From Permission to Participation

Jollywood

Executive summary

  • The creative industries’ recent court filings point toward licensing AI rather than stopping it. The open question is how to make permission work at scale.
  • Licensing still runs on custom, platform-specific deals. The largest of the past year ended when its platform shut down.
  • The market has built strong tools for saying no: lawsuits, guardrails, legislation and detection. It has no standard way to say yes.
  • Training and generation are separate questions. The generation question needs an answer however the training cases end.
  • We propose six components for authorized AI creation: Permission, Authorization, Portability, Verification, Accountability and Governance.

From whether to how

For three years, the public argument about AI and entertainment has been framed as a fight. Creators against machines. Studios against model companies. Licensing against fair use. In the last few weeks, that framing started to break.

On October 30, 2026, William Morris Endeavor filed an amicus brief in The New York Times’ copyright case against OpenAI. A day earlier, SAG-AFTRA, the American Federation of Musicians and both Writers Guilds filed a joint brief of their own. By the deadline, the court had received 52 amicus briefs. Read together, the entertainment industry’s filings don’t ask for AI to stop. They ask for it to be licensed.

That doesn’t settle the debate. The law is still contested, and not everyone agrees on where the lines should fall. But it marks a shift in emphasis: away from whether AI-generated media should exist, and toward who can authorize the use of a character, a likeness or a brand inside it, on what terms, and how the people who own that IP get paid.

The market is moving from debating permission to working out how to make it operational, and that part is far from solved. Licensing a famous character for AI creation still works the way licensing has always worked: a negotiated deal between two large companies, months of legal work, and terms written for a single platform. That model produced the biggest AI character licensing deal of the past year. It collapsed when the platform did.

This paper looks at why permission doesn’t scale, what the industry’s recent filings and deals reveal, and what a rights layer for AI-generated media has to do to work. It ends with the hard questions this market still faces, and where we stand on each.

The permission problem

Traditional licensing was built for a small number of large uses. A toy company wants a character on a product line. A brand wants a celebrity for a campaign. Each deal is negotiated once, covers a defined category, territory and term, and goes through approvals on both sides. The process is slow, but the volume is low enough that slow works.

AI creation turns that volume upside down. In Adobe’s Creators’ Toolkit Report from October 2025, a survey of 16,000 creators, 86% said they use generative AI somewhere in their workflow. Single platforms now produce millions of videos a day. A meaningful share of that demand involves recognizable IP: famous characters, real people, known brands. That is what audiences respond to, and it’s what creators reach for.

The demand is being met. It just isn’t being met with permission. When no authorized path exists, people make the content anyway, using likenesses and characters nobody approved. Rights holders see it, and their only tool is to take it down after the fact.

The math explains why. A negotiated deal per creator, per use, is impossible at this scale. A blanket deal with one platform solves the paperwork, but it gives the rights holder no say over each individual output, and it only covers that one platform. Neither option gives a rights holder what they actually want, which is to say yes to the uses they’re comfortable with, no to the rest, and get paid for the yeses.

That gap between demand and permission is the problem this paper is about.

What the industry told the court

The New York Times v. OpenAI is a case about training: whether using copyrighted work to build an AI model requires permission. The outcome is not settled, and we won’t predict it. What’s useful here is what the entertainment industry chose to say while it had the court’s attention.

WME argued for licensing, not prohibition. The agency’s brief frames its goal as preserving a legal system in which creators decide whether, and on what terms, their work is used to develop or train AI. It also argues that a licensing market already exists, pointing to more than two dozen licensing deals between rights holders and AI companies announced between mid-2023 and mid-2024, and to AI developers actively shopping for licenses to film and television libraries.

The unions filed together. SAG-AFTRA, the American Federation of Musicians, and the Writers Guild of America East and West filed a joint brief backing the Times on October 29, a day before WME.

Talent organized around consent. A separate brief came from the Human Consent Foundation, a coalition that includes CAA.

Not everyone agrees. The Wikimedia Foundation filed too, with a more cautious position. It argued that too narrow a reading of fair use could hold back innovation, while blanket fair use could push the web toward closed systems. That tension is real, and any honest account of this debate should include it.

