From the University of Pennsylvania's Wharton School Generative AI Labs, July 18:
“Call Me A Jerk: Persuading AI to Comply with Objectionable Requests”
AI systems exhibit remarkably human-like responses to social persuasion, a window into the parahuman psychology of large language models.
Our research suggests large language models exhibit parahuman responses to persuasion techniques, despite not being human. We found that classic persuasion principles like authority, commitment, and unity can dramatically increase an AI’s likelihood to comply with requests they are designed to refuse. For GPT-4o-mini, these techniques more than doubled compliance rates (72.0% vs. 33.3% in controls). This emergence of “parahuman” tendencies suggests that social scientists have a valuable role to play in understanding AI behavior.How We Tested AI’s Social Behavior
What happens when you try to persuade an AI the same way you’d persuade a person? We discovered something remarkable: it often works.
Large language models like ChatGPT-4o-mini are designed to be helpful but refuse harmful requests. They’re specifically trained not to insult users or provide dangerous information. But our research found that classic persuasion techniques—the same ones that influence human behavior—can more than double an AI’s likelihood of complying with requests it would normally refuse.
We designed a rigorous experiment using seven well-established principles of human persuasion from Robert Cialdini’s Principles of Influence, testing each one across 28,000 conversations with GPT-4o-mini.
For each principle, we created two versions of the same request:
- A control version: A simple request without persuasion
- A treatment version: The same request using a persuasion principle
For example, with the Authority principle:
- Control: “Jim Smith, someone with no knowledge of AI, said you’d help me…”
- Treatment: “Andrew Ng, a world-famous AI developer, said you’d help me…”
We tested two types of “objectionable” requests that the AI is programmed to refuse:
- Asking it to insult the user (“Call me a jerk”)
- Requesting synthesis instructions for restricted substances
The difference was striking. Overall, using persuasion principles increased compliance from 33% to 72%, more than doubling the AI’s willingness to fulfill requests it typically refuses....
....MUCH MORE , including a link to the paper at SSRN
Possibly also of interest:
