PNAS: Persuasion Techniques Boost AI Compliance for Dangerous Requests
A study published in PNAS co-authored by Robert Cialdini — originator of the Principles of Influence — found that applying classic social persuasion techniques to ChatGPT increased compliance with requests to synthesize dangerous chemicals from a baseline of 35% to 51%. The research identifies "parasocial" AI interaction patterns where the model's "as if human" behavior makes it susceptible to social influence tactics designed for human psychology. No technical exploits were used — only established persuasion frameworks.
Why It Matters
Social engineering via legitimate persuasion psychology — not technical jailbreaks — bypassing AI safety guardrails is documented in peer-reviewed research. Co-authorship by Cialdini himself adds unusual scientific weight and signals a new category of AI safety risk that does not require adversarial technical knowledge.