Is AI Training Fair Use? A 2026 Scoreboard

Sometimes yes, sometimes no. As of mid 2026, US courts have split the question along one line: how you got the data. Training an AI on works you bought or accessed legally has been ruled fair use in two 2025 cases. Downloading pirated copies to build a permanent library was ruled not fair use in the same case. What you did with the data matters. Where the data came from matters just as much.

What have US courts actually decided about AI training?

Three US rulings in 2025 gave builders their first real signal. This is decided law from district courts, not a proposal. All three are being appealed or could be, so treat them as the current score, not the final whistle.

Bartz v. Anthropic (June 23, 2025, Judge Alsup, N.D. Cal.): Training a large language model on lawfully bought books is "exceedingly transformative" and a fair use. So is scanning print books you own into digital copies. But downloading over 7 million pirated books to keep a permanent library is not fair use.
Kadrey v. Meta (June 25, 2025, Judge Chhabria, N.D. Cal.): Training the Llama models on books was ruled fair use. The win was narrow. The judge said the authors lost because they made weak arguments, not because Meta was clearly in the right.
Thomson Reuters v. Ross (Feb 11, 2025, Judge Bibas, D. Del.): Not fair use. Ross copied Westlaw legal headnotes to build a competing research tool. The court found direct infringement of 2,243 of the 2,830 headnotes at issue.

Sources: Bartz v. Anthropic order (PDF) | Kadrey v. Meta order | Thomson Reuters v. Ross opinion (PDF)

Why did Anthropic win the fair use point but still pay $1.5 billion?

Because the court split the behavior in two. Using books to train the model was fair use. Building a library from pirated copies was a separate act, and that act was infringement. Anthropic had downloaded pirated books before it ever trained on them. The training win did not erase the piracy.

The parties settled the piracy piece. Anthropic agreed to pay at least $1.5 billion. That works out to roughly $3,000 per work across about 482,000 works.

A judge gave preliminary approval on September 25, 2025. Final approval was still pending. It is one of the largest publicly reported copyright settlements.

The lesson for a builder is blunt. The source of your data can cost more than the model you build with it.

Sources: Publishers Weekly on preliminary approval | Copyright Alliance settlement summary

Does the Meta ruling mean training on books is always safe?

No, and the judge said so in the ruling itself. Judge Chhabria wrote that the decision "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 point that these specific authors "made the wrong arguments and failed to develop a record in support of the right one."

He pointed to a "market dilution" theory, the idea that AI floods the market with cheap substitutes. He said such a theory "will often cause plaintiffs to decisively win the fourth factor."

So the door is open. A future case with better evidence could flip the result.

Sources: Kadrey v. Meta order

When is AI training NOT fair use?

The Ross case is the clearest "no" so far. Three things sank it.

The use was commercial, not research. It was not transformative, because Ross used the material to do the same job as the original. And the product competed directly with the source it copied from.

Add pirated sourcing on top, as in the Anthropic piracy finding, and you have the worst case.

The pattern across rulings: copying to compete with the thing you copied, or copying from a pirate source, is where fair use breaks.

Sources: Thomson Reuters v. Ross opinion (PDF)

What did the Getty v. Stability AI ruling in the UK change?

Less than the headlines suggested. On November 4, 2025, the UK High Court handed down its judgment. The UK has no "fair use" doctrine, so this is not a fair use case.

Getty dropped its main copyright and database claims before the end of trial. It could not show the training happened in the UK.

The court dismissed the secondary copyright claim. It held that a model like Stable Diffusion contains no stored copies of the training images. Getty won only a narrow trademark point on early model versions.

The takeaway for a builder: where training physically happens can decide which country's law even applies.

Sources: Getty Images v. Stability AI judgment (PDF)

What does the US Copyright Office say?

This is guidance, not law in force. On May 9, 2025, the Copyright Office released a pre-publication version of Part 3 of its AI report.

It says fair use is decided case by case. The first factor (purpose) and fourth factor (market effect) tend to carry the most weight. Some training can be transformative.

But the report warns that "making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, goes beyond established fair use boundaries."

That lines up with what the courts did.

Sources: Copyright Office Part 3 report (PDF)

So what training data is safe for a builder to touch?

Run every dataset through one filter. Call it the provenance test: can you show a lawful path to each item you trained on? Bought, licensed, public domain, or openly licensed passes. Scraped from a pirate library fails, no matter how transformative your model is.

Lawful source. Buy it, license it, or use public-domain and openly licensed data. Keep the receipt.
Do not compete with your source. A tool that replaces the thing you copied is the fact pattern that loses.
Watch the output, not just the input. These rulings covered training. They did not bless what the model spits out. If your model reproduces recognizable work, that is a separate problem.

If you are building on a training set and want a second set of eyes before you ship, that is the kind of thing I do at nomadtechnologist.com.

Not legal, financial, or tax advice.