“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.”
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:
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:
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.