Legalize distillation: a provocative proposal to close the US-China AI gap

Legalize Distillation: Why Ben Thompson Says the US Should Stop Fighting It

Ben Thompson has a proposal that will make every frontier lab's legal team uncomfortable: pass a US law that explicitly legalizes distillation.

Thompson, the analyst behind Stratechery, wants Congress to do two things. First, make clear that collecting data for model training counts as fair use. Second, bar companies from writing terms-of-service clauses that forbid distillation 1.

That second provision is the provocation. Distillation, the practice of training a new model on the outputs of an existing one, is how competitors extract knowledge from frontier models without licensing the underlying data. Major labs include terms-of-service language prohibiting it. Thompson's argument is that enforcement is practically impossible.

"Stopping distillation, which is literally just querying the API, is nearly impossible," Thompson writes. His prescription goes the opposite direction: build a copyright framework that indemnifies labs for training on unlicensed data while guaranteeing that what they learn becomes fuel for further innovation 1.

This proposal arrives as Chinese labs are releasing increasingly powerful open-weight models. Alibaba just published Qwen 3.8 Max, a 2.4-trillion-parameter model released as open weights, nearly as large as the 2.8-trillion-parameter Kimi K3 1. That release was a reversal: Alibaba had decided not to release the previous version, Qwen 3.7 Max, in May. Thompson theorizes that Alibaba's change of heart may have been influenced by a speech from Xi Jinping, who told Chinese industry to "seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing" 1.

The strategic picture Thompson is pointing at: Chinese open-weight releases are accelerating, possibly at the urging of the country's top leadership. US labs are spending billions building frontier models while trying to lock down the knowledge in those models through ToS language that is, in practice, unenforceable. Thompson's thesis reframes the competition entirely. If you cannot stop distillation, treat open innovation as the competitive advantage rather than the vulnerability.

For builders, the implication is direct. If distillation becomes legal by default in the US, the moat for frontier labs shifts from data exclusivity to execution speed and infrastructure. Thompson makes this case through commodity market logic: as intelligence for many economically useful tasks approaches commodity status, the route to profitability is not charging higher prices but building a superior cost structure 2. A team that can distill from any available model and ship improvements in weeks operates on a different cost curve than one waiting for licensing negotiations to conclude.

For investors, the question is whether AI defensibility comes from legal walls or operational velocity. Under Thompson's framework, companies whose primary moat is data exclusivity lose their margin of protection. The beneficiaries are the ones with infrastructure advantage, distribution reach, and the engineering talent to convert open knowledge into shipped product faster than competitors.

There is also the hypocrisy Thompson is naming directly. US labs train on unlicensed data at massive scale while writing ToS clauses that forbid others from doing something structurally similar to their outputs. Thompson's proposal dissolves that tension not by tightening enforcement but by declaring the entire category fair game. Labs get indemnification for their own training practices; in exchange, they accept that their models are, by design, input for the next round of innovation 1.

Xi's open-source push turns this from a domestic copyright debate into a competitive strategy question. If Chinese leadership is directing labs toward open-source collaboration as national policy, the question is not whether openness is risky. It is whether the US can win on a playing field where openness is the default and speed is the differentiator. Thompson's answer is that leaning into that competition, rather than legislating against it, is the stronger position.

The proposal is unlikely to become law in the near term. But as a framing device for how the industry thinks about its own positioning, it matters now. The labs fighting hardest to wall off distillation are the ones most exposed if the walls come down regardless. If the moat was never really the data but what you built with it next, the companies optimizing for speed over secrecy are already ahead.