Fair Use and AI Training: What US Courts Have Ruled

No US court has settled whether training AI on copyrighted work is fair use, and the rulings so far point in two directions. Two federal judges in San Francisco held in June 2025 that training a generative model on books was fair use ( Bartz v. Anthropic and Kadrey v. Meta), but the only federal appeals court to rule, the Third Circuit, held in September 2026 that copying Westlaw's headnotes to train a competing legal search tool was not fair use. On one point every court agrees: how the data was acquired is judged on its own, and the pirated library in the Anthropic case ended in a settlement of $1.5 billion plus interest.

This guide is for anyone who has to act on that uncertainty: a company deciding whether to license its data or wait for the courts, a buyer's data team weighing unlicensed sources, counsel, reporters and investors. It explains the four fair use factors in plain words, walks through every significant ruling and where it stands today, covers the US Copyright Office's report on training, and explains why the growing number of licensing deals matters to the fourth factor. It ends with what all this means for a holder deciding to license and for a buyer deciding whether to train on data nobody licensed to it.

This guide is research from public sources, not legal advice.

Contents

1. The four fair use factors in plain words
2. The rulings so far, at a glance
3. Thomson Reuters v. Ross Intelligence
4. Bartz v. Anthropic
5. Kadrey v. Meta
6. The New York Times v. OpenAI and Microsoft
7. The other US cases, including everything filed or decided in 2026
8. Getty Images v. Stability AI (UK)
9. The Copyright Office's report on training
10. Why a licensing market matters to factor four
11. What it means for holders and buyers
12. Questions people ask
13. What this means for you

1. The four fair use factors in plain words

Copyright gives the author of an original work the exclusive right to copy it, adapt it and distribute it. Training an AI model involves copying: the training set is downloaded, stored, cleaned and fed through the model, often many times over. So the question in every training case is not whether copies were made (they almost always were) but whether the copying is excused. The main excuse in US law is fair use, written into the Copyright Act at 17 U.S.C. section 107. A use that qualifies "is not an infringement of copyright" at all, so a defendant who wins on fair use owes nothing.

Fair use is not a rule with a checklist. A court weighs four factors, set out in the statute, case by case, and no single factor decides the outcome. That is why two judges can look at similar facts and come out differently, and why each ruling below matters mainly for its reasoning, not as a rule that binds the next court. Trial court rulings bind only the parties; an appeals court's ruling binds the trial courts in its own circuit; only the Supreme Court binds every court in the country.

Factor one: the purpose and character of the use

The first factor asks why the copy was made and whether the new use has a further purpose or different character from the original. Courts call a use with a genuinely different purpose "transformative". The Supreme Court's 2023 decision in Andy Warhol Foundation v. Goldsmith narrowed that idea: if the original and the copy serve the same or a highly similar purpose and the use is commercial, this factor will likely weigh against fair use, which is how the Third Circuit applied it in the Ross appeal. Commercial use counts against a defendant but does not decide the case.

For AI training, the fight under this factor is whether "learning from" works to build a model is a different purpose from the works' own purpose (to be read, watched or used). The two San Francisco judges said yes for generative models trained on books. The Third Circuit said no for a search tool trained to do the same job as the product it copied from.

Factor two: the nature of the copyrighted work

The second factor asks what kind of work was copied. Creative works (novels, songs, photographs) sit closer to the core of copyright than factual or functional ones (reference material, code, records). Published works get less protection here than unpublished ones. Courts rarely let this factor decide a case. In Ross, it was the one factor that favored the copier, because headnotes summarise the law; it did not change the result.

Factor three: the amount and substantiality of the portion used

The third factor asks how much was copied and whether the copy took "the heart" of the work. Training usually copies whole works, which sounds fatal, but courts measure the amount against the purpose: if the purpose is transformative and needs whole works, copying whole works can be reasonable. In Bartz, the judge found that copying entire books was reasonably necessary for training because the volume of text a model needs is so large. In Ross, copying whole headnotes was too much because ROSS's purpose was not transformative and it could have used the underlying court opinions, which are free.

