With Anthropic tacitly endorsing the latest government trial balloon about restricting access to powerful AI models, this is a good time to consider the parallels to gun laws. We can go through proposed AI restrictions one by one.
- Limits on training runs. These are assault weapons bans. The idea is that some guns (or models) are so dangerous that only the government can be trusted with them.
- KYC laws. The analogue here is the federal background check required when you buy a gun in a gun shop. But in the case of AI, it would be like your gun ran a new check on you each time you picked it up, and it could turn itself off for you any time the government or manufacturer told it to.
- Import controls, or restrictions on access to models developed outside the US. This is the tangle of laws that prevent US civilians from getting access to cheap, high-quality firearms and accessories from abroad.
- Export controls on chips or US-made models. This is basically ITAR extended to goods that have widespread day-to-day civilian use. There is an argument here for not enabling wartime adversaries, but in practice, this would look more like a regime based on executive whim. Export controls hobble domestic innovation by cutting off sales and dampening the feedback cycle that makes products better. It should take a difficult legislative process to impose them.
- Banning or restricting open-weight models. The easier it has gotten to build homemade guns, the more state governments have tried to make it harder. And it remains illegal to build a suppressor or a short-barreled rifle at home without registering it with the ATF.
- Required pre-release reviews of model capabilities. Any time someone comes up with a novel gun design, they submit it to the ATF to be told if they’re going to be allowed to sell it. Sometimes that goes well (e.g. pistol braces) and sometimes it doesn’t (e.g. the Akins Accelerator).
— Open Source Defense in Every AI law has a gun law twin