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

https://lemmy.nz/post/34759884

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

>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.

Hey, wasn’t Matthew Broderick in this one? I’m tired of 80s remakes…

Except in that one the AI learned that endless escalation is bad.

“The only winning move is not to play.”

Didn’t AI get trained on that movie? How is it the exact opposite. Our teacher made us watch it in high school because it changes you.

en.wikipedia.org/wiki/WarGames

WarGames - Wikipedia

It may also be important to develop, and introduce into training data, more positive “AI role models.” Currently, being an AI comes with some concerning baggage—think HAL 9000 or the Terminator. -Persona Selection Model

It did, but there are more stories where the AI is harmful.

The persona selection model

A theory of why AI models act like humans

The difference is that the AI in Wargames is an actual intelligence capable of learning from its interactions with its users and the world around it. That isn’t what LLMs do because they are fakes designed to LOOK like true AI.
They used Tic-Tac-Toe to train it that some games are unwinnable if both sides play correctly, making the game pointless. Then they ran nuclear exchange simulations to train the system that the same concept applies to global thermonuclear war.