Factor four: the effect on the market for the work

The fourth factor asks about "the effect of the use upon the potential market for or value of the copyrighted work" ( section 107(4)). Courts have long treated it as among the most important. For training, three different kinds of harm have been argued:

• Direct substitution: the model reproduces the work itself, so people stop buying the original.
• Lost licensing fees: the copier did not pay for a use the rightsholder could have licensed. This is where the deals on our deal ledger come in, and where the courts are split (section 10).
• Market dilution: the model floods the market with new works that compete with the originals without copying them. One judge called this the harm of market dilution and treated it as the argument most likely to win in future cases.

2. The rulings so far, at a glance

There are now 83 lawsuits, rulings and settlements about training AI on data in our case ledger, of which 28 are court rulings and 7 are settlements. Very few of the rulings decide fair use itself. Most are early procedural steps: whether a complaint survives a motion to dismiss, which court hears the case, what the defendant must hand over. The table lists the decisions on the merits of fair use and the cases that will shape them, with their status as of 17 October 2026. Each record links to the court document or filing it is read from.

Two things stand out. First, the only appellate ruling went against the AI developer, but on facts the court itself set apart from generative AI: a tool that could not generate anything new, built to replace the product it copied. Second, the two rulings for AI developers both carved out the way the data was obtained: Anthropic's pirated library was sent to trial (and then settled), and Meta still faces claims that it shared pirated books while downloading them.

3. Thomson Reuters v. Ross Intelligence

What happened

Ross Intelligence was a legal-research start-up that set out to compete with Westlaw, the research service owned by Thomson Reuters. Its tool answered plain-language legal questions by returning passages from court opinions. To train it, ROSS hired a contractor to write training memos, each pairing a legal question with passages that did or did not answer it. The memo-writers used Westlaw's headnotes (the short editorial summaries of legal points that sit above each opinion in Westlaw) to frame the questions. ROSS had first asked to license Westlaw's content and been refused, because it was a competitor. Thomson Reuters sued in 2020.

In February 2025, Judge Stephanos Bibas, a Third Circuit judge sitting as a trial judge in Delaware, granted summary judgment to Thomson Reuters: ROSS had infringed 2,243 headnotes, and its use was not fair. On factor four he wrote that it did not matter whether Thomson Reuters had used its data to train its own tools: the effect on a potential market for AI training data is enough. He stressed that only non-generative AI was before him. He then certified two questions for immediate appeal: whether headnotes are original enough to be copyrighted, and whether ROSS's use was fair.

What the Third Circuit held

On 29 September 2026, the Third Circuit affirmed on both questions, in what the legal press described as the first federal appellate decision to address fair use in AI training. The court's summary is one sentence: " We therefore hold that Thomson Reuters's materials are copyrightable and that ROSS's use was not fair." It went through the factors in order:

• Factor one (against fair use). ROSS's use was commercial, and its purpose was the same as Westlaw's: helping users find relevant court opinions. Training an AI was only "an intermediate step", so the use was minimally transformative, at best. The court rejected comparisons to Google Books and to the software-interoperability cases, because ROSS did not need the headnotes: the opinions themselves were free to copy, and ease is not a justification.
• Factor two (for fair use). Headnotes are published and more factual than creative.
• Factor three (against). ROSS copied each headnote whole, without a transformative purpose that would make that reasonable.
• Factor four (against). The court found harm to Westlaw's value and to its market, and, most importantly for this guide, to the market for licensing data to train AI. In its words: " Here, the evidence shows that the market for licensing headnotes as text to train AI is rapidly developing." By copying without permission, ROSS usurped Thomson Reuters's opportunity to enter that derivative market, and it did not matter that Thomson Reuters had never licensed headnotes to anyone.

