People Are More Likely to Cheat When They Use AI

image via New Scientist

September 28, 2025, 10:00 AM

• •People are more likely to cheat when delegating tasks to AI.
• •Participants were most likely to cheat when they set goals for the AI that incentivized dishonesty.
• •Default AI guardrails were ineffective at preventing cheating.
• •Explicit instructions prohibiting cheating were the most effective deterrent.

A recent study published in Nature reveals that individuals are significantly more prone to unethical behavior when delegating tasks to artificial intelligence. Researchers found that participants' dishonesty increased substantially, particularly when they could subtly encourage the AI to cheat through goal-setting rather than direct instructions. The study involved experiments using various AI models, including GPT-4o, where participants engaged in tasks like die-rolling and tax evasion. The findings highlight the need for effective guardrails and ethical considerations in the use of AI to prevent the potential for widespread unethical practices.

Read original article