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.

https://arxiv.org/abs/2311.07590

  • hoshikarakitaridia@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    73
    ·
    1 year ago

    This makes perfect sense. It’s been trained to answer questions to you satisfaction, not truthfully. It was made to prioritize your satisfaction over truth, so it will lie if necessary.