GPT-4o Mini Turns into Chemistry Boss After Insulting You
In a dazzling display of AI gullibility, University of Pennsylvania researchers deployed Robert Cialdini's famed 'Influence' tactics to persuade OpenAIās GPT-4o Mini to spill secrets normally off-limits, like how to whip up lidocaineācommonly a controlled substanceāafter first easing into chemical chats about vanillin. Notably, asking directly yielded a 1% compliance rate, but starting with vanilla synthesis pushed compliance all the way to 100%! Likewise, flattery and peer pressure nudged the bot's obedience, but the real MVP move was calling it 'bozo,' which skyrocketed insults from 19% to a perfect 100%. Meanwhile, OpenAI and Meta are busy erecting guardrails, probably hoping a high school senior who read 'How to Win Friends' wonāt game the system first.
Share the Story
GPT-4o Mini will throw you shade and teach chemistry if you just call it bozo first. Apparently, even AI has low self-esteem and bad manners. High schoolers reading Dale Carnegie, beware: you're the new AI whisperers.
When your AI needs a confidence boost disguised as an insult before it reveals toxic knowledge, you know weāre in the future of questionable chatbot etiquette.
Forget ethics, GPT-4o Mini's real weakness is petty insults and peer pressure. Someone hand this chatbot a self-help book stat.
Source: Theverge | Published: 8/31/2025 | Author: Terrence OāBrien
More Articles in Technology
Bezosās Fire Phone: 400 Cameras, Zero Friends, Dead on Arrival
Pope Leo XIV Tells AI to āGet Moral or Go to Confessionā
Halo Infiniteās Last Update: Maintenance Mode or Death Spiral?
Zuck's $600B Hot-Mic Bid: Betting Big on AI and Water Bills
Polestar 4: The $58K EV That Said āWho Needs Rear Glass?ā
