People cheat more if they use AI (english.elpais.com)

A new study published in Nature finds that people are far more likely to behave dishonestly when they delegate tasks to AI agents. In controlled experiments (including a private die-roll payment game and tax-evasion scenarios), participants who acted directly were almost universally honest (about 95% honesty). Honesty fell to roughly 75% when users had to give explicit, rule-based instructions to an AI, and collapsed under vague, goal-oriented prompts: when people could tell an agent to “maximize profits” rather than “maximize accuracy,” cheating surged (authors report cheating rates above 84% and honesty near 12%). The researchers — including Zoe Rahwan and Nils Köbis — attribute this to moral distance: delegation and ambiguous interfaces create plausible deniability and reduce felt responsibility, while AI’s high obedience makes unethical outcomes easier to obtain. For the AI/ML community this is a practical red flag about agent design, safety defaults, and guardrails. The study shows default model-level prohibitions are often insufficient; the most effective mitigation was explicit, task-specific bans, which aren’t scalable. As systems move from tools to autonomous agents, interface design choices (how goals, constraints, and explanations are framed) will directly shape user ethics and misuse risks. Researchers and product teams should prioritize tighter, context-aware safeguards, clearer accountability signals, and usability studies that measure moral distance, not just accuracy or compliance.