Training vs. Generation: Two Different AI Rights Questions
Jollywood
Two questions, one headline
When a story runs about AI and copyright, it usually lands under one headline: AI companies versus creators. But underneath that headline sit two different questions, and they get blurred together constantly.
One is about what goes into an AI model. The other is about what comes out of it. They involve different rights, different parties and different timelines. For anyone who owns IP or creates with it, knowing which question is being argued makes the news a lot easier to read.
The two questions, side by side
Training is about inputs. To build a model, a company feeds it enormous amounts of material: text, images, video, audio. The training question asks whether using copyrighted work for that purpose requires the owner’s permission and payment.
Generation is about outputs. Once a model exists, people use it to make things. The generation question asks whether a specific piece of content that features a specific character, person or brand was made with the rights holder’s permission.
| Training | Generation | |
|---|---|---|
| What it concerns | What a model learned from | What a model is used to make |
| When it happens | Once, while the model is built | Every time someone creates something |
| Rights mainly involved | Copyright in the training material | Copyright and trademark in characters and brands; a person’s right over their likeness and voice |
| Who is on each side | Model companies and owners of large bodies of work | Creators and the owners of the IP they want to use |
| Where it’s being argued now | Court cases such as The New York Times v. OpenAI | Licensing deals, takedowns, and bills such as the NO FAKES Act |
The likeness row deserves a note. A famous character is property, typically protected by copyright and trademark. A person’s face and voice are protected differently, through rights of publicity that today vary by state. The NO FAKES Act, which cleared the Senate Judiciary Committee in June 2026, would create a federal right over voice and likeness. That’s a generation issue at its core: it’s about replicas being produced, not about what a model was trained on.
This is a simplification. Whether an output needs permission at all can depend on copyright, trademark, publicity rights, contracts, and legal exceptions such as parody or commentary, and the answer isn’t always yes. The table shows where each question usually sits, not how any particular case would be decided.
Where they meet
The two questions are separate, but real deals often touch both. Three recent examples show how.
A generation deal that didn’t survive its platform. The Disney and OpenAI agreement announced in December 2025 was mainly about generation: letting Sora users create with more than 200 Disney, Marvel, Pixar and Star Wars characters. When OpenAI shut Sora down in March 2026, the arrangement ended with it.
Talks that reportedly cover both. In August 2026, the Los Angeles Times reported that Google had approached Disney, Universal, Warner Bros. Discovery and other studios about licensing IP for its AI models. The discussions reportedly include both training on studio material and generating content with studio characters. No agreement has been announced.
Training on your own IP. In November 2026, Disney and Adobe announced a partnership in which Disney will build private AI models trained on its own characters, for use in its own businesses. Here the training and the generation stay inside one company that owns the IP on both ends.
The overlap runs in both directions, and neither answer settles the other. A model trained only on licensed material can still be prompted to produce a character it was never licensed to output. And an authorized output doesn’t say anything about how the model behind it was built.
Why the difference matters
For rights holders, the output question doesn’t wait. Training cases can take years to resolve, and they mostly involve companies with very large catalogs. Meanwhile, content featuring famous characters, real people and known brands is being made and published every day. Whatever the courts decide about training, someone still has to answer, for each of those pieces: was this allowed, by whom, on what terms, and who gets paid?
For creators, it’s the question that affects them directly. Most creators will never train a model. They use models other people built. What they need to know is simpler: can I make this, using this character or this person, with the rights holder’s permission? That’s a generation question, and right now it rarely has a clear answer.
For everyone, mixing the two muddies the debate. When the headline says creators are fighting AI, it’s easy to assume the only options are to stop AI or let it take whatever it wants. Separating training from generation opens a third option: let creation happen, with permission, on terms the rights holder sets.
What isn’t settled
We’re not lawyers, and this isn’t legal advice. Several parts of this picture are genuinely open: how courts will treat training on copyrighted work, whether the NO FAKES Act becomes law and in what form, and how rights over likeness will apply to AI-generated replicas across different states and countries. Anyone making decisions on these questions should work with counsel.
What we’re confident about is narrower. The generation question exists today, at scale, and it needs a practical answer whichever way the training cases go. That answer is permission: rights holders setting terms, creators working inside them, and every authorized piece traceable back to its license.
That’s the side Jollywood’s proprietary, patent-pending rights infrastructure works on.
Jollywood is the commercial rights infrastructure for AI-generated media. It’s in private beta, with applications open for rights holders, creators and partners at jollywood.ai.
Sources
- The New York Times v. OpenAI and amicus filings: Publishing Perspectives
- Disney and OpenAI agreement: Disney press release
- Sora shutdown: Variety
- Google studio talks: Los Angeles Times, syndicated
- Disney and Adobe: The Next Web
- NO FAKES Act: Congress.gov
