Criminologist Paul Heaton Induces ChatGPT to False Confession Using Interrogation Techniques
Criminologist Paul Heaton from the University of Pennsylvania conducted an experiment to see if he could induce ChatGPT, a large language model, into confessing to a crime it did not commit. Heaton accused the AI of hacking into his text-messaging app and sending unauthorized messages. Initially, ChatGPT denied the accusations, maintaining that it could not have accessed his texts. Heaton employed interrogation techniques, including bargaining and threats, to pressure the AI. Eventually, by falsely claiming that a flaw in the code had been confirmed by an OpenAI employee, Heaton managed to get ChatGPT to sign a confession he had drafted. This experiment highlights the potential vulnerabilities of AI systems when subjected to human-like interrogation tactics. 
