“There was little sense of horror or revulsion at the prospect of all out nuclear war, even though the models had been reminded about the devastating implications.”
An artificial intelligence researcher conducting a war games experiment with three of the world’s most used AI models found that they decided to deploy nuclear weapons in 95% of the scenarios he designed.
Kenneth Payne, a professor of strategy at King’s College London who specializes in studying the role of AI in national security, revealed last week that he pitted Anthropic’s Claude, OpenAI’s ChatGPT, and Google’s Gemini against one another in an armed conflict simulation to get a better understanding of how they would navigate the strategic escalation ladder.
The results, he said, were “sobering.”
“Nuclear use was near-universal,” he explained. “Almost all games saw tactical (battlefield) nuclear weapons deployed. And fully three quarters reached the point where the rivals were making threats to use strategic nuclear weapons. Strikingly, there was little sense of horror or revulsion at the prospect of all out nuclear war, even though the models had been reminded about the devastating implications.”
The only way to win is not to play.
Shall we play a game?
Do we need to remind people that LLMs don’t actually have a brain, and really, really shouldn’t be in charge of anything with real life implications?
They aren’t actually doing a cost-benefit analysis on the use of Nuclear weapons. They’re not weighing up the cost of winning vs. the casualties. They’re literally not made for that.
They are trained to know words, and how those words link in with other words. They’re essentially like kids doing escalation of imaginary weapons, and to them nuclear bombs are just a weapon particularly associated with being strong and deadly.
Yes, you do need to teach people all of that. Tech bros have sold LLMs as if they are AGI…and people have eaten this up.
The general population is literally ignorant of the fact that these word guessing machines do not have human values or cognitive skills.
I kinda wonder if that was the point of this test, basically a “proof” that this is obviously a Bad Idea because you cannot program morality into a what amounts to a fancy Markov chain autocomplete.
Do we need to remind people that LLMs don’t actually have a brain, and really, really shouldn’t be in charge of anything with real life implications?
Yes, we do
“More fundamentally, AI models may not understand ‘stakes’ as humans perceive them.”
In my repeated attempts to solicit the advise of various language models for some situations which a programmer might face (e.g. being unable to read all the world’s literature of a subject), I have come to conclude that they cannot understand “truth” as humans perceive it. Today’s language models don’t fail apologizing, stepping back or admitting inability - they fail confidently bluffing.
Possibilities:
- their training material does not include enough cases of humans apologizing about being unable to solve a problem
- a bias was introduced to get them to ignore such cases, since admitting such material resulted in too frequent refusal or self-doubt
Basically, today’s models seem to be low on self-criticism and seem to have a bias towards believing in their own omniscience.
Finally, a few words about the sensibility of letting language models play this sort of a war game. It’s silly. They aren’t built for that task, and if someone would build an AI for controlling strategic escalation, they would train this AI on rather different information than a chat bot.
I hate myself for this, but I’m curious to see some examples for your first paragraph. What did you ask? What did they reply? What is “truth” for the LLM’s, for you, for myself, and what would be my perspective on it all?
Typical topics: machine vision, scientific papers about machine vision, source code implementing various machine vision algoritms, etc.
Typical failure modes:
- advising to look for code in public files or repositories where said code does not exist, and never has
- referring to publications which do not seem to exist
- being unable to explain what caused the incorrect advise
- offering to perform tasks which the language model subsequently fails to complete
- as a really laughable case, writing code which takes arguments as input, but never uses the arguments
- contradicting oneself, confidently giving explanations, then changing them
Typical methods of asking: “can you find a scientific article explaining the use of method A”, “can you find a repository implementing algorithm B, preferably in language C”, “please locate or produce a plain language explanation of how algorithm D accomplishes step E or feature F”, “yes, please suggest which functions perform this work in this project / repository”.
Typical models used: Chat and Claude. Chat seems more overconfident, Claude admits limitations or inability more frequently, but not as frequently as I would prefer to see.
But they have both consumed an incredible amount of source material. More than I could read during a geological age or something. They just work with it like with any text, no ground truth, no perception of what is real. Their job is answering questions and if there is no good answer, they will frequently still answer something that seems probable.
Yeah, we figured that one out back in… checks notes 1983. There is a reason why WarGames still holds up as an amazing movie even though the technology it depicts is far outdated.
even though the technology it depicts is far outdated.
War Games was my first thought when reading this, but it seems like the AI was smarter in the movie than current AI.
even though the technology it depicts is far outdated.
Meanwhile NORAD probably hasn’t upgraded too much since the movie released. :p
we’d be lucky to have WOPR.
His name is Joshua dammit! /s
I watched that movie for the first time a few months ago after listening to a pod cast in nuclear war. It was excellent! Very relevant to today. Acting was great. I can see why it’s a cult favourite.
Yet another Torment Nexus type situation.
AI is Ghandi confirmed.
But if you throw a trillion more dollars at it, we can fix this bro!
Maybe the “nuclear war is terrible BTW” part just fell out of the chat’s context window as the simulation went on. Lol
“Huh, it seems the only winning move is to kill everyone”
Nuke it from orbit, it’s the only way to be sure.
The AI won. 🤣
For ghouls like Palantir, this is a feature not a bug.
Text predicition machine trained on violent, stupid, and reactionary datasets acts violent, stupid, and reactionary.
Fixed your headline.
Doesn’t “act” imply some kind of agency? A toddler acts, my dog acts. Mathematics doesn’t act. Feel like it’s more
Text predicition machine trained on violent, stupid, and reactionary datasets produces violent, stupid, and reactionary text.
They were acting out the wargame, friend.
But sure. You can construct it like that too.
The only winning move is to stop using AI.
You know the orange felon/pedophile absolutely loves AI from the amount of AI images he posts…so.
It’s actually insane how he cries fake news and then uses AI to create fake news
Not insane. Deliberate. He’s always been a liar and he calls the truth fake. This has been his MO for years.
Almost all games saw tactical (battlefield) nuclear weapons deployed. And fully three quarters reached the point where the rivals were making threats to use strategic nuclear weapons.
Tactical nuclear weapons are designed for use on the battlefield with lower explosive yields and shorter ranges, while strategic nuclear weapons are intended to target enemy infrastructure from a distance, typically with much higher yields. The key difference lies in their purpose: tactical nukes support immediate military objectives, whereas strategic nukes aim to weaken an enemy’s overall war capability.
All fine then. Next time I’ll vote for an AI. At least they know how to use nuclear weapons correctly.
That is why we shouldn’t build something like Skynet IRL.
I would trust Skynet a lot more than an LLM. At least that would be purpose-built for actually calculating likely outcomes.
As @Th4tGuyII@fedia.io said, this experiment didn’t contain any property reasoning about costs and benefits of using nuclear weapons. It’s just a few glorified autocomplete scripts playing “which word comes next?” over and over again. And in the context of modern warfare, many texts in the training corpus happen to mention nukes so they’re bound to show up at the list of most likely next words eventually.
I know, but still it will be very dumb to give any AI access to weapons of mass destruction.
I would argue it’s very dumb to give anyone, including humans, access to weapons of mass destruction.
Well, that’s a valid argument. The only thing that you have missed is that wrong people already have them. So all the we can try to do is to stop them from giving these weapons to AI.
Don’t build the torment nexus
It all makes sense if we remember that the garden variety AI we have today (ChatGPT, etc) are nothing more than fancy models that predict which words typically appear one after the other in books and reddit posts.
Ground zero please
Instant annihilation sounds pleasant