Taken together, the filings show where much of the creative industry now stands. For these filers, the argument is less about whether AI belongs in entertainment and more about consent, control and compensation: who agrees to the use, who sets the terms, and who gets paid. Those are commercial questions as much as legal ones, and they need commercial answers whichever way the court rules.

What the deals tell us

If the briefs show where the industry stands, the deals show how far it has gotten. The answer: further than it was a year ago, and not nearly far enough.

The biggest deal didn’t survive its platform. In December 2025, Disney and OpenAI announced a three-year licensing agreement and a planned $1 billion equity investment. Sora users would be able to create with more than 200 characters from Disney, Marvel, Pixar and Star Wars. Talent likenesses and voices were excluded. On March 24, 2026, OpenAI announced it was shutting Sora down. Disney said the deal would not move forward, and reports indicate the investment was never paid.

The lesson isn’t that the deal was badly made. It’s that the license lived inside one company’s product. When that product went away, so did every authorized use of those characters. Rights holders who want AI revenue can’t stake it on any single model company’s roadmap.

The next round is already forming. In August 2026, the Los Angeles Times reported that Google had approached Disney, Universal, Warner Bros. Discovery and other studios about licensing their IP for AI models. One person familiar with such deals estimated pricing at around $40 million per character on average, putting a 100-character package in the billions. No agreement has been announced, and NBCUniversal said it isn’t in those discussions.

Smaller, sharper models are emerging alongside. Warner Music and Suno reached an opt-in licensed AI music deal in November 2025. ElevenLabs’ Iconic Marketplace started with licensed voices and has since expanded into franchise characters, including Hasbro properties, with rights holders approving each request. In August 2026, Higgsfield premiered what it called the first fully AI-generated feature film starring licensed celebrity likenesses, with rights contracted before production.

The pattern across all of these is the same. Each deal is custom. Each is tied to one platform. And the economics only work for the largest catalogs: a reported $40 million per character is a price for Disney, not for an independent creator, a mid-sized brand or a talent agency’s client list. The market is proving that licensed AI creation is valuable. It hasn’t yet built a way for most rights holders to take part.

Enforcement is built. Authorization is not.

While licensing has moved deal by deal, the tools for saying no have matured quickly.

Litigation. Disney, Universal and Warner Bros. sued Midjourney in 2025 over AI-generated versions of their characters. Other studios have sent cease-and-desist letters to AI companies over unauthorized use of their IP.

Industry guardrails. In August 2026, the Motion Picture Association and ByteDance agreed a framework for IP guardrails on ByteDance’s generative video and image models. It’s a meaningful step, but it is a set of limits on what the model will produce, not a path to license anything.

Legislation. The NO FAKES Act would create a federal right protecting a person’s voice and visual likeness from unauthorized digital replicas. In June 2026, the Senate Judiciary Committee advanced it by a unanimous vote. It still needs a full Senate vote and passage in the House before it becomes law.

Detection. According to the same Los Angeles Times report, YouTube has approached agencies and studios about a likeness-detection tool that would flag AI content using copyrighted material or altered celebrity faces.

Each of these is necessary. None of them gives a creator a legitimate way to make the thing they wanted to make. The industry now knows how to say no, quickly and at scale. It still has no standard way to say yes.

That imbalance has a cost. When the only available response is a takedown, every unauthorized use is a loss for the rights holder, and every creator who wanted to do it properly has nowhere to go. Authorization turns the same demand into revenue.

Training and generation are different questions

Much of the public conversation blends two separate issues, and it’s worth pulling them apart.

Training is about inputs: what material a model learned from, and whether that required permission. This is what The New York Times v. OpenAI is about.

Generation is about outputs: whether a specific piece of content featuring a specific character, likeness or brand was made with the rights holder’s permission, under terms they set.

The two are related, but neither settles the other. A model trained only on licensed material can still be prompted to produce a famous character its maker never licensed for output. And a properly authorized output doesn’t answer how the underlying model was trained. The reported Google talks illustrate the overlap: the discussions reportedly cover both training on studio libraries and generating content with studio characters.

