Can AI companies train on your work? Copyright, fair use and opt-outs

If you write, draw, photograph or compose, there is a fair chance your work has already passed through an AI training dataset. This page explains what copyright law says about that, part of our guide to AI and copyright. It covers the fair use test, what judges have actually decided, the settlements and licensing deals that followed, and what an opt-out can and cannot do.

The short version: training is not automatically legal and not automatically infringement. Courts look at the facts of each case under the four fair use factors in the Copyright Act. If you want the factors explained from the ground up first, start with our page on what fair use means and how the four factors work.

01Why training raises a copyright question

Copyright gives an author the exclusive right to make copies of a work. Building a large AI model usually means copying huge numbers of works: downloading them, storing them, cleaning them and feeding them through training. Each of those steps can involve a copy.

AI developers argue that these copies are an intermediate step. The finished model does not hold a readable library, they say; it holds statistical patterns. Authors and artists answer that the model was built from their expression without permission or payment, and that it can produce work that competes with theirs.

Both sides agree on one thing: the legal fight is mostly about fair use. In its Part 3 report on generative AI training, released in pre-publication form on May 9, 2025, the US Copyright Office said several stages of AI development involve copying that can infringe unless a defense such as fair use applies. As of October 2026 that report is still a pre-publication version.

Training is a separate question from ownership. Whether anyone owns what an AI system produces is covered on our page about who owns AI-generated content.

02Fair use and the four factors applied

Section 107 of the Copyright Act lists four factors. Courts weigh them together; no single factor decides a case. Here is how they have been applied to AI training so far.

1. Purpose and character of the use

The key question is whether the use is transformative: does it serve a different purpose from the original, or does it stand in for it? Courts now read this factor through the Supreme Court's 2023 decision in Andy Warhol Foundation v. Goldsmith, which asks whether the new use shares the original's purpose and is commercial. Training a general model that writes new text has been called highly transformative. Training a tool to do the same job as the work it copied has not.

2. Nature of the work

Creative works such as novels and paintings sit closer to the core of copyright than factual works. This factor rarely decides an AI case on its own.

3. Amount used

Training usually copies whole works. Courts have asked whether copying that much was reasonable for the purpose. Where the purpose was transformative, copying entire books was accepted. Where it was not, copying full works counted against the developer.

4. Effect on the market

This is the factor most AI cases turn on. Does the use harm sales of the original, or a market for licensing it? Judges have disagreed on whether "market dilution", meaning a flood of AI-made works that compete with human ones, counts as harm.

03What courts have said so far

Three sets of decisions shape the current picture. None of them settles the question for everyone, and each depends on its own record.

Bartz v. Anthropic (N.D. Cal., June 23, 2025)

In his June 23, 2025 order on fair use, Judge William Alsup held that using books to train Anthropic's models was "exceedingly transformative" and a fair use. Scanning print books Anthropic had bought was also fair use. But the order held that downloading pirated copies to build a permanent central library was not excused by fair use, and sent that part toward trial.

Kadrey v. Meta (N.D. Cal., June 25, 2025)

Two days later, Judge Vince Chhabria granted Meta summary judgment against thirteen authors in an order filed June 25, 2025. He was explicit about its limits: the ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful." The authors lost because they made the wrong arguments and did not build a record on market dilution, which he suggested could be the stronger theory.

Thomson Reuters v. Ross Intelligence (D. Del. 2025; 3d Cir. 2026)

On February 11, 2025, Judge Stephanos Bibas held that Ross's use of Westlaw headnotes to train a legal search tool was not fair use, noting that only non-generative AI was before him. On September 29, 2026, the Third Circuit affirmed, calling Ross's use "minimally transformative, at best" because both products served the same purpose. The court noted that, unlike the generative models in other pending cases, Ross's platform "cannot generate original expression."

For case-by-case detail on the wider litigation, including the New York Times suit against OpenAI and the artists' case against Stability AI, see the case summaries on Stable Diffusion Frivolous. This page sticks to what the decisions mean for creators.

04Settlements and licensing

The largest result so far came from a settlement, not a verdict. After the June 2025 ruling left the pirated-library claims open, the Bartz parties settled as a class action. In an order dated July 20, 2026, Judge Araceli Martínez-Olguín granted final approval of the $1.5 billion settlement.

The order describes an estimated payment of about $3,000 per work, before costs and fees, across a list of 482,460 works. It also notes that by April 16, 2026, claims had been made for 440,490 of those works. The settlement covers past conduct involving the listed books; it does not decide whether AI training in general is lawful.

Alongside the lawsuits, many publishers, news organizations, stock image libraries and music companies have signed licensing deals with AI developers. In the Ross case, the Third Circuit treated "the market for licensing headnotes as text to train AI" as a real market that copying could harm. That reasoning matters for creators: the more a licensing market exists, the stronger the fourth-factor argument against unlicensed copying.

05Opting out: what works and what does not

There is no single opt-out switch for AI training. Different tools do different jobs.

• robots.txt. A text file on a website that tells named crawlers which pages to skip. Many AI companies publish crawler names and say they honor it. In the United States it is a voluntary convention, not a rule written into the Copyright Act, and it does nothing for copies already collected or posted elsewhere.
• Site and platform settings. Some hosting platforms, portfolio sites and social networks offer a setting that asks AI crawlers to stay away or keeps your posts out of the platform's own training. Read the platform's terms; the setting may only cover future use.
• Machine-readable rights reservations in the EU. The EU AI Act requires providers of general-purpose AI models to have a policy to identify and comply with a rights reservation made under Article 4(3) of the 2019 Copyright Directive. Article 53 of the Act (Regulation 2024/1689) sets this duty, and the general-purpose model rules have applied since August 2, 2025.
• Content credentials. Provenance labels can carry a statement about whether you allow AI training. They are a signal, not a lock. Our page on content credentials and C2PA labels explains how they work.

What does not work: a "no AI" line in a caption on its own, or a watermark alone. They may express your wishes, but nothing forces a crawler to read them.

06What creators can do now

1. Register your important works. Registration creates a public record of your claim, and for US works it is generally needed before you can file an infringement suit.
2. Set your opt-outs where you control them. Update robots.txt on your own site and check the AI settings on platforms where you post.
3. Keep records. Dated drafts and originals help if you later need to show what you made and when.
4. Watch for outputs that copy you. If an AI tool reproduces your work closely and someone posts it, you can use the takedown process explained in our guide to sending a DMCA takedown notice.
5. Check class actions and claims deadlines. Settlements set strict dates. If you think your work is covered, read the official notice for that case.
6. Get advice where the stakes are real. If you are thinking about a lawsuit or a licensing deal, talk to a lawyer licensed where you live.

For a wider view of how courts are applying fair use, see the editorial overview at fair use in the age of AI.