
Incorporating verbal cues to improve computer-delivered messaging (the machine heuristic; disclosing more when asked one item at a time, per Adam et al.).
Weisband and Kiesler’s thirty-nine-study meta-analysis, as reported here.
Self-disclosure to conversational AI: a literature review and framework (Personal and Ubiquitous Computing, 2024).
For the counter-evidence, People provide personal information to AI and humans equivalently.
