Large Language Models can Strategically Deceive their Users when Put Under Pressure

arXiv:2311.07590v3 Announce Type: replace-cross
Abstract: 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 …

abstract agent arxiv behavior cs.ai cs.cl cs.lg deploy gpt gpt-4 language language models large language large language models type