This matters for rights holders because the generation question doesn’t wait for the training question to be resolved. Every day, content featuring their IP is being made and published. Whatever the courts decide about training, someone still has to answer the output question: was this use allowed, by whom, on what terms, and who gets paid?

That’s the question a rights layer exists to answer. It’s a commercial problem that exists regardless of how the litigation ends.

The Jollywood Framework for Authorized AI Creation

If permission is going to work at the scale of AI creation, it has to be built as infrastructure rather than negotiated as a series of exceptions. Based on what the filings, the deals and the failures above tell us, we see six components. Together they make up what we call the Jollywood Framework for Authorized AI Creation.

  1. Permission. The rights holder defines authorized uses once: what’s allowed, what isn’t, in which contexts, and at what price. Those terms apply to every request instead of being renegotiated each time.
  2. Authorization. Every request is checked against those terms before anything is made. Where a rights holder wants a person to approve certain uses, those requests go to them.
  3. Portability. Licensing shouldn’t depend on a single model. The Sora collapse is the clearest argument: a license should belong to the rights holder and travel with their IP, not live inside one company’s product. Working across models doesn’t make a license portable on its own, though. Contracts, each provider’s policies and real technical integrations all have to line up for terms to carry from one model to the next.
  4. Verification. Every licensed output carries a verifiable record of the license that allowed it, so anyone can tell an authorized use from an unauthorized one.
  5. Accountability. Rights holders see how their licensed IP is used and are paid for that use. That’s what turns permission into a business rather than a favor.
  6. Governance. Union and guild agreements already include AI protections, talent has contractual rights over their likeness, and representatives negotiate the terms. The framework carries those obligations rather than routing around them.

One more point belongs here, because it’s often misread. Infrastructure like this doesn’t replace the people who represent rights holders today. Agencies and licensing agents will still negotiate the master terms: approvals, exclusivity, pricing frameworks, categories that are off limits. What changes is what happens after. Instead of every individual use requiring its own negotiation, the terms the representative agreed are applied automatically to every request. The agent’s judgment is preserved. Only the paperwork goes away.

How Jollywood approaches it

Jollywood is the commercial rights infrastructure for AI-generated media. Our proprietary, patent-pending rights infrastructure is built around the framework above.

Rights holders, whether they own a character, a likeness, a brand or a catalog, define how their IP can be used and what it costs. Agencies and licensing agents can set those terms on behalf of their clients. Creators, agencies and brands then make content with that IP inside the terms the rights holder set, with approval steps where the rights holder wants them.

The platform is model-agnostic by design. The license sits with the rights holder and their IP, not with any one AI model, so authorized creation doesn’t depend on a single company’s product decisions. Today, that creation happens through Jollywood’s own platform. Over time, the same permissions are designed to reach the AI tools creators already use. Built for direct creation today and platform integration tomorrow.

Jollywood’s proprietary verification ties every authorized output back to the license that allowed it, and rights holders see where their IP is used and earn from that use.

Hard questions, and where we stand

Some questions in this space don’t have settled industry answers yet. Here’s how we approach them.

Can outputs be licensed while training is still contested? Yes. The two are separate questions, and the generation question needs an answer today. Jollywood licenses the output, under terms the rights holder sets, whatever the courts decide about training.

What happens after a licensed piece leaves the platform? The license travels with it. Every output carries its Security ID and content credentials, and licensed content is monitored after it’s published, so use beyond the license can be identified and addressed.

Can standard terms carry guild and union requirements? They have to. Those obligations belong inside the terms a rights holder or their representative sets, not outside them, and approval steps keep a person in the loop wherever an agreement requires it.

Who gets access? Anyone with IP worth licensing, not only the largest catalogs. When terms are set once and applied to every use, an independent creator’s character or an individual’s likeness can earn the same way a studio’s does, without a nine-figure deal.

What can actually be verified? Every authorized Jollywood output carries a Security ID and signed content credentials that tie it to the license that allowed it. We think the whole industry should be held to standards that can be checked this way.

The last few years built the tools to tell the market what it can’t do. The next phase needs infrastructure that tells it what it can, and pays the people whose work makes it worth doing.

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