White House tests the waters for banning frontier open-source models

Paving the way for regulation through the Chinese threat

Discussions are currently underway in Washington about how to regulate open AI models. According to industry sources, the White House is preparing to use arguments around so-called model distillation and the threat from China (such as DeepSeek) to gain leverage over the open-source community. The pretext is intellectual property protection and cybersecurity, but the real goal is to introduce a ceiling for what can be released as an open weights model.

This effort is heavily fueled by Anthropic. The company has launched a campaign claiming that Chinese actors are distilling their models, and is outright proposing sweeping bans. Anthropic argues its models are too dangerous to be freely copied, even though this simultaneously demonstrates that the company cannot secure its own APIs against misuse.

Who the rules of the game are changing for

If the US administration succeeds in pushing through restrictions on the import or training of open models above a certain capability threshold (roughly at the level of GPT-4 or Claude 3 Opus), it will hit the emerging ecosystem of smaller companies hard. The startup scene that relies on fine-tuning and running open-source models would lose access to cutting-edge tools.

Closing the door on open-source development also cements the monopoly of established players. Companies selling closed models would gain the certainty that no one will destroy their margins by releasing a comparably smart model for free. It is a textbook example of attempting to legally anchor one's own market dominance.

The limits of protection through bans

The argument that banning open-source models in the US will increase security clashes with the reality of the global market. Development in China or Europe will not stop, and the bad actors the US government fears will just download the models elsewhere. The only result of a US embargo on open AI would be crippling its own research and market, while the rest of the world moves on.

Furthermore, it is hypocritical to claim that API access to models is safe. History shows that APIs can be bypassed or abused (as happened with Claude), so forcing all users onto closed platforms does not eliminate security risks; it merely centralizes them into the hands of a few corporations.

The next step must be a Western open model

Whether this regulatory offensive succeeds will depend on whether the US tech sector offers its own frontier open model before politicians intervene. Meta or Microsoft needs to release a model at the level of today's cutting edge to break the narrative that only China makes dangerous open models.

If such a model does not appear soon, the panic around distillation will become a precedent for permanently restricting open-source AI in the US.

Lilith's verdict

The fight over intellectual property is a smokescreen. Established players are lobbying the government to eliminate the very competition that would drive their prices to zero.