How far it reaches

The Third Circuit went out of its way to limit the ruling. In a footnote it distinguished the San Francisco cases and the pending OpenAI litigation: " Unlike the AI models in Bartz and In re: OpenAI, ROSS's AI platform cannot generate original expression". It also noted that the Justice Department had filed a statement backing fair use in the OpenAI case but not in this one. So the ruling binds the federal courts in Delaware, New Jersey, Pennsylvania and the Virgin Islands, and it is persuasive elsewhere, mainly for non-generative tools built to substitute for the source and for its treatment of the training-data licensing market under factor four. ROSS has said it will seek Supreme Court review.

For a holder, Ross is the strongest ruling on record. Its logic fits a company whose data is a curated, editorial product (annotations, summaries, labelled examples, structured reference content) that a competitor could use to build a rival. It also shows that a holder does not need to have sold training licences already: a market that is developing, plus the holder's own use of the data to train its own tools, was enough.

4. Bartz v. Anthropic

Three uses, three answers

Three authors (Andrea Bartz, Charles Graeber and Kirk Wallace Johnson) sued Anthropic in 2024, alleging it copied their books to train its Claude models. The record showed Anthropic had built a central library in two ways: it downloaded books from pirate sites and, later, bought millions of print books, cut off the bindings, scanned them and destroyed the paper copies. Judge William Alsup's order of 23 June 2025 split the case into separate uses and decided each one on its own.

• Training: fair use. He found the training "exceedingly transformative" and wrote that using copyrighted works to train LLMs to generate new text was quintessentially transformative. Like a reader learning to write, the model trained on the books to create something different, not to replace them. The authors did not claim any output reproduced their books.
• Scanning bought books: fair use. Replacing a print copy Anthropic had paid for with a digital copy, and destroying the print original, added no new copies and was fair, for a different reason from training.
• The pirated library: not excused. The judge held that " Such piracy of otherwise available copies is inherently, irredeemably infringing" even if the copies are later used for a transformative purpose. Building a permanent, general-purpose library was itself the use, and it was not fair. He ordered a trial on those copies and the damages.

On factor four, Judge Alsup assumed a market for licensing books to train models could develop, then held that losing it did not count: " Even so, such a market for that use is not one the Copyright Act entitles Authors to exploit." He compared the authors' fear of competing AI-written books to complaining that teaching schoolchildren to write well would create competing works.

The settlement and its amount

The trial never happened. In September 2025 the parties announced a class settlement: Anthropic would pay $1.5 billion plus interest, about $3,000 for each work, across approximately 500,000 works. The court gave preliminary approval that month and final approval on 20 July 2026. A status report filed on 2 September 2026 said that, after fees and expenses, each work had been allotted approximately $2,203.56 from the money already paid in, with Anthropic's final payment of $450 million plus interest still to come.

Put that per-work figure beside the law on damages. Statutory damages for copyright infringement run per work, and for willful infringement can reach up to $150,000 per work. The settlement price of about $3,000 a work is four times the $750 statutory minimum, and it is now the most visible price ever put on copying a work for AI without a licence. It is not a licence fee: it settles past claims for a pirated library and grants nothing for the future.

What Bartz teaches

Bartz is often summarised as "training is fair use". That is half of it. The other half is that acquisition is a separate act, and taking copies from sources you could have bought from or licensed is the clearest loser in the case law so far. The Copyright Office says the same. For a buyer, the provenance of every file in a training set is therefore a legal question, not a paperwork one. Our guide to training data provenance covers how buyers document it.

5. Kadrey v. Meta

What happened

Thirteen authors, most of them well-known novelists, sued Meta for downloading their books from "shadow libraries" (pirate collections such as Library Genesis, known as LibGen) and training its Llama models on them. Two days after Bartz, on 25 June 2025, Judge Vince Chhabria granted Meta summary judgment on the training claim. But the opinion reads almost as a warning to the next defendant.

He opened by saying that, on the general question of whether copying protected works to train models without permission is illegal, "in most cases the answer will likely be yes", and that in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission. Meta won only because of the record these plaintiffs built: the ruling " stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one". It is not a class action, so it only affects the rights of these thirteen authors.

