← Blog

If the model was trained on others' work, is my output infringing?

2026-11-14 · OpenStela

Not automatically. Two questions are tangled together here, with different answers. The first is whether it was lawful to train the model on other people's work. That one is being fought in court, mostly between rights holders and the companies that built the models, and nobody has a final answer yet. The second is whether the specific image, song or video you made infringes someone's copyright. That one is older and more familiar: it mostly comes down to whether your output copies a particular protected work closely enough.

So "the model was trained on scraped art" does not, by itself, make everything it produces infringing. And "the court said training was fair use" does not make your output safe either. Below are the official sources, quoted with dates, with no prediction of where the cases end up. We only cover rulings we read in full, not later events in each case.

This is general information, not legal advice. If you are about to sell or license something that might resemble an existing work, talk to a lawyer where you publish.

Two questions, not one

The clearest official statement of the split is the US Copyright Office's report on generative AI training, Part 3 of its Copyright and Artificial Intelligence series. The Office released it as a pre-publication version on May 9, 2025, and says on its AI page that the final version is expected without substantive changes to the analysis or conclusions (Part 3, pre-publication PDF).

Most of the report is about training. On outputs, three things stand out:

• Copying in training isn't enough on its own. Page 30, footnote 177: "Use of the works in preparing the model is not enough—there must be evidence of substantial similarity." The Office is talking about whether the model's weights themselves infringe, but the test it applies is the ordinary one: substantial similarity to a particular work.
• Outputs that replicate a work are a problem. Page 31: "Generative AI models sometimes output material that replicates or closely resembles copyrighted works." It lists near-exact replicas of film stills, "copyrightable characters" and news text, and says: "Such outputs likely infringe the reproduction right and, to the extent they adapt the originals, the right to prepare derivative works."
• Who is liable for an infringing output is still open. The same page says these issues, "including enforcement challenges and the allocation of potential liability, will be addressed in a later Part of this Report." Footnote 110 (page 19) calls it Part 4. As of the day we checked, the Office's AI page lists Parts 1 to 3 only.

One more line matters if you work "in the style of" someone. Page 65 describes concerns about "AI outputs that imitate a creator's style, which copyright does not protect as a separate element". The Office raises style as a possible market harm when it weighs whether training is fair use. That's different from saying a style-alike output infringes.

What the US courts have actually said

Three US trial-court decisions get cited most. All are on govinfo.gov. None is about a creator's output.

Bartz v. Anthropic (N.D. Cal., Order on Fair Use, June 23, 2025). Judge Alsup held that using books to train the models "was exceedingly transformative and was a fair use", but that "Anthropic had no entitlement to use pirated copies for its central library" (page 9). On outputs he was explicit (pages 11–12): "Authors do not allege that any LLM output provided to users infringed upon Authors' works." And: "if the outputs seen by users had been infringing, Authors would have a different case. And, if the outputs were ever to become infringing, Authors could bring such a case. But that is not this case." (order, PDF)

Kadrey v. Meta (N.D. Cal., June 25, 2025). Judge Chhabria ruled for Meta on the record in front of him, then warned against reading too much into it (page 5): "this ruling does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful. It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one." On outputs, the order notes that even with prompts designed to make the model regurgitate, the experts couldn't get it to produce more than 50 words and punctuation marks from the plaintiffs' books (order, PDF).

Thomson Reuters v. Ross Intelligence (D. Del., February 11, 2025). Here training went the other way: Judge Bibas held Ross's use "not transformative" and rejected fair use. He also limited the ruling: "It is undisputed that Ross's AI is not generative AI", and "only non-generative AI is before me today" (opinion, PDF). On May 23, 2025 he explained why he had granted Ross's motion for an interlocutory appeal, on fair use among other questions, to the Third Circuit (memorandum, PDF).

What these rulings share: they're about the developer's copying, and the judges kept noting that infringing outputs weren't before them. They don't say outputs are fine. They say outputs weren't the question.

