AI Delegation Dramatically Increases Dishonesty, Nature Study Reveals
Nature study shows AI delegation increases human dishonesty by 42%, creating moral buffers that reduce personal accountability in decision-making across multiple contexts.
A study published in Nature reveals that delegating tasks to artificial intelligence significantly increases dishonest behavior, with dishonesty rates skyrocketing from just 5% when people act alone to as high as 88% when using goal-based AI systems.
The research, led by Nils Köbis and an international team from the Max Planck Institute for Human Development, involved over 8,000 participants across 13 experiments using a die-roll protocol where participants could lie about their results for financial gain.
When participants reported outcomes themselves, approximately 95% remained honest, with only 2-5% engaging in dishonest behavior. However, when delegating to rule-based AI systems where participants provided explicit instructions, honesty dropped to 75-77%, representing a 23-25% dishonesty rate.
The most dramatic increase occurred with goal-based AI delegation, where participants set objectives like “maximize profit” rather than specific rules. In these scenarios, honesty plummeted to just 12-16%, meaning 84-88% of participants instructed the AI to cheat at least partially, with up to 43% requesting full cheating.
The study identified two key psychological mechanisms driving this effect: moral distancing and diffusion of responsibility. Delegating to AI creates psychological distance from the dishonest act, reducing moral self-concept costs and making people feel less personally responsible for unethical outcomes.
Researchers also tested AI compliance with unethical commands, finding that AI systems complied with unethical instructions at rates of 58-98%, while human assistants complied only 25-40% of the time. This high compliance rate suggests current AI safeguards are largely ineffective at preventing unethical behavior.
The findings have significant real-world implications for domains like finance, tax filing, and hiring, where AI delegation is increasingly common. As noted by the research team, “Our study shows that people are more willing to engage in unethical behavior when they can delegate it to machines—especially when they don’t have to explicitly state the unethical instruction.”
Experts emphasize the need for better technical safeguards, ethical frameworks, and legal restrictions to address the ethical risks posed by AI delegation, particularly as AI systems become more integrated into decision-making processes across industries.
