The authors’ class action against Anthropic in the United States (Bartz v. Anthropic) has ended in the largest copyright settlement of the AI era. On 20 July 2026 the US District Court for the Northern District of California granted final approval to a USD 1.5 billion settlement covering 482,460 books, about USD 3,000 per work before fees. First payments are expected by 15 November 2026. The case is often summed up in one line: “the court allowed AI training and punished piracy.” The real picture is more detailed, and under Turkish law the exposure is potentially greater. This note covers what the court decided and did not decide, the settlement figures, the latest rulings in the US and Europe, and the lessons for companies that build or use models in Türkiye.
How the case started: the 23 June 2025 ruling
Authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson sued in August 2024. The pivotal moment was Judge William Alsup’s partial summary judgment of 23 June 2025. The court identified three different uses of the books and assessed each separately (summary of the ruling):
| Use | Court’s assessment |
|---|---|
| Training large language models on books | Fair use. The court found the use “exceedingly transformative”: the model learned from the works to generate new text, not to copy and distribute them. |
| Scanning purchased print books into digital copies | Fair use. The format change of a purchased copy made storage and search easier without creating additional copies. |
| Downloading more than seven million books from pirate sites (Library Genesis, Pirate Library Mirror, Books3) to build a permanent central library | Not fair use. There is no exception that excuses avoiding payment for works that could be lawfully acquired. This part was sent to trial. |
In July 2025 the court certified a class of rightsholders of books downloaded from LibGen and PiLiMi. The class was limited to works with an ISBN or ASIN that had been timely registered with the US Copyright Office (class definition). That is the main reason why only 482,460 of more than seven million books entered the settlement: under US law, statutory damages are available only for timely registered works. Books3 was left out of the class for lack of metadata.
Did the court rule piracy unlawful?
Not conclusively. The court held that downloading and keeping pirated copies could not be protected by the fair use defence, but the scope of infringement and the amount of damages were to be tried before a jury from 1 December 2025. The settlement came before that trial, so there is no final judgment on infringement or damages for the pirated library. One point the court stressed matters: buying a book after first downloading it from a pirate site does not remove liability for the pirated copy, although it may affect the amount of statutory damages.
The economics of the settlement should be read from there. US copyright law provides statutory damages of USD 750 to 30,000 per work, and up to USD 150,000 for wilful infringement. About USD 3,000 per work is four times the ordinary minimum, and far below the wilful maximum. The settlement amount therefore reflects the litigation risk attached to how the books were acquired, in a case where the court had already treated training as fair use, rather than the theoretical statutory ceiling across hundreds of thousands of works. The settlement was reached without any finding of liability on the pirated library.
The settlement in figures
| Item | Figure |
|---|---|
| Settlement amount | USD 1.5 billion plus interest; non-reversionary |
| Works covered | 482,460 books (the Works List) |
| Gross per work | About USD 3,000 |
| Claims | 440,490 works (about 91%) at final approval; 449,731 works (93.2%) per the 2 September 2026 status report. 350 rightsholders opted out with 1,802 works |
| Attorneys’ fees | 12.5% (USD 187.5 million) requested; the court approved about 6.8% (USD 101,561,111) |
| Objections | 54 objections and comments, all overruled |
| First distribution | By 15 November 2026; about USD 2,203.56 per work (the final USD 450 million instalment is not yet in the fund) |
| Author and publisher shares | 50/50 by default unless otherwise agreed; disputes go to a Special Master |
Sources: Authors Guild, summary of the final approval order, summary of the 2 September 2026 status report.
What the settlement does not cover matters as much as the figures. The release covers only past claims about acquiring and copying Works List titles up to 25 August 2025. Claims about model outputs, future conduct and works not on the list are outside the release. Anthropic undertook to destroy the original files downloaded from LibGen and PiLiMi and copies made from them, and represented that neither dataset was in the training data of its commercially released models. Non-monetary relief sought by objectors, such as model deletion, attribution in outputs or a ban on training with scanned books, was not granted.
A procedural note: Judge Alsup took inactive status at the end of 2025, and the case was reassigned to Judge Araceli Martínez-Olguín, who granted final approval.
