The rule
Commons in, commons out
Any company that trains AI models on material it has not licensed must release the weights of those models to the public, once a year, on a fixed schedule.
The bargain
A trade, not a ban
The most capable AI systems were built on the commons: public research, public-domain works, the open web, and work taken under contested legal claims. The companies that built them now hold that value behind closed doors. We do not propose to stop them. We propose a trade.
Use the commons, and you owe the commons. You keep a lead of about a year, which is enough to earn a return on what you built. Then what you built returns to the public that made it possible.
This is not a ban, a tax, or a licensing regime run by a regulator. It is a clear rule that firms can plan around. Train on licensed data and keep your model. Train on the commons and give it back.
The rule
Four parts
§ 1 · Coverage
Who is covered
Any firm that uses a general-purpose AI model (one that can generate text, images, audio, video, or code across a wide range of tasks) trained, in whole or in part, on unlicensed material. Material counts as licensed only if a written license covers it, from a party with a registered copyright or a documented chain of title. Everything else is unlicensed: public-domain works, openly licensed works, scraped web content, and anything used under a fair-use claim. The developer bears the burden of proof. Material without a provenance record is presumed unlicensed. A firm means the whole corporate group, so models cannot be parked in a subsidiary.
§ 2 · Trigger
What triggers it
Commercial use, direct or indirect. Any use by or for a for-profit company counts, including purely internal use. A hedge fund that trains a model on scraped data and uses it only to pick stocks is covered.
§ 3 · Release
What is owed
Every March 31, each covered firm releases the weights of every covered model it had in commercial use at any time during the twelve months ending the previous September 30, along with documentation of the safeguards in place at release. Every version, fine-tune, and checkpoint put into use counts, so there is no need to argue over what counts as one model. A firm's lead runs from six to eighteen months, about a year on average. Retiring a model early does not take it off the list.
§ 4 · Enforcement
How it is enforced
Covered models must be registered when commercial use begins, because internal use is invisible from outside. Failing to register is a violation in itself. The rule is a condition of market access. It applies to any model offered to users in the jurisdiction through any interface, and to any model whose outputs are used in commercial activity there, wherever it was trained. Distilling a covered model produces another covered model.
The definitions are deliberately strict. Almost every general-purpose model built today would be covered. That is the point. The way out is open to anyone: license the data you train on.
Why
Three principles
| Principle | The concern | What the rule does |
|---|---|---|
| Safety | AI belongs in the same threat class as nuclear weapons and dangerous pathogens, but it is fungible: it copies, leaks, and gets independently rebuilt. | For a technology that cannot be contained, the safest environment is one where defenders have what attackers have. |
| Efficiency and competition | A small number of firms dominate the market for AI, and markets close behind early movers. | Each year, last year's frontier becomes the floor that every competitor, researcher, and new entrant can build on. It forbids nothing and approves nothing in advance. |
| Fairness | The leading models were trained on work whose makers were not asked and were not paid. | No permanent private claim on a public inheritance. The public as a whole is the most practical beneficiary. |
Design
Built to be practical
This rule should be judged by three observable measures: what it costs the government to run, what it costs firms to comply, and how much litigation it produces. Each design choice is meant to keep all three low.
- One date a year. Every covered firm releases on March 31. There are no per-model clocks to track or dispute.
- Bright lines instead of judgment calls. Material is licensed only under a written license from a registered or documented rights holder. No fair-use analysis, no case-by-case weighing.
- No arguments over what counts as a model. Every set of weights put into commercial use is released, so firms gain nothing by renaming or slicing.
- Records only for those claiming an exemption. A firm that accepts coverage needs no provenance records. A firm claiming its data was licensed must prove it.
- Hard cases settled in advance. Worked examples in regulation and binding advisory opinions answer questions before they become lawsuits.
- Registration as simple as registering a domain. An online form, a few minutes, no cost or a nominal fee. The firm submits a cryptographic fingerprint of its weights, not the weights themselves.
- Cheap to enforce. A missing registration is provable on its face. Released weights are checked against the fingerprint filed at registration. The government never holds unreleased weights.
- A clear safe harbor. A firm that registers and releases on schedule is in compliance. Enforcement belongs to the government, with one narrow exception: anyone may sue to compel a release that is overdue, with no damages.
Safety
Why openness is safer
Enclosure does not keep a fungible technology out of dangerous hands. It keeps it out of the hands of defenders, researchers, and the public.
Cybersecurity learned this decades ago. Cryptography became trustworthy only when its algorithms were published and attacked in the open. Open-source software runs most of the internet because many eyes find flaws that a single owner misses.
Firms stay responsible for safety. Every release must come with documented safeguards. What the rule refuses to accept is secrecy offered as a substitute for safety.
Objections
Objections answered
- It will kill investment in AI.
- Firms keep an average of twelve months of exclusive use, and the cost of licensing data stays available as an alternative. Investment follows predictable rules, and this rule is predictable.
- Our model is a trade secret.
- Then license the data. A trade secret built on unlicensed public work is the exact thing the rule ends.
- Creators deserve payment, not the public.
- Agreed in principle. In practice there is no way to trace and pay every creator. The public is the closest practical stand-in, and a compensation scheme can be added without changing this rule.
- Firms will move offshore.
- Compliance is a condition of market access. A model offered here through any interface, or used for commercial activity here, is covered wherever it was trained.
- Firms will hide models internally.
- Internal use is commercial use, and registration is required from day one. Failing to register is a violation in itself.
- Open weights put dangerous capability in bad hands.
- Closed weights leak and get reproduced. Openness puts the same capability in defenders' hands at the same time.