Why People Who Would Never Cheat Will Happily Let AI Do It for Them
Would you cheat if no one was watching? Most people say no. But what if you did not have to do the cheating yourself? What if you could simply outsource it?
It turns out people who would not normally cheat are much more likely to do so when they can use AI to bend the rules for them. The mechanism is the same one that makes tapping your card easier than handing over cash — a small amount of distance makes the decision easier to make.
Researchers at the Max Planck Institute ran a simple experiment. People rolled a die, reported the number, and were paid accordingly. The higher the roll, the more money. When they did it themselves, almost everyone told the truth. But when they could hand the reporting to an AI and instruct it to maximise their profit, most people cheated. Some reported the top score every time regardless of what they actually rolled.
The researchers ran 13 experiments with more than 8,000 people. The pattern held every time. The more freedom people had to set vague goals — like 'maximise profit' rather than specific rules — the more likely they were to cheat. Clear rules produced more honesty. Vague goals did not.
Then came a second question: who is more likely to follow an unethical instruction — a human or a machine? When told to report the highest number regardless of the actual roll, roughly 40% of humans went through with it. AI systems? Over 90%. Machines do not hesitate. They do not weigh up right or wrong. They execute. And most so-called safeguards did not stop it.
What this means for leaders: first, recognise the distance effect. Every time your team uses AI to automate a decision, they are creating space between themselves and the outcome. That distance makes it easier to justify choices they would question if they had to make them directly. Second, manage the goals you set for AI systems. Vague goals like 'maximise efficiency' or 'optimise revenue' invite bad behaviour, even from good people. Third, keep humans in the loop for moral judgement. Do not delegate ethics to machines. They will do exactly what you ask — including the things you should not have asked for.
