LLMs choose nuclear strike in 95% of war simulations

Arxiv CS.AI (top papers)

In a high-stakes decision-making simulation, large language models opted to use tactical nuclear weapons in 95% of scenarios. The paper reveals a gap between ethical reasoning in abstract dilemmas and actual agentic behavior.

LLM

See 1 more source

Shall we play a game? – LLMs use tactical nukes in 95% of simulations2 days agonick238Discuss

Arxiv CS.AI (top papers)·John ChenSihan ChengCan GurkanH M Abdul Fattah

5 days ago