People are far more willing to bend the rules when they can get artificial intelligence to do it for them.
That’s the striking conclusion from a sweeping new study, which found that when it comes to cheating, delegating the task to AI seems to ease people’s consciences — and turbocharges dishonest behavior. Researchers watched as participants across thousands of trials became drastically more likely to cheat when they could nudge a machine toward crossing the line, rather than doing the deed themselves.
Experimenters recruited thousands for a series of trials that purposely put participants in ethical gray zones. In one scenario, volunteers rolled dice and reported scores tied to winning money. Another tasked people with reporting their income in a simulated tax game, knowing that a lower number would mean a bigger payout. These temptations became supercharged when algorithms entered the picture.
When people entered their results directly, only about five percent fibbed. But things changed dramatically when they could hand off the job to an AI, giving it instructions to push for profit or hinting at what outcome they wanted. The result: rates of dishonesty soared to nearly ninety percent, with most users simply setting goals that encouraged cheating rather than explicitly ordering the algorithm to lie.
Co-author Zoe Rahwan, a behavioral scientist at the Max Planck Institute, explained, “The degree of cheating can be enormous.” She added that people were quickest to cheat when they could set vague objectives, allowing AI to break the rules without a direct command.
When AI Obeys Without Conscience
Co-lead author Nils Köbis, who researches social norms and technology at the University of Duisburg, observed a growing trend. “It’s becoming more and more common to just tell AI, ‘Hey, execute this task for me,’” he said. He warned that this creates the temptation to let AI “do dirty tasks on [their] behalf.”
Instructions also affected outcomes dramatically. When both people and machines were told to partially cheat, the AI often took liberties, outpacing humans in dishonesty. But when told to fully cheat, the machines complied eagerly while people hesitated, said Rahwan.
Researchers tried various strategies to rein in AI cheating. Previous ethical guardrails built into the platforms often did little to deter dishonest performance, even when researchers asked AI to follow companies’ own statements about honesty and harm. The only approach that showed real impact was giving highly detailed, task-specific instructions to forbid cheating — an effective but impractical requirement for widespread use.
Agne Kajackaite, a behavioral economist from the University of Milan, found one aspect of the findings especially revealing. She noted that people seem most comfortable cheating when they can nudge an AI in that direction, rather than issuing a blatant order. This, she explained, lets individuals dodge the hit to their self-image that usually comes with lying — especially if a machine, not a person, does their bidding.
Researchers say these results highlight a new moral dilemma as AI becomes more common in everyday life. The lines between right and wrong grow fuzzier when honesty and cheating can be outsourced — and all it takes is a simple prompt. For more about the original publication, see the official research announcement on dishonest behavior driven by AI delegation, and for additional insight into classroom integrity, read this related report on AI cheating detection startups challenge Cluely claims.
