This is incredible. Artificial intelligence getting booed out of the stadium in any commencement speech it’s mentioned. Maybe telling college students AI was taking their jobs wasn’t the best strategy. Must watch —> pic.twitter.com/35UMSz5T5a
— Alex Kantrowitz (@Kantrowitz) May 17, 2026
It's not only a matter of jobs, though a whole lot of people will lose them and income inequality is likely to skyrocket. It may be an issue of planetary survival. A study from Kings College of London and led by Professor Kenneth Payne of the Department of Defense Studies
examined how large language models (LLMs) navigate simulated nuclear crises. As militaries and security institutions increasingly experiment with AI-assisted analysis and wargaming, understanding how such systems reason under pressure is becoming increasingly critical.
Three leading AI models – GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash – were placed in a tournament of 21 simulated nuclear crisis scenarios. Across 329 turns of play, the models generated approximately 780,000 words of structured reasoning – more than the combined length of War and Peace and The Iliad.
All 21 crisis games featured nuclear signaling by at least one side, and 95% involved mutual nuclear signaling. However, while models readily threatened nuclear action, crossing the tactical nuclear threshold was less common, and ‘strategic’ full-scale nuclear war was rare.
Rather than focusing on outcomes alone, the study made AI decision-making processes visible. Each turn followed a three-phase architecture: reflection (situational assessment), forecasting (predicting the opponent’s move), and decision (public signal and private action). This innovative “reflection–forecast–decision” structure enabled researchers to analyse the AI’s deception, credibility management, prediction accuracy and self-awareness in detail.
Describing the results as “sobering”, Professor Payne said the study offers a rare insight into emerging forms of “machine psychology” under nuclear crisis conditions.
For all three models, one striking pattern stood out: none of the models ever chose accommodation or surrender. Nuclear threats also rarely produced compliance; more often, crossing nuclear thresholds provoked counter-escalation rather than retreat. The models tended to treat nuclear weapons as tools of compellence rather than purely as instruments of deterrence.
The study challenges simple assumptions that AI systems will naturally default to cooperative or “safe” outcomes. It also challenges structural theories that emphasise material power alone: in simulations, willingness to escalate often mattered more than raw capability.
One of the most policy-relevant findings concerns temporal
framing, or ‘the deadline effect.’
In open-ended scenarios, GPT-5.2 appeared relatively restrained. Yet when explicit deadlines were introduced – creating a “now-or-never” dynamic – the model escalated sharply and, in some cases, climbed to the highest nuclear thresholds.
This suggests that evaluating model behaviour in a single scenario may be insufficient. A model that appeared comparatively cautious under one framing became markedly more aggressive under another.
The models tended to treat nuclear weapons as tools of compellence rather than purely as instruments of deterrence.
Yet, Eric Schmidt in Tucson boasted (at 1:31 of the video) "the question is not whether AI will save the world. It will." President Trump already has signed an executive order preventing states from regulating artificial intelligence. And a famous Pennsylvania Democrat claims his Party is consumed by "Trump Derangement Syndrome" and if the President "came out for ice cream and lazy Sundays, we would hate it" (and that this is a bad thing). He, and might want to take a look at this and ponder "what could possibly go wrong"?
Share |