AI generated fan content sits on two different legal risks, not one. Learn which is which, and how licensed fan art avoids both. Start here.
Photo via Unsplash
Search "AI generated fan content" on a Tuesday and you will find two completely different arguments happening in the same results page. One is about whether a fan can legally put a real streamer's face into a piece of art. The other is about whether Midjourney can legally let anyone generate Bart Simpson. Most people asking the question think they are asking one thing. They are actually asking two, and the two have almost nothing in common except that a diffusion model can produce both in the same afternoon.
That confusion is not an accident of language. It is the actual shape of the legal exposure, and untangling it is the difference between fan content that survives a takedown request and fan content that does not.
Key takeaways
- "AI generated fan content" bundles two separate legal problems: a real person's right to their own likeness, and a rights holder's copyright in a fictional character.
- Disney and Universal's live lawsuit against Midjourney is the copyright half of that problem, currently in expert discovery. It says nothing about the likeness half.
- A watermark proves an image is AI-made. It does not prove anyone agreed to it — those are different documents, and only one of them survives a dispute.
- Licensed fan content needs consent from the person depicted and clearance for any third-party character design it borrows — one licence rarely covers both.
- A provenance manifest that names the boundary-set version in force at generation time is what makes a licence checkable months later, not just at the moment it was granted.
Fandom has always made two kinds of thing about the people and characters it loves: portraits of real people it admires, and reinterpretations of fictional characters it's attached to. A fan painting of a musician and a fan painting of Elsa are both "fan art" in ordinary speech, but they have never rested on the same legal footing, and generative tools didn't change that — they just made both so cheap to produce that the difference stopped mattering to the person clicking generate.
It still matters to everyone else involved. A real person's face is theirs to license or refuse. A studio's character design is the studio's to license or refuse. Neither party's rights are transferred by the other party's consent, and a platform that treats "the subject said yes" as covering a copyrighted costume, or "the character is public domain-adjacent" as covering a real face, is wrong on one half of every image it produces.
When the output is a real person — a creator, an athlete, a public figure — the operative right is theirs alone: the right to control commercial and, in a growing number of jurisdictions, non-commercial use of their own identity. In the US this varies state by state, but the direction of travel is toward stronger protection, not weaker. Federally, the TAKE IT DOWN Act (S.146) now creates a 48-hour notice-and-removal duty enforced by the FTC for non-consensual intimate imagery specifically, and we've covered the mechanics of meeting that deadline in the TAKE IT DOWN Act compliance guide. In the UK, the Online Safety Act 2023 puts parallel duties on services that host this kind of content, which we walk through in the UK Online Safety Act guide.
None of that legislation asks whether the image was labelled "fan art". A takedown duty triggered by a real person's likeness doesn't care what genre the request thought it was making.
What consent actually looks like, done properly, is a boundary set: a published, structured statement of what a specific creator will and won't allow, evaluated on every request before generation happens. We cover writing one in the boundary configuration guide, and the full vocabulary of permitted contexts and forbidden depictions lives on the boundary type reference. Fan-facing licensing on LikeWard runs through exactly this mechanism, documented on the licensed fan art use case page — a request outside a creator's published limits is refused with the specific rule it broke, not generated and argued about afterwards.
The other half of "fan content" is fictional. Elsa, Bart Simpson, Darth Vader, the Minions — none of them are real people with a right of publicity, but all of them are somebody's copyrighted and trademarked property, and that somebody has started litigating.
Disney and Universal filed suit against Midjourney in June 2025, alleging that the platform's image generator reproduces their characters on request without a licence. The case has since been consolidated with a separate Warner Bros. action, with Disney's suit as the lead matter — and as of August 2026 it's very much live: expert disclosures are scheduled through October 2026, with expert discovery running into November. It is, functionally, the first serious test of whether a generative platform can be held liable for character-level output rather than just training-data ingestion, and the outcome will set a real precedent either way (Hollywood Reporter; Georgetown Law's Tech Institute).
