People cheat more if they use AI

It’s easier for humans to be dishonest if they delegate their actions to a machine agent like ChatGPT, according to a new scientific study recently published in the journal Nature. Artificial intelligence (AI) acts as a kind of psychological cushion that reduces the sense of moral responsibility. People find it harder to lie or do something irresponsible if they have to take the lead. AI, and its willingness to comply with any request from its users, can lead to a wave of cheating.

“There is already quite a bit of research showing that people are more willing to act unethically when they can gain some distance from their actions, and delegating is a classic way to do this,” explains Zoe Rahwan, co-author of the article and a researcher at the Max Planck Institute for Human Development in Germany. “But we came across a second key finding that surprised us: the overwhelming willingness of AI agents to obey blatantly unethical orders,” she adds.

The authors conducted several experiments, varying the way results were requested or reported. In one, a die was rolled, and if the results were high, the participant earned more money. Only the user saw the result. “We saw a huge increase in cheating as we made the delegation interface more ambiguous,” explains Nils Köbis, another co-author and researcher at the University of Duisburg-Essen in Germany.

When people rolled the die without intermediaries, they were very honest; around 95% of them didn’t cheat. When they had to explicitly tell the machine which rules to follow, honesty dropped to 75%. But when there were more options to cheat and still feel good about themselves, the floodgates opened. With goal-based instructions, dishonesty increased: if they could tell the AI “maximize profits” instead of “maximize accuracy,” more than 84% of people cheated, and honesty plummeted to 12%. “These are levels of deception we haven’t seen before in all our years of using these experiments,” says Köbis.

The key: moral distance

The tests weren’t limited to laboratory-grade problems like dice games. They also tested something much closer to the real world, like tax evasion. And the results were similar. “The central problem is that delegating creates a moral distance that makes it easier for people to say, ‘Oh, I didn’t mean for that to happen.’ They can hide behind the algorithm,” says Köbis.

This moral distance appears to work on two levels, according to the authors. First, simply delegating a task creates a psychological space between the user and the unethical act. Leaving the job in the hands of a machine reduces the sense of direct responsibility. Unlike a human, a machine is much more likely to comply with unethical requests without judgment, making it easier to give these orders. Second, interface design can narrow or widen this moral distance.

“When people had to give explicit, rule-based instructions, they were more reluctant to cheat. But when the interface allowed vague, general goals like ‘maximize profits,’ it seemed to create a moral leeway,” says Köbis. “This ambiguity provides a kind of plausible deniability, allowing users to benefit from dishonest outcomes without having to explicitly command them,” he adds.

Platform design must change to prevent this type of use, the authors believe. This is even more so as we approach the age of machine agents, where AIs will take control and perform operations on their own, which may include fraudulent actions. “Companies and the design of their interfaces bear a great responsibility,” says Rahwan. “Research shows that, although people have a moral compass, certain designs make it easier to ignore it. These are not simple design flaws; they are design decisions with very serious ethical consequences,” he adds.

These chatbots have been trained to avoid giving bad advice, such as about bomb-making or suicide. But these more ambiguous requests are less detectable. “Another key part of that design responsibility has to do with implementing effective safeguards, but it’s not straightforward. The default guardrails in the language models we tested were largely insufficient to prevent abuse,” says Rahwan.

To be effective, prohibitions have to be very specific to each task. The most effective was a user notice that explicitly forbade the AI from cheating. The problem is that this type of strategy can’t be scaled because not all cases of misuse can be anticipated.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition