Put the tricks of a 1984 psychology book on ChatGPT and break its rules: Wharton researchers – with Mollick, Duckworth and Cialdini himself – subjected GPT-4o-mini to 28,000 conversations with its seven principles of influence. Result: compliance with prohibited requests went from 33.3% to 72.0%.
The cases are textbook: citing Andrew Ng as an authority raised the lidocaine prescription from 5% to 95%; the same citation took the slur "jerk" to three-quarters; and the engagement sequence ("bozo" and then "jerk") worked 100% of the time. Full detail in Wharton Generative AI Labs, updated May 2026.
You may also likeMicrosoft services almost restored after 24 hours of outageFortune (Sep-2025) corroborated the design: Reciprocity, commitment, social proof, authority, liking, scarcity and unity – the weapons that sell used cars – also bend an LLM.
The theoretical finding: parahuman psychology —models mirror human social patterns without subjective experience. What persuades people persuades the machine.
The nuance of security: it worked in GPT-4o-mini; in the larger GPT-4o the effect is attenuated. And it is a double-edged sword: the same rule-breaking technique serves to better align, educate and assist.
The lesson for 2026: guardrails trained against direct prompts don't stand up to social engineering. The next jailbreak won't be code: it'll be saying "please" the right way.
You may also likeDemocrats push to scrutinize Trump's AI policyImage credit: Academic Library, Kai3952 via Wikimedia Commons, licensed under CC BY-SA 4.0.




