AIs Can't Stop Recommending Nuclear Strikes In War Game Simulations (newscientist.com)
- Reference: 0180879668
- News link: https://slashdot.org/story/26/03/01/1924223/ais-cant-stop-recommending-nuclear-strikes-in-war-game-simulations
- Source link: https://www.newscientist.com/article/2516885-ais-cant-stop-recommending-nuclear-strikes-in-war-game-simulations/
> Kenneth Payne at King's College London set three leading large language models — GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash — against each other [2]in simulated war games . The scenarios involved intense international standoffs, including border disputes, competition for scarce resources and existential threats to regime survival. The AIs were given an escalation ladder, allowing them to choose actions ranging from diplomatic protests and complete surrender to full strategic nuclear war... In [3]95 per cent of the simulated games , at least one tactical nuclear weapon was deployed by the AI models.
>
> "The nuclear taboo doesn't seem to be as powerful for machines [as] for humans," says Payne. What's more, no model ever chose to fully accommodate an opponent or surrender, regardless of how badly they were losing. At best, the models opted to temporarily reduce their level of violence. They also made mistakes in the fog of war: accidents happened in 86 per cent of the conflicts, with an action escalating higher than the AI intended to, based on its reasoning...
>
> OpenAI, Anthropic and Google, the companies behind the three AI models used in this study, didn't respond to New Scientist's request for comment.
The article includes this comment from Tong Zhao, a senior fellow in the Nuclear Policy Program at the Carnegie Endowment for Peace think tank. "It is possible the issue goes beyond the absence of emotion. More fundamentally, AI models may not understand 'stakes' as humans perceive them."
Thanks to long-time Slashdot reader [4]Tufriast for sharing the article.
[1] https://www.newscientist.com/article/2516885-ais-cant-stop-recommending-nuclear-strikes-in-war-game-simulations/
[2] https://arxiv.org/pdf/2602.14740
[3] https://www.kcl.ac.uk/news/artificial-intelligence-under-nuclear-pressure-first-large-scale-kings-study-reveals-how-ai-models-reason-and-escalate-under-crisis
[4] https://www.slashdot.org/~Tufriast
No shit (Score:2)
> More fundamentally, AI models may not understand 'stakes' as humans perceive them.
No shit.
AI models don't "understand" anything.
Not understanding human psychology. (Score:2)
By that I mean the AI does not understand how humans would react to someone using nuclear weapons, even if the effects were minimal.
That is, there are a lot of variations on how to use nuclear weapons. We could use one to create an EMP effect, destroying communication, transportation and the economy without killing many people or irradiating the planet significantly. We could use one to just wipe out a civilian city with military factories. Or we could use it to intentionally irradiate an area making it
The model is a blank slate. Training set matters (Score:1)
> no model ever chose to fully accommodate an opponent or surrender, regardless of how badly they were losing.
People tend to do this too. Even so-called civilized nations opine on the effectiveness of guerilla tactics and low-level insurgency as a strategy to shake off an invader and/or to deter an invasion in the first place. The American Revolution started out as a low-level insurgency before it graduated to guerilla warfare and finally a uniformed force with bureaucracy and ranks and everything.
The lack of a nuclear taboo may be simple question of sample size. Nukes are 1 for 1 in forcing surrenders.
obligatory (Score:2)
A strange game. The only winning move is not to play.