Who owns what AI makes

Every company that uses AI to write, draw or film meets two questions sooner or later. Who owns what comes out. And who is liable if what went in, or what comes out, belongs to someone else.

Courts and authorities have answered parts of both in the last eighteen months. This is where things stand in October 2026. It is an overview, not legal advice.

Output without an author

The clearest answer comes from the United States. In January 2025 the US Copyright Office published the second part of its report on copyright and AI, on what can be protected. Copyright does not extend to purely generated material, or to material where a person had too little control over the expressive elements. With today's technology, prompts alone do not give that control. A person who uses AI as a tool keeps the rights in their own expression that can be seen in the output, in a creative selection or arrangement of it, and in creative changes to it.

In March 2025 the federal appeals court in Washington confirmed that a work must be authored in the first instance by a human being, in a case about a picture whose only named author was an AI system.

Switzerland has no ruling of the same weight. The Swiss Federal Institute of Intellectual Property wrote in September 2026 that it is examining exactly this question, how much human involvement a work in the legal sense requires, and that it is drafting copyright rules for AI after parliament adopted a motion by Councillor of States Petra Gössi.

The practical consequence is the same everywhere. An image made from a prompt and used unchanged may belong to nobody, and a competitor can copy it. What a company can own is the human part: the idea, the choices, the edits, the composition, and the record that shows them.

The training cases

On 29 September 2026 the US Court of Appeals for the Third Circuit decided Thomson Reuters against ROSS Intelligence. ROSS had used Westlaw's headnotes, the short summaries above court opinions, to train a legal search tool. The court held that the headnotes are protected and that the use was not fair, because ROSS built a direct competitor for a highly similar purpose. It also found harm to a developing market for licensing headnotes as AI training data. The court stressed that the ROSS system was not generative and only returned existing passages, so the larger questions stay open.

Those questions sit in California and New York. In June 2025 Judge William Alsup ruled in Bartz against Anthropic that training on books was fair use, but that downloading millions of pirated books for a central library was not. The case settled for 1.5 billion dollars, about 3'000 dollars per book, and the court granted final approval on 20 July 2026.

The New York Times case against OpenAI and Microsoft is the one to watch. In September 2026 the Times asked the court to decide in its favour without a trial, and the US Department of Justice filed a statement arguing that training a large language model is a transformative use. There was no ruling by the end of September.

Europe and Switzerland

On 4 November 2025 the High Court in London ruled largely for Stability AI against Getty Images. Getty had dropped its main claim about training during the trial. On what remained, the judge held that a model which does not store or reproduce the works it was trained on is not an infringing copy, even though a model can count as an article under UK law. Getty has permission to appeal that point.

A week later the Regional Court of Munich decided the other way. GEMA, the German music rights society, showed that ChatGPT reproduced song lyrics on simple prompts. The court held that the model contained copies of the works, that OpenAI should have licensed them, and that the exception for text and data mining does not justify storing and putting out the lyrics.

The EU AI Act adds a duty for the model makers. Since August 2025 providers of general purpose models must have a policy to comply with EU copyright law and publish a summary of the content they trained on. In Switzerland, the institute has said that the use of works for AI will fall under Article 10 of the Copyright Act, and it is weighing models with and without payment to authors, each with an opt out.

For a company that only uses AI, the risk sits less in training than in output. An image or a text that reproduces a protected work is a problem for whoever publishes it.

Faces, voices and brands

Generated images carry two further risks. The first is trademarks. Getty's one success against Stability AI concerned Getty watermarks that appeared in images made with older versions of the model. The judge called her findings historic and extremely limited, but the principle is plain: a brand that appears in a generated image is a brand in use.

The second is people. The FDPIC expects any manipulated face, image or voice of an identifiable person to be clearly marked, and the EU AI Act has required deepfakes to be labelled since August 2026.

Our rule: no real person, no other company's brand and no recognisable work in a generated image unless the rights are cleared, and no prompt that asks to copy a living artist.

What to put in contracts

With clients, say where AI is used and which parts of a deliverable are human work. Transfer the rights you hold, and say plainly that purely generated elements may not be protected, so exclusivity cannot be promised for them. Agree who checks the output for third party rights, brands and faces, and who labels what.

With suppliers, settle three things: whether your prompts and files train their models, who stands behind the output if a third party claims it, and what records they keep. Some large providers offer cover. Microsoft promises to defend commercial customers sued over the output of its Copilot services and to pay adverse judgments, as long as they used the guardrails and content filters built into the products. Google indemnifies customers of several generative services unless they tried to infringe on purpose. The cover follows the conditions, so read them.

Keep the record. Prompts, sources, choices and edits are the evidence of the human part, and the evidence of good faith if something goes wrong.

How we work

Ideas, taste and judgement are the parts of creative work we keep with people, while AI speeds up the making. The law now points the same way: what a person chooses and changes is what a company can own.

Reel builds every style from real references, and a person picks every image and every clip that is used. Each one sits in a media library with every version and the recipe behind it. Studio commits a company's content to its own repository with every publish and keeps every action in an audit log. Twin is delivered as source code the company owns. None of that decides a court case. It keeps the record a court would ask for.

The law on AI output will keep moving through appeals in London, Philadelphia and New York. A company that knows what it made, how, and with whose material, is ready for whichever way they go.