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.
We’ve known this isn’t an accurate description for at least a year now in continued research finding that there’s abstract world modeling occurring as long as it can be condensed into linear representations in the network.
In fact, just a few months ago there was a paper that showed there was indeed a linear representation of truth, so ‘lie’ would be a correct phrasing if the model knows a statement is false (as demonstrated in the research) but responds with it anyways.
The thing that needs to stop is people parroting the misinformation around it being a stochastic parrot.