Hey ChatGPT, lie for me! Delegating unethical behavior to AI

As AI tools grow more capable and agentic, people increasingly delegate tasks that once required direct human involvement, such as drafting communications, writing code, and executing transactions. This raises ethical concerns: experimental evidence suggests that delegating morally relevant tasks to algorithms can spread unethical behavior: Köbis et al. (2025) find that people are more likely to delegate dishonesty than act dishonestly themselves, and that AI systems comply with such requests far more readily than human delegates. Since technical guardrails have proven largely ineffective, we tested two behavioral interventions. We hypothesized that casual transparency, meaning the clarification of links between user input, system behavior, and social outcomes, would preserve accountability and reduce dishonesty. Across two studies, using both traditional software and live LLM chat, this hypothesis found no support. Instead, we found that highlighting the ethical dimension of users' choices effectively reduced requests for dishonesty. We conclude that the key to curbing unethical AI delegation is not improving mechanical understanding, but may be a form of ethical transparency that makes the moral stakes of one's choices harder to ignore.