Tan's Distillation Gambit: Open-Weight AI's Public Good?

Y Combinator's Garry Tan wants U.S. open-weight labs to distill frontier models, arguing that because those models were trained on public human knowledge, access to capable AI should be 'a form of public good.' That single sentence puts him against OpenAI, Anthropic, and Google—and squarely inside the unresolved question of whether model outputs are fair game.

Y Combinator's Garry Tan said U.S. open-weight AI labs should be allowed to distill frontier models, framing capable AI access as 'a form of public good.'
The claim rests on the idea that frontier models were themselves trained on public human knowledge, so their outputs carry a public obligation.
The tension: no U.S. law currently requires frontier labs to permit distillation, and closed labs treat outputs as proprietary.
This article resolves whether Tan's framing is a policy opening or a rhetorical dead end—and who benefits either way.

What Exactly Did Garry Tan Propose?

TechCrunch reported on September 11, 2026 that Y Combinator president Garry Tan wants U.S. open-weight AI labs to 'distill' frontier models, and that Tan argues frontier models themselves were trained on public human knowledge—so access to capable AI should be 'a form of public good.' The proposal is not a bill or a filing; it is a norm-setting argument from one of Silicon Valley's most visible startup accelerators. That matters because Y Combinator's portfolio includes a long list of AI startups that depend on cheap, capable models. Tan is not a neutral commentator. He is an interested party whose argument, if adopted, lowers the cost base of the companies he funds.

Why Is Distillation the Flashpoint?

Distillation is the practice of training a smaller or cheaper model on the outputs of a larger one. It is how many open-weight models close the gap with frontier systems without spending frontier-level compute. Closed labs—OpenAI, Anthropic, Google DeepMind—have terms of service that restrict using their outputs to train competing models, though enforcement is patchy and largely contractual rather than statutory. According to TechCrunch, Tan's argument is that the frontier models' own training on public knowledge creates a reciprocal obligation. That is a moral claim, not a legal one. U.S. copyright law does not currently recognize a 'public good' exception for model outputs, and no court has ruled that distillation from a closed API is fair use.

Tans Distillation Gambit: Open-Weight AIs Public Good or Free Ride?

Who Wins and Who Loses If Tan Gets His Way?

Open-weight labs and their downstream users win immediately. A legal or normative green light on distillation compresses the capability gap between frontier and open models, which lowers inference prices and weakens the pricing power of closed labs. Y Combinator-backed startups win twice: cheaper models and a policy environment friendlier to building on top of them. Closed labs lose leverage. Their moat is partly capability, partly data, and partly the contractual fence around outputs. Tan's argument attacks the third leg directly. If 'public good' becomes a widely accepted norm, that fence becomes politically expensive to maintain.

Approach	Position on Distillation	Primary Interest	Likely 2027 Outcome
Y Combinator / Garry Tan	Permit it; frame as public good	Lower model costs for portfolio	Norm gains traction in policy debate
OpenAI	Restrict via terms of service	Protect capability moat and API revenue	Lobbies for contractual enforcement
Anthropic	Restrict; emphasize safety provenance	Control over model behavior lineage	Pushes safety-based distillation limits
Google DeepMind	Restrict; Gemini API terms	Protect Gemini ecosystem	Quietly tightens API terms
Open-weight labs (Mistral, EleutherAI)	Permit; expand access	Compress capability gap	Release distilled models openly
Verdict	Tan's framing wins the argument, not the law	Open-weight labs gain ground; closed labs keep contractual control	No statute by 2028; norms shift toward openness

Is 'Public Good' a Legal Argument or a Political One?

It is political, and Tan likely knows it. TechCrunch's framing—'wants U.S. open-weight AI labs to distill frontier models, too'—places the proposal inside a broader debate about whether American open-weight labs should keep pace with Chinese open models like DeepSeek, which have already been accused of distilling Western frontier outputs. The 'too' is doing real work: Tan is arguing that if the practice is happening anyway, U.S. labs should not be disadvantaged by unilateral restraint. That is a competitiveness argument dressed as a public-good argument. It may be more effective than either alone. Lawmakers respond to both fairness and national competition, and Tan is offering them a version of each.

What Would Actually Change the Rules?

Not a blog post or a conference panel. The binding levers are contracts, courts, and Congress. Contracts: frontier labs can tighten API terms and add output watermarking or rate-limit anomalies that signal distillation. Courts: a fair-use ruling on distillation would settle the question faster than legislation. Congress: any AI bill that defines 'public good' access would be the real win for Tan's side. According to Y Combinator's public positioning, the firm has consistently argued for lower barriers to AI capability. That consistency is a signal of intent, but it is not evidence that Tan's specific distillation argument has legal traction.

Thesis: Tan's distillation argument is morally coherent and legally empty—and that combination is exactly why it will shape the next two years of AI policy without ever becoming law.

In the short term, nothing changes. Frontier labs will keep their terms of service, and open-weight labs will keep distilling where they can and disputing the terms where they are caught. The immediate consequence is rhetorical: Tan has given U.S. open-weight labs a public-interest vocabulary to use in Washington, and that vocabulary is cheaper to deploy than litigation.

In the long term, the argument matters because it reframes who owes whom. Closed labs took public text to build private capability. Tan is saying the public is entitled to a return—not in cash, but in access. That framing is hard to rebut on fairness grounds and easy to rebut on property grounds, which is why it will live in hearings rather than courtrooms. My concrete prediction: by Q2 2028, at least one major U.S. frontier lab will publicly revise its API terms to create a narrow, documented distillation exception for academic and open-weight researchers—not because of Tan's argument alone, but because the political cost of refusing has risen. The losers are closed labs' pricing power; the winners are open-weight labs and the startups built on them.

Predictions

By Q2 2028, OpenAI will publish an updated usage policy creating a limited distillation carve-out for academic and open-weight researchers, under pressure from the public-good framing.
No U.S. federal statute defining frontier model outputs as a public good will pass before the 2028 election; the debate will remain in agency guidance and contract law.
At least one open-weight lab—most likely Mistral—will release a model by mid-2027 that it explicitly describes as distilled from a U.S. frontier system, forcing a public legal test.

September 2026
Tan's distillation argument goes public
TechCrunch reports Garry Tan wants U.S. open-weight labs to distill frontier models, framing capable AI as a public good.
Q4 2026
Frontier labs review API terms
Closed labs are expected to tighten output-use restrictions and detection in response to distillation pressure.
2027
Expected open-weight distillation release
An open-weight lab is likely to release a model explicitly described as distilled from a U.S. frontier system, forcing a legal test.
Q2 2028
Predicted limited carve-out
At least one major U.S. frontier lab is predicted to create a narrow distillation exception for academic and open-weight researchers.

Estimated open-weight vs. frontier capability gap, 2024–2027 (estimated)

Article Summary

Tan's proposal is a norm-setting move, not a legal one; its power is in reframing public obligation, not in enforceable rights.
Distillation is the pressure point because it directly attacks closed labs' contractual moat around outputs.
The 'too' in TechCrunch's headline signals a competitiveness argument: U.S. open-weight labs should not restrain themselves while Chinese labs do not.
Real change will come through contracts and courts, not conferences; watch API terms and fair-use litigation, not op-eds.
Open-weight labs and YC-backed startups are the structural winners even if Tan's specific argument never becomes law.