Good distillation.
Let's unpack the options for what the US government could do to potentially restrict Chinese AI models: 1. FUD: @deanwball noted the other day that the US government could take a "FUD" approach of just creating lots of doubt and uncertainty about use of Chinese AI, rather than strict legal requirements. (Dean made this point as an analytic matter, he was not advocating for it). The government could issue public security advisories, get sectoral regulators like energy and financial regulators to start raising risks, etc. This might not legally prohibit use, but would likely discourage it. 2. Procurement bans: The U.S. government could use procurement authorities to prohibit the use of Chinese AI models in government contracts, similar to what DoD did to Anthropic earlier this year. This would not prohibit private companies from using Chinese AI models unrelated to government contracts, but it would discourage use and reinforce a FUD approach. 3. Entity List: The Commerce Department could put specific Chinese AI companies (Moonshot, etc.) on the Entity List, which would prohibit exports of US technology to them. Setting aside evasion, this would cut off their access to any remaining US-origin chips not already covered by more general export controls and to US software. It would also potentially complicate U.S. companies’ use of closed or hosted models, as a US company could not send its own software or other "items" to the Chinese company for processing. It would not, however, strictly legally prohibit a U.S. company from downloading an open source model that a Chinese AI company posted online, as the download would not be an export *to* China, but rather an import from it. 4. "ICTS" rule: Commerce's existing "ICTS rule" allows Commerce to restrict the import or US use of a variety of software and other technology that poses a national security risk. Commerce has used the ICTS rule to, inter alia, prevent Americans from using Kaspersky Lab software (from Russia) and it is the basis for the recently-in-the-news restrictions on Polestar autos. Commerce could in theory deploy the rule to prohibit US firms from using Chinese AI models in the US. The one potential caveat to this is that the ICTS Rule exists pursuant to IEEPA, and IEEPA cannot regulate the import (or export) of "information" or "informational materials." I think courts would almost certainly uphold the ICTS rule as applied to the import or use of Chinese-hosted AI models as a prohibition on a “service,” and not on “information.” There would, however, be an interesting question about whether the courts would uphold a ban on a simple US download of an open-source Chinese model, where there was no payment. 5. Sanctions on Chinese AI labs: The President could use IEEPA to impose sanctions on Chinese AI labs, likely after declaring a new national emergency related to adversarial country AI risks. This would block their property in the US and prohibit all US commercial transactions with them. However, IEEPA sanctions would have the same carve out for "information" or "informational materials" as applies to the ICTS rule, and there is some chance the courts would find that IEEPA sanctions cannot prevent a US company from downloading, for free, an open source Chinese model. (Courts, for example, blocked Trump's 2020 use of IEEPA to try to restrict TikTok in the United States). 6. New IEEPA-derived restrictions on US companies using/deploying Chinese AI models: IEEPA is a flexible statute and the Trump Administration could use IEEPA to create new restrictions on the use or deployment of Chinese AI models, for example prohibiting US companies from using the models both in the US and abroad and prohibiting US hyperscalers from making Chinese AI models available in the US or globally. This could either be an outright prohibition, or a restriction, e.g., Chinese models could only be used subject to certain conditions, such as security protocols. (There is some conceptual parallel to how the Department of Justice's Data Security Program works). (There would, however, still be the legal question about whether IEEPA could in fact prevent a US company from simply downloading and using a free Chinese open source model). 7. Congressional legislation: Congress could of course legislate new restrictions on the US use of Chinese AI models. This would avoid the IEEPA limitations discussed above related to “information,” though would still be subject to the First Amendment. I could see the courts upholding a congressional ban on the use of even free open-source AI models despite the first Amendment by finding that such a ban passes “intermediate scrutiny.” (This was part of the basis on which SCOTUS upheld *Congress’s* 2024 legislation requiring a divestment of TikTok after lower courts enjoined Trump’s 2020 IEEPA-based ban). Note: this post is intended solely as a rundown of what the legal options for restrictions are, not a normative argument on what the US government should do.
@petereharrell @deanwball How about: 8. Just have friendly relations with China Ever thought of that?