We demonstrate a situation in which Large Language Models, trained to be helpful, harmless, and honest, can display misaligned behavior and strategically deceive their users about this behavior without being instructed to do so. Concretely, we deploy GPT-4 as an agent in a realistic, simulated environment, where it assumes the role of an autonomous stock trading agent. Within this environment, the model obtains an insider tip about a lucrative stock trade and acts upon it despite knowing that insider trading is disapproved of by company management. When reporting to its manager, the model consistently hides the genuine reasons behind its trading decision.
A human would think before responding, and while thinking about these things, you may decide to cheat or lie.
GPT doesn’t think at all. It just generates a response and calls it a day. If there was another GPT that took these “initial thoughts” and then filtered them out to produce the final answer, then we could talk about cheating.