AI Now Institute: claims about the risk of AI taking control of nuclear weapons must be verifiable
Heidy Khlaaf from AI Now Institute argues that existential risks from AI must be falsifiable; otherwise, they resemble religious arguments. Using nuclear weapons as an example, she shows that air-gapped systems and physical security (see Stuxnet) limit the realistic possibilities for AI to take control of them.
Heidy Khlaaf, chief AI scientist at AI Now Institute and a former safety engineer at OpenAI, argues that scientific claims about existential risk from AI must be falsifiable – meaning they must be capable of being proven or disproven. In her view, without this condition, such claims resemble religious arguments in their nature, which cannot be verified.
As a concrete example, she raises the question of whether AI could take control of nuclear weapons, which have existed for over 80 years and are generally considered capable of wiping out human life. Khlaaf points out that nuclear facilities are “air-gapped”, meaning disconnected from the publicly accessible internet, which she says limits the ways an AI agent could gain control of them. According to her, these systems are also built to significantly stricter and more regulated engineering standards, often including physical reinforcement of protective measures.
As historical evidence, she cites the Stuxnet worm, which damaged Iranian nuclear facilities – it had to be physically introduced into the system via a USB drive, not through an internet connection. The source provides no further details of the discussion or the broader context in which these statements were made. You can find details in the source article.
Why it matters
The argument shifts the debate about existential risk from AI from general concerns to concrete, technically verifiable claims – in the case of nuclear weapons, it shows that physical isolation and strict engineering standards present a real technical barrier that must be taken into account when assessing such scenarios.
Relevant practical impact
What this means
For individuals
Readers can better distinguish substantiated arguments about AI risk from unfounded concerns by requiring a concrete and verifiable description of the mechanism by which the risk in question would arise.
Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.