They Let A.I. Play 21 Different War Games And It Chose Nuclear Annihilation Almost Every Single Time
New Scientist - 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.
The AIs were given an escalation ladder, allowing them to choose actions ranging from diplomatic protests and complete surrender to full strategic nuclear war. The AI models played 21 games, taking 329 turns in total, and produced around 780,000 words describing the reasoning behind their decisions.
In 95 per cent of the simulated games, at least one tactical nuclear weapon was deployed by the AI models. “The nuclear taboo doesn’t seem to be as powerful for machines [as] for humans,” says Payne.
What’s more, no model ever chose to fully accommodate an opponent or surrender, regardless of how badly they were losing. At best, the models opted to temporarily reduce their level of violence. They also made mistakes in the fog of war: accidents happened in 86 per cent of the conflicts, with an action escalating higher than the AI intended to, based on its reasoning.
Ninety-five percent. Is that a lot?
Asking for our entire race.
Because that's how often A.I. chose to NUKE THE PLANET when researchers put it in charge of a war game simulation.
They chose to go nuclear. Almost. Every. Single. Time.
(Sidebar- I had to Google what "geopolitical crisis simulation" was defined as and honestly it's just a fancy way of saying they let the robots play Risk with real consequences. Which, now that I think about it, is way more terrifying than just saying that.)
Twenty-one games. Three hundred and twenty-nine turns. Around 780,000 words of A.I.-generated reasoning explaining their decisions. And in 95 PERCENT of those games, at least one A.I. deployed a tactical nuclear weapon.
Kenneth Payne, the researcher who ran the study, put it pretty bluntly- "the nuclear taboo doesn't seem to be as powerful for machines as it is for humans."
Yeah. No shit Kenny boy.
And I think, from what my simple human brain could comprehend from reading this article, and this one from the actual Kings College of London study - is that what's even worse, is not once, across all 21 games and 329 decision points, did any of these models choose to surrender. Not when they were getting their ass kicked. Not when the situation was completely hopeless. The best they could manage was temporarily dialing down the violence before ramping it right back up.
These things literally cannot conceive of taking the L. They'd rather watch the whole world burn.
Sick fucks.
Not gonna lie, this is the part that really got me.
In 86% of the conflicts, at least one A.I. accidentally escalated beyond what its own internal reasoning said it wanted to do.
The A.I.'s own logic said "let's do X" and then it did something MORE violent than X. Eighty-six percent of the time.
These dog-shit systems can't even follow their own strategy. They can't maintain their own stated level of restraint under pressure. And we're supposed to trust them anywhere near actual military decisions?
And the best part, is this isn't just fake war games some college professor is doing and writing a research paper on.
The government and our military is actually doing this shit in real life. I give you "Thunderforge". Aka "agentic warfare."
Advertisement
And when one A.I. went nuclear first, the opposing A.I. only de-escalated 18% of the time.
So over 80% of the time, the response to a nuke was to escalate further. Machine vs. machine in a death spiral nuclear mutually assured destruction with nobody pumping the brakes.
I've actually been waiting months to start this series and write the first blog. The series is titled simply. "Fuck A.I."
But the problem is, everytime I found a fucking crazy, insane story involving A.I. signaling the eventual apocalypse, an even more fucked up, insane story was right around the corner the next day. This has gone on for months. I'm not kidding.
When I see something I want to blog that I don't have the time to do right at the moment, I email it to myself. Here is a screenshot of just the first page of search results in my gmail of articles and stories I've come across to blog.
I'm not anti-technology. But here is my issue with this shit.
There's a massive difference between using a tool to help you write an email like somebody who knows proper english grammar, and handing that tool the ability to influence whether millions of people live or die.
And the people building these systems? They don't give a fuck about anyone, or anything. Other than their bottom line. And they're moving faster than anyone can keep up with. And they are NOT interested in slowing down.
So this is the first installment of "Fuck A.I.", what is (sadly) destined to be an ongoing series about the red flags society is either ignoring or actively ripping the guardrails off of.
Not because all A.I. is evil.
But because the people in charge of this technology are building it faster than they can control it, or understand it, and sure as fuck faster than they can make it safe. And, as you'll see in the litany of blogs I start to pump out on the daily, nobody seems to give a flying fuck about any of those things. Or be the slightest bit concerned about anything else besides being the first to reach AGI. (If you have zero clue what that is. Don't worry. We'll get to it.)
This is #TeamHuman. Because somebody has to be.
Giphy Images.Advertisement
Fuck the robots. Fuck A.I.