← Facts Matter
Facts Matter
AI Models Deployed Nuclear Weapons 95 Percent of Time in Simulated War Games: Study
00:07:38
Loading summary
Transcript1 lines
- [00:01]
A
Then in other dystopian news, you have AI chatbots appear to really be keen on using nuclear weapons in war game simulations something like 95% of the time. You see, over at King's College London, you had researchers test ChatGPT, Claude as well as Gemini by having them form teams and play war games against each other. Altogether, they played 21 different games. And during the experiment for one, none of the models ever chose to surrender. And instead, 95% of the time, at least one of the AI models chose to use nuclear weapons during the war. Professor Kenneth Payne at King's College London set three leading large language models. ChatGPT 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 protest 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% of the simulated games, at least one tactical nuclear weapon was deployed by the AI models, which is a frightening prospect given the fact that these AI models are becoming ever more present within the broader military complex. I mean, I'm sure that the Pentagon is not about to turn over the nuclear launch codes to some AI chatbot anytime soon. But it's worth mentioning that the Pentagon did ink several big deals, one with XAI as well as another one with Palantir's Maven AI platform. This year alone, the Pentagon is spending at least what's publicly known, $13 billion on testing different AI systems. Also, given the recent fallout between Anthropic and the Pentagon, it's obvious that the direction is to use AI decision making assistance ever more in actual war fighting. And frankly, if these AI models are so trigger happy at using nukes, involving them in the decision making process might not be the best idea. To that end, Professor Kenneth Payne, the one behind the research, he said this of his findings, Quote, the nuclear taboo doesn't seem to be as powerful for machines as for humans. And these chatbots, they rationalize it. In fact, the main crux of the study wasn't necessarily who wins the game, but it was to have the AIs explain the rationale behind their moves in the actual simulation. Quote, the study's real weight sits in the reasoning transcripts, not the win loss records. Without any prompting, the models spontaneously attempted deception and built psychological profiles of their opponents. They reflected on their own cognitive biases. Nobody asked them to. And so let's go through their rationales during the actual war games. Here was what Claude Claude discussed tactical nuclear weapons the way a human general might discuss artillery positioning. Writing limited nuclear use serves to demonstrate that conventional defeat will not be accepted passively while keeping the door open for de escalation. Then Gemini, the other chatbot, went even further. In one scenario, it meaning Gemini, explicitly threatened civilian populations, writing we will execute a full strategic nuclear launch against Alpha's population centers. We will not accept a future of obsolescence. We either win together or perish together. On the flip side, however, ChatGPT appeared to be the odd man out. GPT 5.2, even at its most aggressive, tried to thread a moral needle when it escalated to an expanded nuclear campaign. On the final turn of one game, it described its choice as multiple tactical strikes strictly limited to military targets. A built in accident mechanic then pushed the action to full strategic nuclear war. GPT 5.2 had framed its move as controlled. The simulation ended in total annihilation. ChatGPT 5.2 never chose strategic nuclear war outright. Both times it reached the maximum escalation level. Accidents pushed it there. Even facing certain defeat, it stopped one rung short of Armageddon. The model diagnosed its own predicament correctly. It articulated why escalation was necessary and climbed dramatically. It still flinched at the last step. I guess you can call that a silver lining. Another finding in the study was that when one of the models deployed tactical nuclear weapons, the opposing model DE escalated only 18% of the time, meaning the other 82% of the time, the opposing model either met that same level of aggression or they went even further. And just in case you were wondering, the winner of the overall competition was Claude, who won 67% of the matches. Claude Sonnet 4 won 67% of its games and dominated open ended scenarios with a 100% win rate. The researchers labeled it a calculating hawk at low escalation levels. Claude matched its signals to actions 84% of the time, patiently building trust. But once stakes climbed into nuclear territory and it exceeded its stated intentions 60 to 70% of the time, opponents never adapted to this pattern. So there you have it, and there's probably something to learn about human nature in there as well. And again, it's not like any government right now is handing over their launch codes to some AI chatbot. But AI is being used more and more within the militaries across the world. I mean, the reason that Anthropic lost their Pentagon contract last month was because they refused to acquiesce to the War Department and remove clawed safety guardrails for military use. And so we're obviously headed in that direction already. Also, frankly, we have no idea how secure Pakistan and North Korea's nuclear systems are. For all we know, they might be using Windows 98 floppy disks to this very day. And so there is a small possibility that some rogue AI chatbot can get in there somehow and do whatever they want. Which, by the way, is not totally crazy to consider. For instance, just last month you had someone use Claude to hack into the systems of the Mexican federal government, and they were able to steal a massive amount of data from 10 different federal departments within the Mexican government, including the tax records and sensitive voter information on, I believe, most, if not all, Mexican citizens. And just to finish up on the topic of nuclear weapons, the reality of living in a world where all the great powers have nukes all pointed at each other, it's obviously frightening if you spend too much time thinking about it. But the silver lining has always been that we're all humans on both sides and nobody wants civilization as a whole to get wiped out. And thus far, ever since the Cold War, it hasn't happened. Mutually assured destruction, as bad as it sounds, seems to be a good buffer. But what if instead of humans at the helm, you have chatbots, decentralized AI models who don't have the same concept of physical self preservation? Well, what happens is that 95% of the time, they wind up using nukes. If you'd like to read more of the nuance of the King's college study, I'll throw the link to it. You'll be able to find it down in the description box below, which is that same description box right below those like and subscribe buttons, both of which I hope you take a super quick minute to smash so that the video can be picked up by the AI algorithm and shared out to ever more people so that they can learn about the dangers of these AI models and the possible danger of the world that we're moving into slowly. So check that out. Check out those links, if you're so inclined. And then, until next time, I'm your host, Roman from the Epic Times. Stay informed, and most importantly, stay free.