AI and copyright: who owns what, explained simply

Products, laws and numbers here change fast. Checked October 2026.

In plain words

Copyright gives creators control over copies of their books, songs, photos and code. Generative AI raises two new questions about it, and the world has not fully answered either.

First, the input. AI models are trained on enormous amounts of text and images, much of it copyrighted and used without asking. AI companies say learning from work is allowed, like a student reading in a library. Many authors, artists and news publishers say it is copying at industrial scale and they deserve permission and payment. Dozens of court cases are under way.

Second, the output. If you make a picture with an AI tool, who owns it? In the US, courts and the Copyright Office say a human must be the author, so a picture made purely from a prompt usually gets no copyright. Your own edits and arrangement can. Other countries are still working this out. We are not lawyers; this page is not legal advice.

Everyday example

The music student and the cover band

A student learns sitar by listening to thousands of recordings, then composes her own raga. Nobody calls that theft. A band that records a near-exact copy of a hit song and sells it needs a licence. AI sits awkwardly between the two: it learns from huge numbers of works like the student, but it does so by copying them into a computer, and sometimes it can repeat parts almost exactly.

Where the example breaks: A student learns from a few thousand pieces over years. An AI copies millions in weeks, as part of a business, which is exactly why courts are unsure whether the student comparison holds.

Go deeper

How it actually works

As of October 2026, here is the state of play. Courts move case by case, so these are snapshots with dates.

US: training and fair use
US law allows some unlicensed use under "fair use". In June 2025, a federal judge in Bartz v. Anthropic held that training on books was fair use, calling it "exceedingly transformative", but keeping millions of pirated books was not. The case settled for $1.5 billion, approved in July 2026. In September 2026, a US appeals court ruled against a legal AI start-up, Ross, that trained on Thomson Reuters' Westlaw headnotes, short case summaries. The New York Times' case against OpenAI and Microsoft is still pending.
US: who owns AI output
The Copyright Office said in January 2025 that prompts alone are not enough for authorship, because they "function as instructions that convey unprotectable ideas". In March 2026 the Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the rule that a work made entirely by AI has no human author and no copyright.
EU
EU law lets rights holders opt out of text and data mining. Under the AI Act, from 2 August 2025 makers of general-purpose models must have a copyright policy and respect machine-readable opt-outs such as robots.txt. In November 2025, a Munich court ruled largely for the music rights group GEMA against OpenAI over song lyrics; the ruling is not final.
India
In ANI v. OpenAI, the Delhi High Court on 24 July 2026 refused the news agency an interim injunction. Its early view, as reported: storing the material fell under an exception in the Copyright Act, and ChatGPT's answers were not substantially similar to ANI's articles. This is not a final ruling; ANI has appealed and the case continues. Separately, a DPIIT working paper (December 2025) proposed a mandatory licence for AI training with royalties paid to creators. It is a proposal, not law.
Memorisation
Models sometimes reproduce training text or images closely. That output side is often the strongest part of a creator's case, and why AI makers add filters against repeating copyrighted material.

Largest settlement so far: About $1.5 billion in Bartz v. Anthropic, around $3,000 per book, approved July 2026
Human authorship (US): Required. Supreme Court declined review in Thaler v. Perlmutter on 2 March 2026
India's first big case: ANI v. OpenAI, Delhi High Court. Interim relief refused 24 July 2026; appeal listed for 8 December 2026
Checked: 28 October 2026. New cases are filed every month

Should AI companies need permission to train on copyrighted work?

One view

Yes, ask and pay

Creators' work is the raw material that makes these products valuable. Training without consent or payment takes that value, and AI output can then compete with the very people it learned from.

The other view

No, learning is not copying

Models learn patterns, not stored copies, much as people learn from what they read. Requiring permission for every work would make training impractical and favour only the largest companies.

Most likely: Courts in different countries are splitting the difference: training itself is sometimes allowed, while pirated sources and near-copies in outputs are not. This is an open legal and political question, and the answer may differ by country.

Questions people ask

Can art made with AI be copyrighted?

In the US, not the parts the AI generated from a prompt alone. People can protect their own creative contributions, such as edits, selection and arrangement. Rules in India and elsewhere are less settled. We are not lawyers; take advice for anything commercial.

Is it legal for AI to train on someone's work?

It depends on the country and is being decided in court. A US judge found training on lawfully bought books to be fair use in one major 2025 case. The EU lets creators opt out of text and data mining. India's position is still forming.

Can a website owner stop AI from training on their site?

Partly. Owners can block known AI crawlers in the site's robots.txt file, and in the EU, makers of general-purpose AI models must respect such machine-readable opt-outs. It does not remove what was collected before, and not every crawler obeys.

Who owns the output of ChatGPT?

OpenAI's terms say it assigns its rights in output to the user, and many AI services do similar. But a company can only pass on what the law allows, and in the US purely AI-generated content may not be protected by copyright at all.

What is the ANI vs OpenAI case?

Indian news agency ANI sued OpenAI in the Delhi High Court in 2024, saying ChatGPT was trained on and reproduced its news. In July 2026 the court refused an interim injunction. The main case and ANI's appeal continue.

AI and copyright