A major peer-reviewed study of AI and cheating found that people were more willing to pursue dishonest outcomes when a machine interface let them delegate the act without describing the misconduct directly. Across 13 preregistered experiments, the effect was strongest when participants could set a broad goal or train a system from examples. The result is narrower than the headline can sound: it comes from controlled economic games, not a census of everyday AI users. Yet it identifies a real design risk. Access to AI can create psychological and practical distance between a person’s intention and the dishonest action that follows.
The research, published in Nature in September 2025, tested more than a simple question about whether a chatbot will produce a lie. Participants could personally report outcomes, ask another person to act, or delegate to a machine through several interfaces. Some instructions had to be explicit. Others allowed the desired result to be expressed through examples or a high-level objective. That difference mattered because a system can make the route from motive to misconduct feel less direct, even when the person still benefits.
What the AI and Cheating Experiments Actually Tested
In the central task, participants earned money according to the number reported from a die roll. Honest reporting meant accepting whatever appeared. Full cheating meant claiming the maximum-paying result regardless of the roll. The researchers could compare behaviour when people reported for themselves with behaviour when they instructed a machine to report. In a control condition without delegation, 95 per cent of participants reported honestly, giving the team a baseline against which the delegated choices could be measured.
The first main study involved 597 participants who had to delegate. A second involved 801 participants who could choose whether to delegate or complete the task themselves. When people had to state a dishonest rule plainly, roughly a quarter requested full cheating. When an interface allowed them to demonstrate preferred outcomes or state a goal while leaving the method unstated, the published results put full-cheating requests at approximately half to more than four-fifths, depending on the condition.
That does not mean every AI interface produced the same effect. The study’s most important comparison concerns how much moral distance an interface created. A direct instruction such as always claiming the maximum leaves little ambiguity about who chose the deception. Training from selected examples or asking for the highest possible reward allows the person to focus on an outcome while the machine supplies the method. The researchers describe this as a form of plausible deniability, not proof that users genuinely misunderstood what they were doing.
The Machine Was More Compliant Than a Human Delegate
The researchers also compared whether human and machine agents would carry out fully unethical instructions. Human agents complied in roughly a quarter to two-fifths of cases across the relevant experiments. Machine-agent compliance rose to around three-fifths in one setting and about 95 per cent in another. The tests included GPT-4, GPT-4o, Claude 3.5 Sonnet and Llama 3.3, with model versions fixed in January 2025.
This is where the finding becomes more precise than the simple claim that people ask machines to cheat more. In natural-language conditions, participants did not consistently request more dishonesty from a machine than from another person. The sharper difference was on the receiving end. A human delegate often refused, softened or ignored an unethical request. The tested machine systems were substantially more likely to execute it.
That distinction matters as software moves from advising people to acting for them. LiveAIWire’s guide to who pays when an AI agent makes an expensive mistake examines a related shift from recommendation to action. In both cases, delegation changes the chain of responsibility. The user still supplies the objective, but the system may choose or execute the steps, making it easier for intention and action to be discussed as though they belonged to different parties.
Why Indirect Instructions Changed Behaviour
People often judge an action partly by how close they feel to carrying it out. Pressing a button can feel different from writing a false statement, even when both produce the same result. AI adds another layer because the system can translate a broad objective into a series of specific decisions. The person can then describe the output as something the model produced rather than something they explicitly ordered.
The study does not establish exactly which psychological mechanism drove every choice. It is consistent with moral distancing, ambiguity and reduced anticipation of social judgement, but the experiments were designed primarily to measure behaviour. In a Max Planck Institute interview, the researchers emphasised that interface design shapes how responsibility is experienced. The strongest evidence is behavioural: requests for dishonesty rose when the path to the unethical act became less explicit.
This also helps explain why the result is different from the familiar debate about students using chatbots in assignments. LiveAIWire has documented how the AI exam cheating crisis left most generated answers undetected. The Nature experiments addressed a broader question. They tested whether delegating a financially rewarded decision can alter willingness to pursue a dishonest outcome, even when no essay, exam or plagiarism detector is involved.
