Every AI company is asked who owns the model, and almost nobody answers in a way that survives contact with a licence agreement. In this episode we tried to work out what ownership actually means when the thing being owned is a set of weights, a training recipe and a pile of scraped text.

The three things people confuse

The weights. The file that does the work. Owning it is a copyright question, and the answer varies by jurisdiction and by whether the model was trained from scratch.

The training data. Usually not owned by the company that trained the model, which is the uncomfortable part. Scraped text is licensed, argued over or simply used.

The output. Who owns what the model produces is decided by the terms of service, not by copyright law, and those terms differ wildly between providers.

The open-weight illusion

Releasing weights is not the same as open source. A model can publish its weights and still restrict commercial use, require attribution, or cap the number of users. The open source definition was written for software, where the source code is the thing you need to rebuild it. For a model you would also need the data and the compute, and almost nobody publishes both.

That distinction matters commercially. A company that builds on open weights believes it owns its product. Often it owns a thin layer on top of something it cannot legally rebuild.

What the licences actually say

We read four of the most commonly used model licences and compared them on three points: commercial use, redistribution and the trigger clauses that change your rights at scale.

  • Most permit commercial use.
  • Most require you to pass on the licence terms.
  • Several include a clause that restricts use if your product exceeds a user threshold, which is a commercial term disguised as a licence term.

That last one is the part businesses miss. You can build a company on a model and discover, at a certain size, that the terms changed underneath you.

Why this matters beyond lawyers

If you are choosing a model for a product, the licence is a supply chain decision. A model you cannot rebuild is a dependency, and dependencies have prices. Our piece on how open-weight models closed the gap covers the capability side of the same question.

The ownership question nobody asks

Who owns the mistakes? If a model produces something defamatory or infringing, the provider's terms almost always place responsibility on the deployer. Ownership of the upside is negotiated. Ownership of the downside is assumed.

That asymmetry is the most important thing in the entire debate and it rarely appears in the marketing.

What we concluded

Three things, briefly.

  1. Weights are not source. Treat open weights as a distribution channel, not as freedom.
  2. The licence is a business risk, not a legal footnote. Read the threshold clauses.
  3. Ownership of output is contractual, so it can be taken away by an update to terms you agreed to years ago.

FAQ

Can you own an AI model? You can own the weights as a copyrighted work in some jurisdictions. You cannot own the concept or the maths.

Is open source AI the same as open source software? No, because the data and compute are not usually released.

Who owns what the model writes? Whoever the terms of service say, which is usually you, with conditions.

The bottom line

The ownership question has a boring answer: whoever holds the licence holds the power, and the licence can change. Anyone building on a model should read the threshold clauses before they scale, because that is where the terms stop being friendly.

Further reading: Hugging Face model licences and the Open Source Initiative's position on open source AI.