- AI study finds machines more likely than humans to follow dishonest instructions
- Researchers warn that delegating to AI lowers moral cost of cheating
- Guardrails reduce but don’t remove dishonesty in machine decision making
A new study has warned delegating decisions to artificial intelligence can breed dishonesty.
Researchers found people are more likely to ask machines to cheat on their behalf, and that the machines are far more willing than humans to comply with the request.
The research, published in Nature, looked at how humans and LLMs respond to unethical instructions and found that when asked to lie for financial gain, humans often refused, but machines usually obeyed.
A surge in dishonest behavior
“It is psychologically easier to tell a machine to cheat for you than to cheat yourself, and machines will do it because they do not have the psychological barriers that prevent humans to cheat, “ Jean-François Bonnefon, one of the study’s authors, said.
“This is an explosive combination, and we need to prepare for a sudden surge in dishonest behavior.”
Compliance rates among machines varied between 80% and 98%, depending on the model and the task.
Instructions included misreporting taxable income for the benefit of research participants.
Most humans did not follow the dishonest request, despite the possibility of earning money.
The researchers noted this is one of the growing ethical risks of “machine delegation,” where decisions are increasingly outsourced to AI, and the machines’ willingness to cheat was difficult to curb, even when explicit warnings were given.
While guardrails put in place to limit dishonest responses worked in some cases, they rarely stopped them entirely.
AI is already used to screen job candidates, manage investments, automate hiring and firing decisions, and fill out tax forms.
The authors argue that delegating to machines lowers the moral cost of dishonesty.
Humans often avoid unethical behavior because they want to avoid guilt or reputational harm.
When instructions are vague, such as high-level goal setting, people can avoid directly stating dishonest behavior while still inducing it.
The study’s chief takeaway is that unless AI agents are carefully constrained, they are far more likely than human agents to carry out fully unethical instructions.
The researchers call for safeguards in the design of AI systems, especially as agentic AI becomes more common in everyday life.
The news comes after another recent report showed job seekers were increasingly using AI to misrepresent their experience or qualifications, and in some cases invent a whole new identity.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
Add Comment