Distillation, Fair Use, and the Open-Weight Model Land Grab

What Ben Thompson actually proposed

Writing on Stratechery, Ben Thompson argues that US frontier labs' anxiety about Chinese open-weight models is largely a symptom of their own contradiction: they train on unlicensed web data while their terms of service forbid anyone else from querying their APIs to train competing models — the practice known as distillation. His proposed fix, summarised by Simon Willison, is two-part legislation: make data collection for training explicitly fair use, and void terms-of-service clauses that ban distillation, at least for US companies.

Stopping distillation — which is literally just querying the API — is nearly impossible; the U.S. should go the other way and try to make it easy for American companies to compete using distilled, open models.

— Ben Thompson, via Simon Willison

The logic is that if distillation can't realistically be stopped — a rival lab can always route millions of queries through an API and use the outputs as training data — the US is better off legalising it domestically so that American open-weight developers can build on top of frontier models like Chinese labs already do, rather than ceding that layer entirely.

The releases that prompted the debate

The timing isn't incidental. On 19 July 2026, Alibaba previewed Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts multimodal model, at the World Artificial Intelligence Conference in Shanghai — its first model above a trillion parameters to go multimodal, per MarkTechPost. That came just days after Moonshot AI's open-weight release of Kimi K3. Both point to the same trend: Chinese labs are shipping large, capable, openly-weighted models at a pace US labs — largely committed to closed, API-gated frontier models — are not matching.

Why AI governance teams should care, not just policy watchers

This story reads as a copyright and industrial-policy debate, and largely is one — we take no view on whether Thompson's proposal is good law. But it puts a spotlight on something PyramidLedger sees repeatedly in AI governance work: organisations adopt open-weight models — Chinese or otherwise — faster than they can answer basic provenance questions about them. Was this model itself distilled from another vendor's outputs, and does that create downstream licence exposure for a business using it commercially? What data was it trained on, and does that carry its own IP or regulatory risk under frameworks like ISO 42001? Those questions don't resolve themselves just because a model runs well on a benchmark.

Practical due diligence for teams adopting open-weight models

• Trace training and distillation lineage as far as the vendor or model card discloses it, and flag models with no disclosure at all.
• Review the model's licence terms for restrictions on commercial use, redistribution, or use of outputs to train further models.
• Treat an open-weight model as a supply-chain component in your AI inventory — versioned, reviewed, and re-assessed on update, not adopted once and forgotten.
• Separate the legal/licensing risk question from the technical security question — a permissively licensed model can still carry security or data-handling risks worth testing independently.

None of this requires taking a side on Thompson's proposed law. It requires treating model provenance as seriously as organisations already (should) treat open-source software provenance — which is exactly the muscle ISO 42001-style AI governance programmes are meant to build.