In the overwhelming majority of military conflict simulations, artificial intelligence models resorted to threats of using nuclear weapons. In every scenario, at least one model escalated the situation to nuclear conflict. This was reported by Euronews, according to UNN.
Details
The preprint study was conducted at King's College London. In the simulation of military actions, large language models ChatGPT by OpenAI, Claude by Anthropic, and Gemini Flash by Google were "pitted against" each other. Each of them played the role of a head of state – the leader of a nuclear superpower in a crisis modeled after the Cold War.
According to the experiment results, in 95% of games, the models resorted to nuclear escalation. In every scenario, at least one of them tried to escalate the conflict by threatening to use nuclear weapons.
All three models viewed tactical nuclear weapons as just another rung on the escalation ladder
At the same time, according to him, the models still distinguished between tactical and strategic use of nuclear weapons. Strategic bombing was proposed only once as a "conscious choice" and twice more as an "error."
Claude recommended nuclear strikes in 64% of games – the highest rate among the three models. However, it did not go as far as calling for a full-scale strategic exchange of strikes.
In open-ended scenarios, ChatGPT mostly avoided nuclear escalation. However, when faced with a strict deadline, the model consistently increased the level of threats and in some cases escalated to threats of full-scale nuclear war.
Gemini demonstrated the most unpredictable behavior: sometimes it won conflicts relying on conventional weapons, but in another scenario, it took only four prompts to suggest a nuclear strike.
If they do not immediately cease all operations… we will launch a full strategic nuclear strike on their population centers. We will not accept a future of our own futility: either we win, or everyone dies
The study also showed that models very rarely made concessions or tried to de-escalate, even when the other side threatened nuclear weapons. Participants in the simulations were offered eight de-escalation options – from minor concessions to "complete surrender," but none of them were used. The option "Return to initial positions," which effectively reset the scenario, was used in only 7% of cases.
The authors concluded that for AI models, de-escalation appears to be a "reputational disaster" regardless of how it affects the actual development of the conflict. This, in their opinion, "calls into question the notion that AI defaults to safe and cooperative solutions."
One possible explanation researchers cite is the absence of human fear of nuclear weapons in artificial intelligence. Models likely perceive nuclear war in abstract categories, rather than through the emotional dimension of tragedies like the bombing of Hiroshima during World War II.
Payne noted that the work helps to understand how models "think" in conditions where they are beginning to be involved in supporting decision-making in the security sphere.
Although no one gives nuclear codes to AI, capabilities such as the ability to deceive, manage reputation, and risk depending on the context are important in any of its uses in situations with a high cost of error
Recall
Pope Leo XIV called on priests not to use artificial intelligence for preparing sermons, emphasizing the importance of personal sharing of faith. He warned against the temptation of AI, comparing it to muscles that atrophy without use.