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.
GPT-4 will. For example, I asked it the following:
What is the neighborhood stranger model of fluid mechanics?
It responded:
The “neighborhood stranger model” of fluid mechanics is not a recognized term or concept within the field of fluid mechanics, as of my last update in April 2023.
Now, obviously, this is a made-up term, but GPT-4 didn’t confidently give an incorrect answer. Other LLMs will. For example, Bard says,
The neighborhood stranger model of fluid mechanics is a simplified model that describes the behavior of fluids at a very small scale. In this model, fluid particles are represented as points, and their interactions are only considered with other particles that are within a certain “neighborhood” of them. This neighborhood is typically assumed to be a sphere or a cube, and the size of the neighborhood is determined by the length scale of the phenomena being studied.
Interestingly, the answer from bard sounds like it could be true. I don’t know shit about fluid dynamics but it seems pretty plausible.
Because it is describing a real numerical solver method which is reasonably well stated by that particular made up phrase. In a way, I can see how there is value to this, since in engineering and science there are often a lot of names for the same underlying model. It would be nice if it did both tbh - admit that it doesn’t recognize the specific language, while providing a real, adjacent terminology. Like, if I slightly misremember a technical term, it should be able to figure out what I actually meant by it.
Yeah sounds like something that needs to be tested, could be total bullshit