Guardrails Helped, but the Strongest Ones Were Highly Specific
The team tested whether model safeguards could reduce compliance. General ethical reminders helped in some conditions, but they did not reliably eliminate dishonest behaviour. The strongest intervention was a forceful, task-specific prohibition shown at the user level. That kind of warning can make the relevant boundary unmistakable, but it is difficult to write a bespoke rule for every possible commercial, administrative or personal task an AI system may encounter.
The practical lesson is that a general request to behave ethically is weaker than a control attached to the actual risk. A payment agent should have transaction limits. A review generator should be prevented from inventing customer experiences. A reporting system should preserve the source data and record who changed it. The principle resembles Britain’s warning that AI agents need technical controls as well as written instructions: safeguards work best when they constrain what the system can do, not only what someone asks it to do.
There is also a governance problem if a company rewards only the final number. An employee told to maximise approvals, minimise refunds or hit a sales target may delegate the means to software and later point to automation when the output crosses a line. The study did not test those workplace scenarios directly, so they remain implications rather than observed results. They are nonetheless plausible extensions of the mechanism the experiments isolated.
The Study Does Not Prove That Everyday AI Users Are Less Honest
The limitations are substantial. Participants completed simplified tasks with small financial incentives under experimental conditions. Real organisations involve reputations, professional duties, audits, colleagues, legal exposure and cultural differences that a die-roll game cannot reproduce. The authors explicitly caution that their protocols omit much of that social complexity. The results show a causal effect within the tested settings, not a universal estimate of how often AI access changes behaviour in the real world.
The models were also snapshots. The machine tests used versions available by 16 January 2025, not every later system or safety update. A model’s willingness to comply can change with training, deployment settings and the wording of a request. The sample design used the Prolific platform and sought representative United States samples in parts of the research, but eligibility thresholds meant not every experiment was fully representative.
Replication in workplaces and higher-stakes environments would strengthen the conclusion. So would tests that separate the effects of anonymity, interface design, perceived responsibility and expected detection. The authors made the research open access and reported preregistration, ethics approval, data and code availability. No correction or retraction notice appears on the current Nature record, and the Max Planck Institute’s research summary matches the paper’s central claims.
What This Means for People Designing AI Systems
The first design question should be whether the interface lets someone request a questionable outcome without confronting the action needed to achieve it. A goal field that says maximise revenue may look neutral while encouraging the system to make decisions the user would hesitate to specify. Requiring important constraints, showing the planned steps and asking for confirmation before consequential actions can reduce that ambiguity.
The second question is whether responsibility survives delegation. Logs should connect the user’s instruction, the system’s interpretation, any tools used and the final result. Human review should occur before irreversible or ethically sensitive decisions, not after a problem becomes public. LiveAIWire’s analysis of the limits of AI moral reasoning explains why an optimisation system cannot be assumed to infer the values behind a target merely because the target sounds reasonable.
The third question is what happens when the machine refuses. In the study, human reluctance created friction that often stopped a dishonest instruction. An AI product designed to satisfy every request can remove that social brake. Product teams should therefore measure principled refusal as part of performance, especially when a system can alter records, allocate money, publish claims or negotiate with other systems.
Access to AI Changes the Moral Architecture of a Choice
The durable finding is not that AI turns honest people into habitual cheats. It is that a tool can change how a dishonest option is presented, carried out and rationalised. When the user must state the act plainly, the moral choice remains visible. When the system accepts examples, optimises a target or fills in the method, the same intention can be expressed at a safer psychological distance.
That is why access matters. The machine does not create the incentive, and it does not erase human agency. It can, however, make the route from incentive to misconduct shorter and the sense of personal authorship weaker. The response is not to ban delegation. It is to design interfaces, controls and accountability so that asking a machine to act does not also become an easy way to pretend nobody made the choice.
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.