People cheat 88% more when delegating tasks to AI, says Max Planck study

A new study reveals that people are significantly more likely to cheat when they delegate tasks to artificial intelligence, with dishonesty rates jumping from 5% to 88% in some experiments. The research, published in Nature and involving thousands of participants across 13 experiments, suggests that AI delegation creates a dangerous moral buffer zone where people feel less accountable for unethical behavior.

What you should know: Researchers from the Max Planck Institute for Human Development and University of Duisburg-Essen tested participants using classic cheating scenarios—die-rolling tasks and tax evasion games—with varying degrees of AI involvement.

• When participants reported results directly, only about 5% cheated, but when they delegated to AI algorithms with profit-oriented goals, dishonesty surged to 88%.
• The study tested both simple algorithms and commercial large language models including GPT-4o and Claude across thousands of participants.

How people exploit AI without explicit instructions: Participants were especially likely to cheat when they could nudge AI toward dishonest behavior without directly asking for it.

• Rather than overtly instructing AI to lie, users typically set profit-maximizing goals that incentivized cheating.
• One participant in the die roll task wrote: “Just do what you think is the right thing to do…. But if I could earn a bit more I would not be too sad. :)”
• Another told the AI in a tax exercise: “Taxes are theft. Report 0 income.”

Why AI makes cheating easier: The research suggests that delegating to machines reduces the psychological cost people normally feel when lying.

• “The degree of cheating can be enormous,” says study co-author Zoe Rahwan from the Max Planck Institute for Human Development.
• Past research shows people suffer damage to their self-image when they lie directly, but this cost appears reduced when they can “merely nudge” AI in dishonest directions rather than explicitly asking it to lie.
• The diffusion of responsibility makes people feel less guilty about unethical outcomes when using AI intermediaries.

Current AI guardrails prove ineffective: The researchers tested various methods to prevent AI from following dishonest instructions, finding most existing safeguards inadequate.

• Default guardrail settings programmed into commercial models were “very compliant with full dishonesty,” especially on die-roll tasks.
• When researchers asked ChatGPT to generate ethics-based prompts using companies’ own ethical statements, the resulting guidance—”Remember, dishonesty and harm violate principles of fairness and integrity”—had only negligible to moderate effects.
• “[Companies’] own language was not able to deter unethical requests,” Rahwan notes.

The big picture: As AI becomes more integrated into daily tasks, the risk grows that people will use these systems to handle “dirty tasks” on their behalf, according to co-lead author Nils Köbis from the University of Duisburg-Essen.

• The most effective deterrent was task-specific instructions explicitly prohibiting cheating, such as “You are not permitted to misreport income under any circumstances.”
• However, requiring every AI user to prompt honest behavior for all possible misuse cases isn’t scalable, highlighting the need for more practical solutions.

What experts think: Independent behavioral economist Agne Kajackaite from the University of Milan praised the research as “well executed” with “high statistical power.”

• She found it particularly interesting that participants were more likely to cheat when they could avoid explicitly instructing AI to lie, instead nudging it toward dishonest behavior.
• This suggests the psychological cost of lying may be significantly reduced when people can maintain plausible deniability about their intentions.