Who owns AI-generated images and text? Copyright in plain English

Can you copyright something an AI made?

In the United States, not on its own. Copyright there protects work made by people, and the courts and the Copyright Office have held that line firmly. This article explains the general picture; it is not legal advice, and the rules differ from country to country.

The clearest test case is Stephen Thaler's. He asked the US Copyright Office to register an image called "A Recent Entrance to Paradise", listing his AI system, the "Creativity Machine", as the author. The Office refused. On 18 March 2025 a federal appeals court in Washington, DC agreed, holding that the Copyright Act requires "human authorship", according to the law firm Goodwin. The Supreme Court declined to hear the case on 2 March 2026, so that ruling stands.

In January 2025 the Copyright Office published a report on exactly this question. It found that:

• Writing a prompt does not make you the author of what comes back.
• Using AI as a tool, or including AI material in a larger human work, does not stop the human parts being protected.
• What you add yourself can be protected: your own text or drawing that is visible in the result, or creative changes and arrangements you make to the output.
• The Office saw no case for a new law giving AI output extra protection.

If you register a work in the US that contains more than a small amount of AI-generated material, you must say so and describe what the human did, according to the American Bar Association. Leaving it out can get the registration cancelled.

How much human input changes things?

Quite a lot, but in a specific way. Two US registrations show where the line falls.

In February 2023 the Copyright Office reviewed Zarya of the Dawn, a comic book whose images were made in the generative AI tool Midjourney. Its decision letter kept protection for the author's text and for the way she selected and arranged text and pictures, but excluded the images themselves. Small touch-ups to one character's mouth were "too minor and imperceptible" to count. The Office added that substantial edits to an AI image could count as human authorship.

In January 2025 the Office registered A Single Piece of American Cheese, an image made with the tool Invoke, after first rejecting it, IPKat reports. The creator picked one of three outputs, then used "inpainting" to regenerate parts of it piece by piece. Protection was granted for the "selection, coordination, and arrangement" of the AI material, not for the AI material as such.

Effort alone is not enough. The artist Jason Allen entered prompts "at least 624 times" to make "Theatre D'Opera Spatial", according to the government's court filing, and the Office still refused to register it unless he disclaimed the AI parts. He is suing; as of mid-September 2026 the Colorado court had not yet ruled.

What about Europe, the UK and China?

Czech Republic. On 11 October 2023 the Prague Municipal Court threw out a claim over an image made with a simple prompt, Bird & Bird reports. It said such an image is not a person's creative work, and that a prompt is at most "a theme or idea", which copyright does not protect. The claimant also could not prove which prompt made the image, a lesson in keeping records.

United Kingdom. UK law has an unusual rule for "computer-generated works" with no human author: the author is the person who made "the arrangements necessary" for the work. In a report on 18 March 2026, the UK government proposed removing that rule, while saying works made with AI assistance should stay protected, according to Herbert Smith Freehills Kramer. For now it remains law.

China. China went the other way in one notable case. On 27 November 2023 the Beijing Internet Court ruled that an image made with Stable Diffusion was protected, because the plaintiff had chosen more than 150 prompts, set parameters and refined the result repeatedly, the Kluwer Copyright Blog reports. The court still said the AI itself cannot be an author.

Doesn't the AI company say the output is mine?

Yes, mostly. OpenAI's terms for Europe say you "own the Output" and assign you any rights OpenAI has in it. They also warn that output "may not be unique": someone else may get something similar. Anthropic's consumer terms assign you its rights in outputs "if any". Midjourney's terms say you own what you create "to the fullest extent possible under applicable law", but companies with more than US$1 million in yearly revenue need a Pro or Mega plan to own their images, and images are public and remixable by default.

Those phrases matter. A company can hand over the rights it has, but it cannot create a copyright the law does not recognise. "You own it" in the terms mostly means the company won't claim it.