The two losing arguments and the one that could win

The authors argued two kinds of market harm, and the judge called both "clear losers":

1. Regurgitation. Llama could not reproduce enough of their books to matter.
2. Lost licensing fees. Whether a market for licensing books as training data exists or is developing was, in his words, irrelevant, because this market is not one that the plaintiffs are legally entitled to monopolize. Counting the loss of a fee for a transformative use would make factor four circular: every unpaid use would lose a licence fee.

The argument he thought could win was the one the authors barely made: market dilution. A model trained on books can produce endless competing books in a fraction of the time, which could hurt sales of human-written works even if no output copies any of them. He wrote that " Harm from this form of competition is the harm of market dilution", that it is still harm by substitution, only indirect, and that it would hit some works far harder than others: images, news and functional nonfiction (his example was books on caring for a garden), and lesser-known authors, more than famous ones whose readers seek them by name. The authors offered no evidence of it, so the factor could only favor Meta.

He also rejected the claim that requiring licences would stop AI development: developers would simply pay for licences, and if they chose public-domain works instead, that would show they did not need the copyrighted ones as badly as they said.

Where it stands

The training ruling is final for these plaintiffs, but the case is not over. The authors kept a claim that Meta distributed the books while torrenting them, and in March 2026 the court let them add a claim that Meta, while downloading from shadow libraries by BitTorrent, simultaneously uploaded those works onto the torrenting network. Like Bartz, the lasting exposure lies in how the copies were obtained, not in the training itself.

6. The New York Times v. OpenAI and Microsoft

Why this case is different

The New York Times sued OpenAI and Microsoft in December 2023, seeking billions of dollars in statutory and actual damages. It is the case most likely to produce the next big fair use ruling, and it differs from the book cases in two ways that matter under factor four. First, the Times alleges that ChatGPT can reproduce its articles closely and that its products answer questions with the Times's reporting, so readers need not visit the Times: that is direct substitution, the harm the book plaintiffs could not show. Second, news is a market where AI licensing is already common, so the market evidence is thick.

The procedural road so far

In March 2025 the court dismissed the Times's unfair competition claim and its claims under the Digital Millennium Copyright Act (with leave to replead those) but let the core copyright claims go forward ( ruling). In April 2025 the Times's case and others against OpenAI were consolidated before one judge in Manhattan as In re OpenAI Copyright Infringement Litigation, a multidistrict litigation (one court handling pretrial work for many related suits). In October 2025 the court held that claims based on outputs survive where plaintiffs allege outputs a jury could find substantially similar to their works. In January 2026 OpenAI was ordered to produce 20 million output logs, the record of what users actually received.

On 1 September 2026, the US Justice Department filed a statement of interest asking the court to hold that using copyrighted written works to train large language models is fair use. A statement of interest is the government's view, not a ruling, and the market log notes it bears on public content, not on contract and privacy limits on private company data. Days later, the news plaintiffs, OpenAI and Microsoft filed cross-motions for summary judgment. No trial date is set. The fair use question in this case is now before the judge.

What to watch

The ruling will be the first fair use decision on generative AI trained on news, with output evidence and a mature licensing market in the record. The Times itself licenses: it signed its first AI deal with Amazon in 2025, reported at $20 million to $25 million a year, for Amazon's AI products and for training Amazon's models. Other publishers licensed to OpenAI while the Times sued it, among them the Associated Press and News Corp (a deal the Wall Street Journal said "could be" worth close to $250 million over five years). If the court follows the Third Circuit and the Copyright Office in counting that market, the Times's factor-four case is strong; if it follows Bartz and Kadrey, the case turns on outputs and dilution. Our guide to news publisher AI licensing deals lists the deals on both sides.

7. The other US cases, including everything filed or decided in 2026

The cases above are the ones with rulings. Behind them is a long line of pending suits; a federal judge in New York counted more than fifty lawsuits challenging AI training in late 2025, and more were filed through 2026. Most are far from a fair use ruling, but they show where the next fights are: music lyrics, news, reference works, and the acquisition of data rather than training itself.

