People may be more willing to behave dishonestly towards an AI than towards a person because they expect less social embarrassment from being caught by a machine. That is the central finding of a 2026 study in the Journal of Business Research, which ran one field experiment and five scenario-based experiments. The researchers found that unethical consumer behaviour increased in interactions with AI agents, and that the effect weakened when the AI appeared more competent or seemed to look at the user. The published paper describes this mechanism as lower “anticipatory face loss”.
In plain English, people often avoid cheating not only because it is wrong or risky, but because they do not want another person to see them as dishonest. A machine may not trigger the same social pressure. If the customer feels that nobody meaningful is watching, the psychological cost of bending a rule can fall.
Six experiments tested the idea in different ways
The research record summarises the study as six experiments across human and AI interactions. In the first experiment, 159 valid participants in China read a scenario involving a human or AI agent. A second study used 180 valid participants and tested whether anticipated loss of face explained the difference. A third recruited 120 participants in the United States through Prolific to compare face loss with guilt.
Later experiments changed how the AI was presented. One study compared a human agent with control, competent and warm AI versions. The final experiments examined eye gaze. In the field-style experiment, participants completed a knowledge task and had an opportunity to misreport their performance for an additional reward. The aim was to move beyond attitudes and observe behaviour with a real incentive.
Across the programme of studies, the researchers concluded that lower anticipated face loss helped explain why people were more willing to act unethically towards AI. Guilt did not explain the difference as strongly in the comparison they ran.
Making the AI seem competent changed behaviour
The effect weakened when the AI agent was presented as competent. That is notable because competence is not the same as warmth or friendliness. A user may care more about being judged by an entity that appears capable of understanding what is happening than by one that simply sounds pleasant.
Eye gaze also mattered. When the AI agent appeared to look at the user, the tendency towards unethical behaviour weakened. The result fits a much older feature of human behaviour: cues of observation can change what people do even when the observer has limited power. In an AI interface, a pair of eyes may create enough social presence to make the interaction feel less consequence-free.
This is not a reason to put fake eyes on every chatbot
The obvious design response would be to make AI systems look more human and more watchful. That would be too simple. An interface that deliberately creates the feeling of surveillance can introduce new problems, especially if users do not know what is actually recorded. A visual cue should not imply that a system can see, remember or report something when it cannot.
The finding is better treated as evidence that social design changes behaviour. LiveAIWire has covered how people apply social norms differently to AI and how flattering AI can alter trust and influence. The new study adds dishonesty to that list. Users are not responding only to the answer an AI gives. They are responding to what kind of social actor they believe the AI is.
Businesses have a practical problem if customers exploit AI
Customer-service agents increasingly handle refunds, vouchers, complaints and eligibility decisions. A human employee can notice suspicious stories, ask an unexpected question or communicate that a request crosses a line. An automated agent may follow a predictable policy. If users feel less shame about exploiting it, companies can face losses even when the AI is technically following its rules correctly.
The answer is not necessarily tighter automation. Overly rigid systems can punish honest customers who have unusual circumstances. A better approach may combine clearer rules, anomaly detection, escalation to a person and interface cues that make responsibility visible without pretending the machine has human emotions.
The cultural and experimental limits matter
Several studies used participants in China and one used a US sample. Concepts related to face and social reputation can operate differently across cultures, so the size of the effect should not be assumed to be identical everywhere. Most of the experiments also used scenarios rather than long-term real-world customer relationships. The field-style experiment strengthens the evidence, but it still does not reproduce every setting in which people interact with automated services.
The experiments also focus on specific kinds of unethical consumer behaviour. They do not show that people become broadly less ethical simply because an AI is present. The more precise claim is that, in the situations tested, interacting with AI reduced anticipated social face loss and that this was associated with more willingness to break rules or misreport behaviour.
AI can remove the social friction that keeps people honest
That is what makes the result useful. Automation is often sold as a way to remove friction. Yet some friction performs a social function. A human at a counter makes a dishonest request feel different from clicking through a chatbot. LiveAIWire’s coverage of how people react when humans and AI make decisions shows the same broader point: replacing a person can change behaviour even when the formal rule stays the same.
For designers, the task is to preserve accountability without manipulating users. A system can state that claims may be checked, make escalation visible and explain that fraud policies apply equally to automated interactions. Those cues may be more transparent than trying to make the AI feel artificially human.
The study’s most memorable result is the eye gaze, but the deeper lesson is about social presence. People behave partly according to who they think is watching. As AI takes over more service interactions, companies will need to understand not only what the system says, but what its presence quietly tells users about whether normal social rules still apply.
Honesty cues have to be genuine
The eye-gaze result does not mean companies should simply give every chatbot a pair of animated eyes and expect people to behave better. If users are led to believe they are being watched when no such monitoring exists, the interface itself becomes deceptive. The more defensible lesson is that accountability cues matter. A system can clearly state when activity is logged, when suspicious behaviour may be reviewed by a person and what rules apply to a transaction.
That approach is especially relevant in settings such as expenses, insurance claims, marketplace disputes or employee systems where people may be tempted to exploit what they perceive as a less socially consequential machine. A well-designed interface can make the chain of responsibility visible without pretending the AI is a person. It can also escalate unusual cases to human review rather than trying to solve an ethics problem with anthropomorphic design alone.
The experiments also suggest that perceived competence changes behaviour. If people regard an AI as capable of detecting manipulation, the psychological distance between dealing with a machine and dealing with a person can narrow. In practice, however, designers should not overstate what an automated system can detect. False claims about surveillance or detection could damage trust and may create legal problems of their own.
The broader point is that automated interactions are social environments even when no human is visible on the other side. People infer whether anyone will notice, whether anyone could be embarrassed and whether a rule feels enforceable. As AI takes over more routine decisions, organisations will need to design those environments deliberately. The goal should be clear responsibility and truthful feedback, not a theatrical imitation of human judgement.
Future research can test that distinction more directly by separating truthful accountability signals from human-like appearance. For example, an interface could tell users that claims are recorded and may be audited without adding a face or gaze. If behaviour changes in the same direction, the important mechanism may be perceived accountability rather than the visual impression of being watched. That would offer organisations a clearer route to ethical design.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.