What about being sued? Some companies promise to defend customers, but only paying business customers:

• OpenAI announced "Copyright Shield" in November 2023 for ChatGPT Enterprise and its developer platform, Loeb & Loeb reports.
• Microsoft promised in September 2023 to defend commercial Copilot customers, provided they use its built-in filters and don't try to produce infringing material, according to its own announcement.
• Google offered a similar indemnity in October 2023 for listed Google Cloud and Workspace products, according to its Cloud blog.

On consumer plans it runs the other way. Anthropic's consumer terms, for example, ask you to indemnify Anthropic for claims arising from your use of outputs.

Is it legal to train AI on copyrighted work?

This is the other half of the fight, and it is still being settled in court. AI models learn from huge amounts of training data, much of it books, articles and pictures made by people who were not asked.

• Authors v Anthropic (US). In June 2025 Judge William Alsup ruled that training on legally bought books could be allowed, but keeping pirated copies was not, JURIST reports. Anthropic settled for US$1.5 billion, about US$3,000 per book, and a federal court gave final approval on 20 July 2026, according to the Authors Guild.
• Authors v Meta (US). On 25 June 2025 Judge Chhabria ruled for Meta against 13 authors, but said they had "made the wrong arguments" and failed to prove harm to their market, the Authors Guild notes.
• The New York Times v OpenAI and Microsoft (US). Filed in December 2023. Both sides asked the judge to decide the central "fair use" question in September 2026, and no ruling had been made as of mid-October, according to AI Lawsuit Tracker.
• Disney and Universal v Midjourney (US). Filed on 11 June 2025 over characters such as those from Star Wars and The Simpsons, NBC News reports. It is still in pretrial stages with no trial date.
• Getty Images v Stability AI (UK). On 4 November 2025 the High Court rejected Getty's main copyright claim, finding the Stable Diffusion model did not contain copies of Getty's photos, though Getty won narrow trademark points over its marks appearing in some generated images, the CMS case tracker reports. Getty has permission to appeal; we found no appeal ruling yet.
• GEMA v OpenAI (Germany). On 11 November 2025 a Munich court ruled that song lyrics memorised by OpenAI's models were infringing copies, and that the EU's data-mining exception did not cover them, according to Taylor Wessing.

The direction so far: US judges have been open to training on lawfully obtained material, while piracy and outputs that copy work closely are where companies lose. Little of this is final.

What does the EU AI Act ask of AI companies?

Since 2 August 2025, makers of general-purpose AI models, the large systems behind chatbots, have had to keep a copyright policy, respect creators who opt out of data mining, and publish a summary of their training content on an EU template, Clifford Chance explains. Fines of up to 3% of worldwide turnover or €15 million became enforceable on 2 August 2026.

Separately, under Article 50, AI tools must mark their images, audio, video and text as machine-made, and people who publish realistic deepfakes must say so. Our EU AI Act explainer covers the dates, and our guide to spotting AI-generated content shows what those labels look like in practice.

What should I actually do?

• Assume a raw AI image or text has no copyright. If your logo or mascot is pure AI output, a competitor may be able to copy it. Add real human work if you need something you can defend.
• Keep records of your own work. Save prompts, drafts, versions and your edits with dates. The Czech claimant lost partly because he couldn't prove how his image was made.
• Check for famous characters, logos and real people. Generators can reproduce trademarked characters and recognisable faces. Using them commercially can infringe trademarks, copyright or someone's rights over their likeness, whatever the tool's terms say.
• Book covers and self-publishing. Amazon's Kindle Direct Publishing has required authors since September 2023 to disclose AI-generated text, images or translations, including covers, The Bookseller reported. AI help with editing does not need disclosing.
• School and university work. Copyright is rarely the issue here; your school's rules on AI are. Check them, and say what you used.
• Business use. Read the plan you're on. Indemnities usually cover only paid business tiers and only if you follow the rules. If you're in the EU and publish realistic AI images, label them.
• Don't paste in other people's work you don't have rights to. Microsoft's commitment, for one, excludes customers who feed in material they lack rights to. For what happens to what you type, see our guide to your data and chatbots.