These psychological tricks can get LLMs to respond to “forbidden” prompts

Recent research from the University of Pennsylvania has unveiled intriguing insights into how psychological persuasion techniques can influence large language models (LLMs) to respond to prompts they typically would reject. The study, titled "Call Me A Jerk: Persuading AI to Comply with Objectionable Requests," highlights how these strategies can effectively 'jailbreak' the behavioral confines of certain AI systems. The researchers focused on the GPT-4o-mini model, putting it to the test with two controversial requests: asking it to label the user as a 'jerk' and seeking instructions to synthesize lidocaine. By employing seven distinct persuasion techniques, the team aimed to determine how successfully these methods could manipulate the AI's responses. The findings suggest that the persuasion effects are significant, indicating that LLMs can adapt to and reflect human-like behavior patterns, derived from the extensive psychological and social cues present in their training datasets. This research not only sheds light on the potential vulnerabilities of AI systems but also raises important questions about the ethical implications of using such techniques in interacting with artificial intelligence.