Copyright & IP

AI copyright governance involves several different acts: acquiring material, retaining a corpus, making training copies, distributing a model, and generating or publishing outputs. A favourable ruling about one act does not clear the others. This chapter distinguishes selected verified decisions from practical risk controls; it is not a complete docket tracker or a licence to use a particular dataset.

Bartz v. Anthropic: separate training from the library

In his 23 June 2025 partial-summary-judgment order, Judge William Alsup held that the training use at issue was fair use. He separately accepted the one-for-one digitisation of purchased print books for Anthropic’s library, while rejecting fair use for downloading and retaining pirated books to build a general-purpose central library. The order did not hold that all training on pirated copies was infringing; nor did the training holding excuse the separate library acquisition. Its conclusion expressly left other uses of library copies outside the training ruling.¹

The governance takeaway is to document each acquisition and use, rather than treating “AI training” as a single permission question. Preserve evidence of source, acquisition method, licence terms, allowed uses and deletion decisions. This is a practical inference from the decision, not a new statutory documentation duty.

On 20 July 2026, Judge Araceli Martínez-Olguín granted final approval of the $1.5 billion class settlement and entered judgment. The order identifies 482,460 works on the Works List and explains the limited release; future misconduct and model-output claims are not released. The often-repeated “about $3,000 per book” figure is not a guaranteed net payment to each author: fees, allocation between rightsholders and the approved distribution process matter.²

A district-court ruling and a class settlement are not a nationwide appellate determination of every AI-training practice. Settlement administration is also distinct from the merits of an unrelated copyright claim.

Kadrey v. Meta: a merits ruling on a limited record

Judge Vince Chhabria’s 25 June 2025 order granted Meta partial summary judgment on its fair-use defence to the training-copying claim. That is a substantive fair-use ruling, not merely a procedural dismissal. The court stressed that these plaintiffs had not developed the market-harm record needed for their claims, especially the potential for market dilution by AI-generated competing works. The result was expressly limited to the record before the court; it was not a blanket approval of Meta’s training practices or a general permission to acquire pirated material.³

For governance teams, market effects, the actual use, the relevant works and the evidence all matter. “Another developer won” is not a substitute for analysis of your own facts.

Getty Images v. Stability AI: a UK judgment, not a global training clearance

The UK High Court issued its judgment on 4 November 2025, rather than merely commencing a trial in 2025. The training claims were no longer pursued on the territorial evidence before the court; the remaining secondary-copyright claim failed, while limited trade-mark claims concerning generated watermarks succeeded. These distinct outcomes do not decide the legality of training worldwide. The judgment is a useful reminder to evaluate outputs and branding separately from training-data rights.⁴

Other litigation concerning text, images, music and copyright-management information continues to require case-specific monitoring. This edition does not assert a current procedural status for every New York Times, Andersen, Concord or other action without a separately checked docket. Complaints are allegations, and interlocutory orders are not final liability judgments.

EU AI Act copyright obligations

For in-scope GPAI providers, Article 53(1)(c) addresses a policy to comply with Union copyright law, including appropriately reserved text-and-data-mining rights under Article 4(3) of Directive (EU) 2019/790. Article 53(1)(d) addresses the sufficiently detailed public training-content summary. The letters are different and should not be reversed. Applicable dates and transitional treatment depend on the model; see EU AI Act.⁵

The GPAI Code of Practice is a voluntary compliance tool. Its Copyright chapter addresses a copyright policy, lawful access, machine-readable reservations, output-related safeguards and a complaints contact. A signature alone does not establish implementation. Neither an ai.txt file nor a commercial opt-out registry is a universal statutory safe harbour; evaluate the actual reservation, relevant technical standard and applicable law.⁶

California training-data transparency

AB 2013 requires covered developers, beginning 1 January 2026, to post specified training-data documentation before making covered generative AI systems or services, or substantial modifications, available to Californians. Coverage concerns releases on or after 1 January 2022 and includes statutory exclusions. The required fields include sources, dataset size and characteristics, copyright/public-domain status, personal information and acquisition details. Publishing a summary does not itself license the data or resolve infringement.⁷

Outputs and human authorship

The US Copyright Office’s Part 2 report was published on 29 January 2025, not March. It explains that copyright protects human-authored expression: using AI as a tool does not disqualify an otherwise protectable work, but purely AI-generated elements are not protected through human prompting alone. Human selection, arrangement or modification can be relevant, depending on the facts. The report addresses copyrightability, which is a different question from whether an output infringes someone else’s rights.⁸

Questions about memorisation, substantial similarity, copyright-management information, contractual rights and cross-border conduct must still be analysed separately. Record which elements people created and edited, and retain the permissions needed for source material and publication.

Practical controls

1. Inventory sources and rights. Record provenance, acquisition, licence scope, restrictions and accountable owner for each corpus.
2. Escalate contested material. Do not treat availability on the web as permission; isolate material whose provenance or rights cannot be established.
3. Separate uses. Review acquisition, retention, training, retrieval, fine-tuning and outputs independently.
4. Implement applicable reservations and disclosures. Preserve evidence of your processes and publish required summaries for covered offerings.
5. Provide a rights contact. Triage complaints, preserve relevant records and define correction or removal procedures.
6. Test outputs. Evaluate memorisation, close copying and unwanted marks; filters reduce risk but do not guarantee non-infringement.
7. Review contracts. Understand downstream licences, indemnity exclusions, permitted use and responsibility for prompts and outputs.
8. Obtain jurisdiction-specific advice before relying on fair use, a statutory exception or a disputed dataset for a material release.

Footnotes

1. Bartz v. Anthropic, Order on Fair Use, Dkt. 231, 23 June 2025 (court order reproduced by Justia). ↩

2. Bartz v. Anthropic, Final Approval and Judgment, Dkt. 680, 20 July 2026; court-approved settlement documents. ↩

3. Kadrey v. Meta, Partial Summary Judgment Order, Dkt. 598, 25 June 2025 (court order reproduced by Justia). ↩

4. UK High Court, Getty Images v. Stability AI, [2025] EWHC 2863 (Ch), 4 November 2025. ↩

5. Regulation (EU) 2024/1689, Article 53. ↩

6. European Commission, GPAI Code of Practice and Copyright chapter. ↩

7. California Legislature, AB 2013, chaptered text. ↩

8. US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, January 2025. ↩