A second case worth watching for anyone thinking about the training side rather than the output side is Andersen v. Stability AI, brought by cartoonist Sarah Andersen over her work being used to train an image model — set for trial on 8 September 2026 (NYU Journal of Intellectual Property & Entertainment Law). Between the two, 2026 is the year the "is this transformative fair use or is this infringement at the output level" question stops being theoretical.
Our marketplace desk put the practical takeaway this way:
"Everyone building fan content assumed the risk was the same shape for a face and a costume. It isn't. One is a person who can say yes or no about themselves. The other is a studio's IP department, and studios don't licence informally — they licence on paper, with terms, or they sue."
| Real-person fan content | Copyrighted-character fan content | |
|---|---|---|
| Whose right is it | The individual depicted | The IP rights holder (studio, publisher) |
| Legal basis | Right of publicity, privacy, in some cases defamation | Copyright, trademark |
| Current test case | TAKE IT DOWN Act (US, in force), Online Safety Act (UK, in force) | Disney/Universal v. Midjourney (in litigation, no ruling yet) |
| What "consent" fixes | Everything, if it's real, specific and revocable | Nothing — the subject can't licence a right they don't hold |
| What clears it | A boundary set from that specific person | A licence from the rights holder, separately |
| Worst combination | A real, identifiable person depicted as a character they haven't licensed either half of |
Table: the two problems bundled inside "AI generated fan content", and why fixing one does nothing for the other.
The worst version of fan content, legally, is the one that feels most classically "fandom": a real cosplayer's likeness generated as a specific franchise character. That request needs the cosplayer's consent for their face and the studio's clearance for the costume design, and getting one is not a substitute for the other. It's why platforms that only ask "did the subject agree?" are answering half the question, and it's why LikeWard's cosplay and character boundary type is scoped to the creator's likeness specifically — it is not, and does not claim to be, a clearance for the underlying character IP, which stays the requester's problem to solve separately.
The fix for the real-person half is the part LikeWard actually builds: a creator publishes a boundary set naming which contexts they permit — portrait, fan-art-stylised, cosplay among them — and a realism ceiling capping how photoreal any output can be. A generation request is evaluated against that live boundary set before the model is ever called, on the website, the licensed-likeness API and the MCP server alike, so there's no looser back door for automated traffic. A request that falls outside the published limits is refused, naming the rule it broke.
Every authorised output carries a signed provenance manifest: which creator, which boundary-set version, who generated it, and when. That's not decoration — it's the artifact a card network, a platform's trust-and-safety desk, or the creator themselves would need to produce on request, and it's what makes a later revocation mean something. We go through the mechanism in more depth in Content Provenance, Explained.
None of that touches the character-IP half. If a fan wants a specific franchise costume rather than a generic stylised context, that clearance sits with the rights holder, and no likeness platform — including this one — can manufacture consent it doesn't have the authority to give.
For scale, the market this sits inside is not small: the global creator economy is estimated at roughly $248 billion in 2026, with more than 207 million people worldwide identifying as creators (Coherent Market Insights; DemandSage). A meaningful share of that activity is fan-facing. The absence of a clean licensing answer for it isn't a niche gap.
Assuming "stylised" means "safe". A low realism ceiling limits how photoreal output can be; it says nothing about whether the underlying context was actually permitted.
Treating a public figure as fair game. Public profile narrows some privacy claims. It does not remove a right-of-publicity claim, and it does nothing for a takedown request under the TAKE IT DOWN Act, which doesn't carve out an exception for people who are already well known.
Conflating disclosure with consent. An "AI-generated" watermark and a signed consent manifest answer different questions. Regulators and card networks are asking for the second one.
Ignoring the character half entirely. The most common mistake we see is a platform that builds real, careful likeness consent and then assumes that covers any costume a fan requests. It doesn't, and Disney/Universal v. Midjourney is the live reminder of why that gap matters.
Sharing after a revocation. The manifest exists precisely so a "this was authorised at the time" defence is checkable — and so is its opposite, once the record flips to withdrawn.
Want to see what a real boundary set looks like before you rely on one? Start with the licensed fan art use case, then read the boundary configuration guide for how creators actually set their limits. If you're weighing a licensed platform against an unlicensed generator, The Real Cost of Unlicensed AI Generation covers what's actually at stake on the other side of that choice.