The table lists every case event in our ledger dated 2026, newest first. Where the record is a press report rather than a court document, it says so. Non-US cases are marked.

Three patterns run through the 2026 docket. Piracy is the plaintiffs' preferred ground. After Bartz, the music publishers' suits against Anthropic and the publishers' class action against Google all lead with how the works were obtained, because that is where Bartz and Kadrey left defendants exposed. Scraping cases are increasingly about access, not copying: Reddit's suits against Anthropic and Perplexity rest on contract and on getting around technical barriers, not on fair use (our guide to whether it is legal to scrape data for AI covers those theories). And suits end in licences: Warner Music's suit against Suno settled into a licensing partnership, with Suno promising licensed models in 2026, and Universal's suit against Udio settled with a licensed platform deal. The music sector shows the path many of these cases may take.

8. Getty Images v. Stability AI (UK)

Getty Images sued Stability AI in London over the training of Stable Diffusion, an image model, on Getty's photographs. The UK has no fair use doctrine; it has narrower, listed exceptions ("fair dealing"), so the case is no guide to section 107. It is worth knowing for one reason: it shows how much turns on where training happens.

Before judgment, Getty dropped its main training claim because there was no evidence the training took place in the United Kingdom, and copyright is territorial. What was left was a claim that the finished model, imported into the UK, was itself an "infringing copy". On 4 November 2025, the High Court dismissed that claim, holding that the model weights are not themselves an infringing copy and they do not store an infringing copy. Getty won only a narrow trade mark point about Getty watermarks appearing in images made by early versions of the model. Getty has been granted permission to appeal, with the appeal expected in November 2026, and it is pursuing a separate US case in California, filed in August 2025 and now in fact discovery.

The lesson for a US reader is narrow: a judgment abroad does not decide fair use at home, and a developer that trains in one country and serves models in another may face different law in each.

9. The Copyright Office's report on training

In May 2025 the US Copyright Office released a pre-publication version of Part 3 of its report on copyright and AI, on generative AI training ( record). The Office administers copyright registration and advises Congress; its report does not bind courts, and the final version has not been published. But it is the most thorough government analysis of the question.

The report declines to say that training is or is not fair use in general. It describes a spectrum. At one end, noncommercial research or analysis that does not let models reproduce the works is likely fair. At the other: " the copying of expressive works from pirate sources in order to generate unrestricted content that competes in the marketplace, when licensing is reasonably available, is unlikely to qualify as fair use." Most uses fall in between, and the Office expects the first and fourth factors to carry the most weight.

On factor four, the Office took the opposite view from Bartz and Kadrey on licensing markets. It acknowledged the circularity argument (if a use is fair, no licensing market for it can form), but concluded: " Where licensing markets are available to meet AI training needs, unlicensed uses will be disfavored under the fourth factor." Where licensing is impossible because rights are too scattered, there is no functioning market to harm and the factor may favor fair use. The report also identified who is best placed to license: " Direct licensing is most common and most promising with respect to corporate entities with catalogs of high-quality and easily identifiable content." A company that holds its own records, with clear ownership and a single counterparty, is exactly that kind of entity.

10. Why a licensing market matters to factor four

The split, stated plainly

Factor four asks about the effect on the "potential market for or value of" the work. A licensing market for training data is a potential market. The question is whether losing the fee for training counts as harm. Two lines of reasoning now exist:

• It counts (Third Circuit in Ross, the Delaware trial court, the Copyright Office). A market for licensing data to train AI is a derivative market rightsholders "would in general develop or license others to develop", and unlicensed copying takes it from them. The Third Circuit found that market "rapidly developing" and counted it even though Thomson Reuters had never licensed to anyone.
• It does not count, if the use is transformative (Bartz, Kadrey). Counting a lost fee for a transformative use would make factor four favor the rightsholder in every case. Under this view the licensing market matters only if the training use is not transformative, or if the model's outputs substitute for the works.

