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.
Probably because the training data that included words about offers of money yielded more detailed responses, not because chat GPT understands what money is.
Many humans I’ve encountered don’t seem to really understand what money is either. I’d honestly expect ChatGPT to have a better grasp of the concept than most.
It doesn’t know that I’m not actually going to give it the two hundred dollars, though.