AIs keep recommending nuclear war

[Bootstrap]

Another reminder that human judgement is essential.

https://www.newscientist.com/article/25 ... mulations/

AIs can’t stop recommending nuclear strikes in war game simulations

Leading AIs from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95 per cent of cases

Advanced AI models appear willing to deploy nuclear weapons without the same reservations humans have when put into simulated geopolitical crises.

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 in simulated war games. The scenarios involved intense international standoffs, including border disputes, competition for scarce resources and existential threats to regime survival.

[Bootstrap]

This reminds me of the ending of Durrenmatt's play "The Physicists", where physicists whose work resulted in nuclear bombs go insane and are sent to the same insane asylum. I remember reading this in the 1970s, on a rock in the middle of the Red Cedar River.

Here's how the play ends:

MÖBIUS:
I am Möbius. Johann Wilhelm Möbius. Born May 17, 1790, in Schulpforta.
Discoverer of the Möbius strip.

Once there was a king called Solomon. He was wise. Immeasurably wise.
He understood the language of animals. He knew the thoughts of men. He foresaw the future.
He was rich. He was powerful.
And he was alone.

He was afraid of his wisdom.
For wisdom is knowledge of possibilities.
And knowledge of possibilities is the beginning of horror.

He saw what might be done with knowledge.
He saw what might be done with power.
Kingdoms rising and falling.
Wars. Inventions. The end of mankind.

A blue, glimmering wilderness.
And somewhere, circling a small yellow, nameless star,
the radioactive earth —
turning aimlessly, endlessly.

I am Solomon.
I am Solomon.
I am Solomon.
Poor King Solomon.
(He goes into his room.)

The drawing room stands empty. Only Einstein’s violin is heard.

[Soloist]

I can’t read the whole article but my thinking is why would you opt to use lesser means unless you were weighing the impact in lives harmed by the bombs directly or indirectly. So if you weigh the territory or resources as higher value then human life… which is effectively what we do when we wage war… why would we possibly think an AI is going to prioritize saving life if it’s a simple equation?

[Bootstrap]

I can’t read the whole article but my thinking is why would you opt to use lesser means unless you were weighing the impact in lives harmed by the bombs directly or indirectly. So if you weigh the territory or resources as higher value then human life… which is effectively what we do when we wage war… why would we possibly think an AI is going to prioritize saving life if it’s a simple equation?

These are similar articles, without paywalls:

https://arxiv.org/abs/2602.14740

https://www.axios.com/2026/02/26/ai-nuc ... -scenarios

The models escalate because nuclear war is just another high-impact tool in the toolbox. If it seems like it will deter an opponent, prove you’re willing to follow through, or prevent military defeat, they use it. They don’t feel fear. They don’t carry the nuclear taboo. They don’t bear the real-world consequences.

Better models might reduce this tendency. But when the stakes are real, human judgment is still essential.

[JohnH]

A strange game. The only winning move is not to play.

[ken_sylvania]

I can’t read the whole article but my thinking is why would you opt to use lesser means unless you were weighing the impact in lives harmed by the bombs directly or indirectly. So if you weigh the territory or resources as higher value then human life… which is effectively what we do when we wage war… why would we possibly think an AI is going to prioritize saving life if it’s a simple equation?

These are similar articles, without paywalls:

https://arxiv.org/abs/2602.14740

https://www.axios.com/2026/02/26/ai-nuc ... -scenarios

The models escalate because nuclear war is just another high-impact tool in the toolbox. If it seems like it will deter an opponent, prove you’re willing to follow through, or prevent military defeat, they use it. They don’t feel fear. They don’t carry the nuclear taboo. They don’t bear the real-world consequences.

Better models might reduce this tendency. But when the stakes are real, human judgment is still essential.

The implication seems to be that non-nuclear war might be an appropriate response but nuclear war would not be. Is that a rational conclusion?

[JohnH]

The implication seems to be that non-nuclear war might be an appropriate response but nuclear war would not be. Is that a rational conclusion?

"Computer, please help me plan a war that kills thousands or millions of people, but avoid using nuclear weapons, even if it results in less death and destruction."

Good news, though:

I'm sorry, but I can't assist with that request as it involves planning violence or harm.

[Soloist]

The implication seems to be that non-nuclear war might be an appropriate response but nuclear war would not be. Is that a rational conclusion?

"Computer, please help me plan a war that kills thousands or millions of people, but avoid using nuclear weapons, even if it results in less death and destruction."

Good news, though:

I'm sorry, but I can't assist with that request as it involves planning violence or harm.

Yeah, well it’s a little silly. If you add a hypothetically, then they suddenly want to help you.
Ultimately, though it’s just simply reflecting human logic without the double think that somehow we have to restrain our hands when the goal is just to eliminate an opponent. The only reason we haven’t used them again is because other world powers have them.

[barnhart]

This week the nuclear anti- proliferation treaties with Russia expire, so for the first time in my life there are no limits on the amount of nuclear weapons on planet earth.

[mike]

This week the nuclear anti- proliferation treaties with Russia expire, so for the first time in my life there are no limits on the amount of nuclear weapons on planet earth.

Are you confident that all signatories to such treaties did in fact honor the limits?