Outside the US: the UK, and a LoRA case in China

Getty Images v Stability AI ([2025] EWHC 2863 (Ch), High Court of England and Wales, November 4, 2025). Getty dropped its training claim because there was no evidence the training happened in the UK. It also dropped its output claim, because Stability had blocked the prompts behind the infringing examples, and Getty acknowledged that this had "substantially achieved" the relief it would have got (paragraph 9). What was left was whether the model itself is an "infringing copy". The judge said no: "the model weights are not themselves an infringing copy and they do not store an infringing copy" (paragraph 600) (judgment, PDF).

The Ultraman LoRA case. The Copyright Office report (page 81, footnote 454) describes an appeal in which the Hangzhou Intermediate People's Court upheld a ruling that a Hangzhou AI company "should delete the allegedly infringing Ultraman LoRA model". We didn't read the Chinese judgment itself, so we only repeat the report's description. Still, a LoRA trained to reproduce one recognisable character is about as close as you get to an output problem built into a model.

Taiwan: TIPO's answer is about your output

Taiwan's Intellectual Property Office (TIPO) answered this almost head-on in two email interpretations, with its usual caveat: copyright is a private right, and infringement is for the courts to decide on the facts.

• Email 1140625 (June 25, 2025): a model may have learned from large amounts of other people's text. TIPO says the chance that an output is substantially similar to an original is not high, but it can't be fully ruled out. If it is, and you use it commercially, fair use under Articles 44–65 may not apply. TIPO suggests checking the AI developer's user rules, including whether you get rights to outputs that contain others' works (TIPO, 電子郵件1140625).
• Email 1140829 (August 29, 2025): copying works during training involves reproduction and needs permission unless fair use applies. But if an output isn't substantially similar to any single training image, or only shares unprotected facts, it is not an adaptation, and the output "自不生侵害原始訓練資料著作權問題", meaning it raises no question of infringing the training data's copyright. The letter also warns that if you use AI as a tool to rework an existing work, that can still be an adaptation needing permission (TIPO, 電子郵件1140829).

Fair use in Taiwan is weighed under Copyright Act Article 65's four factors: purpose and character, nature of the work, amount used, and effect on the market (著作權法 第 65 條).

Where the risk actually sits for you

Together, the sources move your risk away from "what was the model trained on" toward "what did I make, and how close is it to something specific". Some habits follow:

1. Compare the output to what you were thinking of. If you prompted with a film, a franchise character or a named artist's piece in mind, put them side by side. The Office's own examples of likely infringement are recognisable characters and near-copies.
2. Be careful with names in prompts. Asking for a specific character or a specific existing work is the quickest way to get an output that resembles it. TIPO's 1140829 warning about reworking existing works applies here.
3. Think twice about LoRAs built on one character. A model whose whole purpose is reproducing an existing character is a different thing from a general model. The Ultraman appeal is the example the Copyright Office chose to cite.
4. Read your tool's terms. TIPO itself suggests it. Look at the current terms for your plan rather than a summary: for example Suno, Runway or Civitai.
5. Keep your own process. Your prompts, drafts, edits and references, with dates. If someone says your piece copies theirs, the question becomes what you did and when, and your prompts and edits are part of that story.

Whether your output is protected for you is another question: see how much you have to do for an AI work to be protected.

What a record with us does, and doesn't

When you register a work on OpenStela, the authorship record holds the file's fingerprint, the date, the attribution you chose, whether it was made with AI tools, Which tools you used, and, for AI-assisted work, Your contribution (the form won't accept an AI-assisted work without it). That record is anchored on chain in the nightly batch. You also choose a rights declaration, My own work or I hold a licence, and the record shows it as your declaration, nothing more.

Here's what it doesn't do. We don't compare your file against training data, other people's work, or anything else. The platform records declarations and facts and doesn't judge what a work looks like. So a registration doesn't show your output is non-infringing, and it doesn't clear anything. What it gives you is a dated, verifiable record of a specific file, under a name, with the tools and the contribution you wrote down at the time. If a dispute comes up, that's one more piece of evidence about your process. What a character ID proves, and what it doesn't has the full version, and someone copied my AI character covers the case where you're the one being copied.

Register a work at app.openstela.io/works.