The US debate is not over: Kadrey v. Meta and Thomson Reuters v. Ross
Bartz is a district court ruling, not binding precedent. Two days later, on 25 June 2025, Judge Vince Chhabria of the same court ruled for Meta in Kadrey v. Meta, but on different reasoning: the authors had failed to prove market harm, and he stressed that the argument that a language model can produce competing works at a speed and scale no human can match (“market dilution”) could change the result in another case. Claims about whether Meta shared books via torrent are still pending; the summary judgment hearing has been set for February 2027.
The first appellate ruling came on 29 September 2026. In Thomson Reuters v. Ross, the US Court of Appeals for the Third Circuit held that using Westlaw headnotes to train a competing legal research tool was not fair use (summary). The court stressed two points: the use served the same purpose as the original product, in a tool that does not generate new expression; and there is a “rapidly developing” market for licensing AI training data, from which the unlicensed use took revenue. The court expressly left the generative AI question open. The upshot: in the US today, “training is fair use” is not a general rule but a fact-dependent defence.
No fair use in Europe: the TDM exception, the Munich rulings and the AI Act
EU law has no general, flexible defence like fair use; exceptions are limited in number. For model training the key one is the text and data mining (TDM) exception in Articles 3 and 4 of the Digital Single Market Directive: commercial mining is permitted where there is lawful access to the source and the rightsholder has not reserved its rights in a machine-readable way.
The Munich Regional Court has drawn the limits of that exception in two rulings. In GEMA v. OpenAI (11 November 2025) it treated the memorisation of song lyrics in model weights as reproduction and held that the TDM exception may cover the copies necessary to assemble a dataset, but not the retention of the work inside the model. In GEMA v. Suno (31 July 2026) the same chamber held that six compositions were memorised in model versions stored on servers in Germany and that this infringed; applying US law to the training copies made in the US, it also rejected fair use, distinguishing Bartz and Kadrey because the outputs were substantially similar to the originals. Neither ruling is final.
The EU AI Act also enters the picture. Providers of general-purpose AI models placing models on the EU market must adopt a copyright policy, honour machine-readable rights reservations and publish a summary of training content (Article 53(1)(c) and (d)); under Recital 106 this applies regardless of where training took place. Models placed on the market before 2 August 2025 have until 2 August 2027 to comply. The copyright chapter of the Code of Practice goes a step further: signatories commit to excluding from crawling websites that EU or EEA courts or authorities have found to infringe copyright persistently and on a commercial scale (summary). In Europe too, a pirate library is not an acceptable “data source”.
The same facts under Turkish law: no fair use, and criminal exposure
Türkiye is far less flexible than the US. The Law on Intellectual and Artistic Works No. 5846 (FSEK) recognises no general fair use defence; its limitations are listed one by one, and there is no TDM exception of the EU kind (see our note on training data and FSEK). Applying the three uses in Bartz to FSEK gives the following picture:
- Training: FSEK Article 22 defines reproduction to cover any copy, “temporary or permanent”, “direct or indirect”. With no exception covering copies made for training, training that one US district court treated as fair use would, as a rule, require the rightsholder’s permission in Türkiye. The private-use exception (Article 38) applies only to non-commercial personal use and does not cover a company training a model.
- Scanning purchased books: the format change treated as fair use in the US is reproduction under FSEK; buying a copy is not permission to reproduce it.
- Pirate library: this is where the consequence is harshest. FSEK Article 71(1) provides for one to five years’ imprisonment or a judicial fine for anyone who “possesses or stores, other than for personal use” unlawfully reproduced works. A company keeping books downloaded from pirate sites on its servers for training may fall within this offence. Prosecution depends on a complaint by the rightsholder (FSEK Article 75), and because companies cannot themselves be criminally liable under Turkish law, exposure falls on the managers and employees responsible. The same paragraph also criminalises unauthorised reproduction, so scanning and training copies could in principle carry criminal exposure too.
On civil damages, the rightsholder may claim up to three times the fee it could have asked under a licence, or three times the market fee (FSEK Article 68). This “up to three times” rule, which, like US statutory damages, can exceed the actual loss, becomes a calculable risk as a licensing market forms: the market for AI training data licences recognised by the Third Circuit could support the market fee before a Turkish court. There is no published Turkish appellate ruling on model training yet; but the wording of the statute suggests that the distinction drawn in Bartz would be difficult to rely on in Türkiye.
