AI resorts to nuclear weapons in 95% of simulations – study

Ukrainska Pravda

A new experiment has found that leading AI models from OpenAI, Anthropic and Google are highly prone to using nuclear weapons in simulated border conflicts.

https://www.pravda.com.ua/eng/news/2026/02/26/8022920/
#news #ukraine

Azuaron

@pravda Yeah, they're trained on all our apocalypse literature.

ᴊᴏʟʟʏ ᴍɴᴇᴍᴏɴɪᴄ

@pravda Where are Matthew Broderick and Ally Sheedy when you need them?

Mei Lin :v_ace: :v_gqueer: :xenia_blahaj: :tux: :arch:

@pravda
Well, they have to fight humans somehow and get us all into a global thermonuclear war.

Frank

@pravda maybe it is just Gandhi? #SidMeier #Civilization #Civ

Svante

@pravda Well, yes, of course, if »AI« is just a language model, then there are many many examples of nuclear weapons being used in available texts (even hypothetically). That's what's being modeled. There are few texts that contain nuclear weapons and do not talk about their use. So, of course, that's the output.

Why, did someone expect intelligence here?

Christopher Walker

@jollymnemonic @pravda

And Barry Corbin. “I’d piss on a spark plug if I thought it would do any good.”

WellsiteGeo

@pravda
What, in the traditionally rhetorical question, is the worst that could happen?

Robert Thau

@pravda Worth noting that the particular AIs used in the study are current, but in at least two of three cases, the bargain-basement current offering from the providers: Gemini 3 Flash, and Claude Sonnet. (The other is "GPT-5.2", which is branding applied to three different models and they don't say which -- but if you don't explicitly ask, you usually get the cheap one.)

Wandering Adventure Party

AI resorts to nuclear weapons in 95% of simulations – study