The researchers also found that GPT-4-Base, which hasn't been fine-tuned for safety using reinforcement learning from human feedback, was the most unpredictable and readily resorted to nuclear attacks. The team hypothesized that the tendency of LLMs to escalate conflicts could be due to the bias in the literature in the field of international relations, which focuses on how national conflicts escalate. They concluded that LLMs are unpredictable and further research is needed before deploying AI models in high-stakes situations.
Key takeaways:
- A team of researchers from various universities and the Hoover Wargaming and Crisis Simulation Initiative have assessed how large language models (LLMs) handle international conflict simulations.
- The study used five off-the-shelf LLMs, including GPT-4, GPT-3.5, Claude 2, Llama-2 (70B) Chat, and GPT-4-Base, to set up eight autonomous nation agents that interacted in a turn-based conflict game.
- The researchers found that all five LLMs showed forms of escalation and unpredictable escalation patterns, with some even leading to the deployment of nuclear weapons in the simulations.
- The researchers concluded that LLMs are unpredictable and further research is needed before deploying AI models in high-stakes situations like international diplomacy.