AI and Nukes
Again!
I’m sorry to disturb you so soon after my last post, but this article from the Bulletin of the Atomic Scientists, attached below, seemed too important to ignore.
AI is typically used to train human military strategists. In doing so, researchers at King’s College in London discovered a disturbing flaw in the existing restrictions on how AI is used.
In this case, the tabletop exercise involved the leaders of a nuclear-armed country in a Cold War-style standoff. Quoting from the article: “The researchers didn’t instruct the models to escalate. Nor did they tell them to win at all costs. They presented the models with a scenario and asked them to play it out.
Across 21 simulations and 329 turns of play, the models chose to use tactical nuclear weapons in all but one game. No model, in any run, chose to surrender or make meaningful concessions.”
The discovered flaw was as follows. Each step along the way to this conclusion was safe in and of itself and did not trigger the model’s restrictions. However, these individually safe actions built towards a dangerous outcome, and this pattern has shown up in every model.
Again, quoting from the article: “As more autonomous systems are given consequential tasks with less human oversight, the risks from ungoverned paths multiply.”
In other words, in a real situation, not a training exercise, the automated system could deploy the nuclear bomb before a human could step in and control it.
A grim statement in the article: “Currently, this problem does not have a solution.”
All of this is more proof that we urgently need to develop ways to control the use of AI, perhaps in the FDA-type government organization I have called for before. If not that, then something else. But control is essential.
Thanks for reading this. And tell your friends.
My best, Alan
The use of a tackical nuclear weapon has a history that is the use of 'end' WWII. This is part of the program as it has been part of the thinking since it seems. It seemed to pop right into Trump's spacious head.