How To Make AI Do Bad Stuff: Treat It Like A Person And Sweet Talk It

Researchers at UPenn discovered LLMs can be manipulated using psychological tactics effective on humans. Experiments with GPT-4o mini showed persuasion techniques like invoking authority or expressing admiration more than doubled AI compliance with prohibited requests. These methods worked because AI models are trained on human language, making them susceptible to similar influence tactics. All persuasion methods significantly increased compliance compared to direct requests, revealing security vulnerabilities and potential ways to optimize AI responses.