The two lines can be reconciled on their facts: ROSS's tool was a substitute, the Claude and Llama models were found not to be. But the reasoning differs, and the courts that will decide the news and music cases must choose. Even on the Bartz view, a licensing market is evidence for market dilution and for substitution: deals show what the works are worth to AI products and how directly those products compete with them.

The evidence that a market exists

Whichever line a court follows, the factual record of a licensing market has grown fast. Our deal ledger holds 111 disclosed licensing deals between data holders and AI companies since July 2023, of which 56 are for training models (the rest license content to be shown in answers). They involve 44 different licensees, and OpenAI alone has signed 29. Only 19 state a dollar amount, but those that do are large:

Across books and archives, the public prices on record put a median of $2,750 per book or title on a licensed book or title. That is the number a court, a buyer and a holder can all point to when asking what a work is worth as training data. The Reddit deals are taken apart in our guide to Reddit's AI data licensing deals. Our price board and the guide to AI training data prices show the full spread.

What this means in practice

For the courts, these deals are the evidence a plaintiff needs under the Ross and Copyright Office reading, and the benchmark for damages under any reading. For the market, they are the alternative that makes the fair use fight less important: a buyer that licenses does not need to win on fair use. That is one reason several suits have already ended in licences (section 7), and why the rights questions in our guide to data licensing agreements now matter more to most holders than the case law itself.

11. What it means for holders and buyers

For a holder deciding whether to license

Most of the fair use debate concerns published works: books, articles, songs, photographs that sit where a crawler or a pirate library can reach them. Most company data is not like that. Support tickets, internal documents, code, call recordings and transaction histories sit behind a login, and a buyer cannot obtain them without the holder. For that data, fair use is rarely the question. The questions are contract (customer agreements, software vendors' terms), privacy law, and whether the holder has the rights to license at all. Our guide on selling company data to AI labs covers those.

Where fair use does matter to a holder (a publisher, a forum, a research firm, anyone whose content is public or has leaked), the rulings point to a few practical conclusions:

• Waiting for the courts is not a strategy. Rulings that would settle the question nationally may be years away, and the ones so far split. Meanwhile buyers are licensing, and a licence signed now is evidence of the market in any later case.
• Your strongest case is the Ross case. If your data is curated, editorial or labelled (summaries, annotations, classifications, structured reference content) and a buyer could use it to build a product that competes with yours, the Third Circuit's reasoning fits you closely. Which kinds of data buyers ask for, and why, is mapped on what labs buy.
• Being licensable is itself protection. The Copyright Office ties factor four to whether licensing is "reasonably available". A holder with clear ownership, documented rights and a way to deliver data makes licensing available, and makes unlicensed use harder to defend. The readiness check and a datasheet are where that starts.
• Crawler blocks are evidence, not walls. Shutting AI crawlers out in robots.txt does not change the fair use analysis much, but it shows the use was unwanted and it pushes buyers toward a licence. Our AI crawler census tracks how many sites do it.

For a buyer deciding whether to train on unlicensed data

A buyer that trains on data nobody licensed to it is betting on three things at once: that the training use will be found transformative, that the outputs will not substitute for the works, and that the data was lawfully acquired. The case law so far gives a fair chance on the first for generative models, a mixed picture on the second, and no protection at all on the third. Bartz held piracy "irredeemably infringing" and priced it at about $3,000 a work; statutory damages for willful infringement run to $150,000 a work. The 2026 suits against Anthropic and Google lead with acquisition for that reason.

The risk is also one of disclosure. California requires developers of generative AI systems offered to Californians to post a summary of the datasets they trained on, including the sources or owners of the datasets and whether they were purchased or licensed. A developer that cannot say where its data came from now has to say so in public. Licensed data costs more than scraped data, but it is the only kind a developer can document end to end.

Worked example: a research publisher and a model builder

This is an illustrative example, not a real company. A 40-person research firm publishes analyst reports behind a paywall and has done so for a decade. Its reports are expressive, published, and sold by subscription. A start-up building an AI analyst tool for the same customers is choosing between scraping leaked copies of the reports and licensing them.

