From https://metaculus.com/questions/17102/us-export-ctrls-for-generative-ai-before-2026/
As of May 2023, US law does have export controls on the sale of software services that include powerful generative AI. As such, OpenAI, Anthropic, Google, and Microsoft can sell LLM-assisted document writing, image generation, video generation, etc. to enterprises worldwide (up to more general export controls not focused on AI.)
In Oct 2022, the US implemented export controls, that, roughly speaking, bans export of semiconductors that involve US in their manufacturing chain to China.
As the capabilities of generative AI systems grow, and as enterprise customers learn more how to extract certain types of knowledge (e.g. how to create Deep Fakes or create armies of online bots), one way US policymakers might address risks would be to apply exports control to software that are similar to the export controls for hardware used for training AI.
1,000
3.00How do you guys expect this to work? Deepmind is based in London and is therefore not, by default, bound by US laws - and according to Google (admittedly a biased source), they have an AI that performs better on benchmarks than GPT-4. If the US tries to claim extraterritorial jurisdiction, Google could just hive off Deepmind into a separate company.
We've already seen the lengths Sam Altman was willing to go to to subvert governance attempts. If AI is valuable enough, I don't see a Google demerger to evade export controls as beyond the bounds of possibility. So, such a law seems kind of pointless. And this is not even considering competition from other countries, such as France (Mistral AI) and China.