What Turkish AI companies and model users should do
Model developers: record the provenance of training data from today: which dataset, which licence, when, from where. Keep no files from shadow libraries (LibGen, Z-Library, Anna’s Archive and the like) on your servers; under FSEK Article 71, not using them for training does not make storing them innocent. If you scrape the web, log robots.txt and machine-readable rights reservations; if you sell into the EU, the AI Act’s copyright policy and training-content summary duties bind you too.
Users of third-party models: read the scope, exclusions and cap of the copyright indemnity (IP indemnity) in your contract with the model provider. The Bartz settlement preserved claims about model outputs; output-related disputes may surface in your product.
Investors and acquirers: in due diligence on AI companies, the data source list, licences, deletion records and provider contracts should be a separate workstream. Bartz showed that how training data was acquired can turn into a liability measured in billions.
Authors and publishers: the Bartz claims deadline passed on 30 March 2026, and the class covered only works timely registered with the US Copyright Office. Rights in works not on the list are preserved, and claims under FSEK in Türkiye are unaffected by this settlement. As the licensing market grows, scope, term, output restrictions and audit rights will be the negotiating points in training licences. Our intellectual property team advises on designing these contracts and on data provenance reviews.
Frequently asked questions
Did the court rule that AI training is lawful?
In Bartz, the district court held that using books to train its models was fair use under US law, but that downloading pirated copies to build a permanent central library was not. That ruling is not binding precedent. On 29 September 2026 the Third Circuit found no fair use on different facts, for a non-generative tool, expressly distinguishing generative models; that ruling may also be taken further.
Was Anthropic found liable for the pirated books?
No. The court held that a library built from pirated copies could not be protected by fair use, but the parties settled before the trial on infringement and damages. The settlement received final approval on 20 July 2026.
Who receives the USD 3,000 per book?
The rightsholders of the 482,460 works on the Works List. USD 3,000 is the estimate before fees and costs; the first distribution, by 15 November 2026, is expected to be about USD 2,203.56 per work. Where an author and a publisher both claimed the same work, the amount is split 50/50 by default.
Can a company in Türkiye train a model on books?
FSEK has no general fair use defence and no text and data mining exception. Reproduction for training requires the rightsholder’s permission as a rule, and storing pirated copies other than for personal use may carry one to five years’ imprisonment or a judicial fine under FSEK Article 71.
Does the settlement cover model outputs?
No. The release covers only past claims about acquiring and copying the works up to 25 August 2025. Claims about outputs, future conduct and works not on the list are preserved.
Sources. Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.): partial summary judgment of 23 June 2025 (summary), class certification (summary), final approval order of 20 July 2026 (Authors Guild; Litigation Sentinel), status report of 2 September 2026 (summary); Thomson Reuters v. Ross Intelligence, US Court of Appeals for the Third Circuit, 29 September 2026 (summary); LG München I, 42 O 14139/24, 11.11.2025 (LTO) and 42 O 763/25, 31.7.2026 (LTO); Regulation (EU) 2024/1689 (AI Act), Article 53, and the GPAI Code of Practice (summary); FSEK No. 5846, Articles 22, 38, 68 and 71 (in Turkish).
Author
More from Vircon Insights
Türkiye Allows Electronic Informed Consent via e-Nabız and SMS OTP: What the New Regulation Means for Hospitals and Health-Tech
October 27, 2026Türkiye Raises the Qualified Investor Threshold to TRY 50 Million and Ends the Carve-Out for VC and Real Estate Funds
October 26, 2026Who Owns the Sword? Virtual Items, User-Generated Content and Mods Under Turkish Law
October 24, 2026Anatomy of a Game Publishing Agreement Under Turkish Law: Revenue Share, Recoupment, Minimum Guarantee and Who Keeps the IP
October 19, 2026Technopark and R&D Centre Incentives for a Turkish Startup: Law 4691, Law 5746 and the Three Per Cent Venture Obligation
October 13, 2026Türkiye Bans Social Media for Under-15s from 1 November 2026: What Platforms Must Change (Law 7578)
October 12, 2026Related Practice Areas
Intellectual Property
Trademark, patent, copyright and trade secret advisory.
View service →Gaming & Esports Law
Studio funding rounds, chain of title, publishing and platform agreements, esports.
View service →ICOs, Crypto & Blockchain
Crypto-asset regulation, token offerings, exchange and custody licensing.
View service →