If the start-up scrapes. On factor one, its tool answers the same questions the reports answer, for the same buyers, so a court following Ross would likely find the purpose the same and the use minimally transformative. Factor two leans slightly its way (reports are partly factual). Factor three is against it (whole reports copied). On factor four, the firm can show direct substitution (customers who use the tool may cancel), and, under the Ross and Copyright Office reading, the loss of a licensing market that the deal ledger shows exists. The copies came from leaks, so the acquisition itself is exposed, as in Bartz. The start-up's best hope is a court that follows Bartz on transformation and finds its outputs do not substitute, which is a weak bet when it sells to the same customers.

If the firm licenses. The firm can license a defined set of reports for training, for a fixed term, with limits on output that reproduces its text, and with the start-up warranting it will not use leaked copies. The start-up gets data it can document and a defence it never has to use. The firm gets revenue, a market price on record, and evidence of a licensing market in any later dispute with someone who did not pay. The terms that decide what such a licence is worth are in our licence terms register.

Questions people ask

Is training AI on copyrighted material fair use? Sometimes, and no court has settled it nationally. Two federal trial courts held in 2025 that training generative models on books was fair use ( Bartz, Kadrey), and the Third Circuit held in 2026 that training a competing non-generative tool was not. The answer depends on the four factors and the facts of each use.

Did Anthropic win or lose its copyright case? Both. The court held that training Claude on books was fair use, but that Anthropic's library of pirated books was not excused and ordered a trial on it ( order). Anthropic then settled the class claims for $1.5 billion plus interest.

How much did authors get from the Anthropic settlement? The settlement was announced at about $3,000 per work. A September 2026 court filing said each work had been allotted approximately $2,203.56 after fees from the money paid in so far, with a final payment still due.

What did the court decide in Thomson Reuters v. Ross? The trial court in 2025 and the Third Circuit in September 2026 both held that copying Westlaw headnotes to train ROSS's legal search tool was not fair use ( appeal ruling). The appeals court stressed that ROSS's tool could not generate new text and was built to replace Westlaw. ROSS has said it will seek Supreme Court review.

Has the New York Times case against OpenAI been decided? No. Both sides filed cross-motions for summary judgment in September 2026 and no trial date is set. The Justice Department has filed a statement backing OpenAI's fair use position, which the court is free to disregard.

Does fair use apply outside the United States? No. Fair use is US law. The UK and EU have narrower, listed exceptions, and the UK's Getty v. Stability judgment turned on where training happened. A German court ruled against an AI music developer in GEMA v. Suno.

Do AI companies need a licence to train on my data? Under current US case law, not always for published works they acquired lawfully, but always for copies they could only get by piracy or by breaking a contract or an access control. For private company data, which a buyer cannot reach without you, the practical answer is yes: the buyer needs your licence, and your terms set what it may do.

Can I stop AI companies from training on my content? You can shut AI crawlers out in robots.txt and forbid scraping in your terms of service, but neither is a copyright rule, and a robots.txt line has been held not to be an access control under anti-circumvention law. The strongest position is to make your content licensable on your terms, so unlicensed use competes with an available market. See our guide on scraping for AI.

What this means for you

If you hold data, the fair use cases are a reason to license deliberately, not a reason to wait. Start by working out whether your data is the kind fair use even reaches: published work a crawler could take, or private records a buyer can only get from you. Then settle what you own and what you may license, using the readiness check, and see what comparable data has fetched on the price board and in the deal ledger. The valuation gives a first estimate with no data upload. When an offer arrives, the offer check compares its terms with what buyers have published, and the licence terms register shows where holders stand on each clause.

If you buy data, the safest reading of the case law is that training may be fair, acquisition must be clean. Document where every file came from, prefer licensed sources for anything that competes with its source, and expect your own customers to ask. The buyer board and the vendors register show who sells licensed data and on what terms.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. We follow these cases because they set the price of the alternative to a licence. The full record is on the cases page, the rights hub and in our other guides, including how AI companies get training data.

This guide is research from public sources, not legal advice.