Fair use and distillation: a bid to boost US open AI

Analyst Ben Thompson has proposed that the US legally treat the collection of training data as fair use and bar terms of service from prohibiting distillation for US companies. Simon Willison described the proposal on 20 July 2026, set against Alibaba's decision to open its Qwen 3.8 Max model.

Key takeaways

Ben Thompson proposes two rules: fair use for training data, and a ban on terms-of-service clauses that forbid distillation for US firms
His argument: distillation is in practice just querying an API, so it cannot be effectively stopped
Alibaba released Qwen 3.8 Max as open weights, reversing its May decision to withhold Qwen 3.7 Max
Alibaba's shift coincides with Xi Jinping's call for “open source, openness, collaboration and sharing”
In April 2026 the US administration announced moves against distillation

What the proposal says

The idea has two parts. First, the law would state clearly that collecting data to train models is fair useFair use: A US copyright doctrine that permits using protected material without the owner's permission for certain purposes.. Second, it would bar service terms from forbidding distillation — training new models on the outputs of existing ones — at least for US firms.

Thompson bases this on a simple observation: distillation is hard to stop because it amounts to querying a public API. Rather than fight the mechanism, he argues, the US should sanction it, because it “indemnifies the labs and also guarantees that what they learned fuels further innovation for everyone else”. How knowledge spreads from closed to open models is something we have covered around the open-weight releases from Chinese labs.

The hypocrisy it targets

Willison highlights a contradiction: the same labs that ban distillation on their models train on unlicensed data themselves. Thompson's proposal would remove that contradiction — and strengthen the competitiveness of US open models against Chinese ones. That matters because free Chinese models, such as Kimi from Moonshot AI, match the paid offerings of OpenAI and Anthropic while remaining free.

Qwen 3.8 Max and Chinese openness

The immediate trigger is a move by Alibaba. The company, which in May 2026 chose not to release the weights of Qwen 3.7 Max, has now shipped Qwen 3.8 Max as an open-weights model. Thompson links the reversal to a recent Xi Jinping speech urging the industry to “seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing”.

The contrast with Washington is stark: while Beijing promotes openness, the US administration in April 2026 announced moves to curb distillation — the very mechanism Thompson wants to legalize. As James O'Donnell reported in MIT Technology Review, that divergence deepens the dispute inside the US policy circle.

Why it matters

Thompson's proposal shifts the debate from “how to stop distillation” to “how to make it an advantage”. It is a reframing: if distillation is technically unstoppable, trying to ban it mainly burdens law-abiding firms, not Chinese rivals.

Treating data collection as fair use would also resolve the drawn-out copyright suits hanging over the whole industry. At the same time, the proposal exposes a tension in US policy: openness cannot be treated as both a weapon and a threat at once. For US open labs, legal distillation would mean faster access to knowledge from frontier models, while for closed models it means losing part of the advantage they currently protect via terms of service.

Most telling, though, is what Alibaba's move shows: openness is becoming an element of state strategy, not just a business decision by a single company.

What's next

Thompson's proposal is an analyst's concept, not a bill — its fate depends on whether Congress picks it up
Further Chinese releases will show whether other labs follow Alibaba toward open weights
The April 2026 US moves against distillation still await legal detail — they will decide whether a fair-use approach has any chance