Copyright and Intellectual Property

Copyright and intellectual property (IP) in AI concerns two sets of legal questions: whether copying protected works to train AI models, or models reproducing such works in outputs, infringes the rights of the owners; and whether AI-generated material, or AI-assisted inventions, can themselves be protected by copyright or patent. The answers are set by national law and are being decided case by case.

Generative models (see [CAP-GENAI-001]) are trained on very large collections of text, images, code, music and video, much of it protected. Rights holders have sued developers in several countries. The disputes turn on different doctrines in each system: the open-ended, four-factor fair use defence in US law; closed lists of exceptions (for example text-and-data-mining exceptions) in EU member states and the UK; and, in the UK, whether a trained model is an "infringing copy". Because the questions are new and first-instance courts have not agreed, the law is unsettled. A second set of questions concerns authorship: US law, as applied by the courts so far, requires a human author.

Privacy aspects of training data are covered in [SAFE-PRIV-001]; dataset provenance in [TECH-DATA-001]; creative-industry effects in [APP-MEDIA-001].

Before generative AI, copyright disputes over machine copying concerned search engines and digitisation, decided largely under US fair use. An early US ruling on fair use for AI training data was Thomson Reuters v. Ross Intelligence, decided on 11 February 2025 by Judge Bibas (sitting by designation in the District of Delaware) in favour of Thomson Reuters, in a case involving a non-generative legal-search tool; the Third Circuit affirmed on 29-30 September 2026, reported as the first federal appellate decision on fair use in AI training. In June 2025 two judges in the Northern District of California reached more favourable conclusions for developers on training itself (Bartz v. Anthropic, 23 June 2025; Kadrey v. Meta, 25 June 2025). In 2025 outside the US, the UK High Court decided Getty Images v Stability AI (4 November 2025), a Munich court decided GEMA v OpenAI (11 November 2025) and the Hamburg appeal court decided Kneschke v LAION (10 December 2025). Section 6 summarises each.

#4.1 Where copying occurs

Copyright analysis distinguishes (a) copying in building the training set (downloading and storing works); (b) training, in which model parameters are adjusted; (c) memorisation inside the model, where parameters may encode portions of works; and (d) outputs that may reproduce or resemble protected works. Technical evidence of extractable memorisation (Carlini et al., 2021; Nasr et al., 2023; see [SAFE-SEC-001] and [SAFE-PRIV-001]) bears on (c) and (d); it does not by itself determine the legal characterisation.

#4.2 US fair use

17 U.S.C. section 107 lists four factors for fair use: the purpose and character of the use, including whether it is commercial; the nature of the copyrighted work; the amount and substantiality of the portion used; and the effect of the use on the potential market for or value of the work (U.S. Copyright Office, Title 17). Courts weigh the factors case by case. In AI cases, the debate centres on whether training is "transformative" and on market harm, including a "market dilution" theory in which AI outputs compete with the authors' works without copying them.

#4.3 EU text-and-data-mining exception and AI Act duties

The EU AI Act requires providers of general-purpose AI models to "put in place a policy to comply with Union law on copyright and related rights", including honouring rights holders' text-and-data-mining opt-outs under Article 4(3) of the 2019 copyright directive, and to "draw up and make publicly available a sufficiently detailed summary about the content used for training" using an AI Office template (Regulation (EU) 2024/1689, Art. 53(1)(c)-(d); quoted wording checked against the Official Journal text, read 2026-10-18). Application dates of the AI Act were amended in 2026; see [GOV-EU-001].

Each item gives jurisdiction, court, date and status as of 2026-10-18. Items marked "secondary" rely on law-firm or press summaries; items marked "primary" cite a court document that was read.

United States

• Thomson Reuters v. Ross Intelligence (D. Del., No. 1:20-cv-613, Judge Bibas, 11 February 2025): partial summary judgment for Thomson Reuters on direct infringement of 2,243 Westlaw headnotes and rejection of the fair use defence; the tool was not generative AI (secondary). On interlocutory appeal (3d Cir. No. 25-2153; argument reported for 11 June 2026) the Third Circuit panel (Montgomery-Reeves, Restrepo, Bove, JJ.) affirmed. The opinion carries a docket stamp of 30 September 2026 (primary, docket stamp only); press reports date a first, sealed version to 29 September 2026. Per press reports of the opinion, it held the headnotes copyrightable, held ROSS's use "minimally transformative at best" and not fair use, found market harm including to a market for licensing headnotes as AI training data, and distinguished ROSS's non-generative system from LLMs (secondary: Courthouse News, LawNext). ROSS has said it will seek en banc rehearing and Supreme Court review (secondary); the case returns to the district court on remaining claims. The opinion PDF on the court's site could not be read in this review, so these details rest on the press reports named above; a law-firm alert cites it as a precedential opinion (secondary), and I found no ruling on a rehearing request.
• Bartz v. Anthropic PBC (N.D. Cal., No. C 24-05417 WHA, Judge Alsup, Order on Fair Use, 23 June 2025; primary, order read): the court held that using the books to train Claude was "exceedingly transformative" and a fair use; that Anthropic's digitising of print books it had purchased was also a fair use (for a different reason, format shifting of lawfully owned copies); and it denied summary judgment on the central library built from pirated copies, ordering a trial on that library and damages. The class action then settled for USD 1.5 billion (plus interest) and the court (Judge Araceli Martinez-Olguin, successor to Judge Alsup) gave final approval on 20 July 2026, with a fee award of about USD 101.6 million (Authors Guild report; Jurist; secondary). Reports give roughly 480,000 works; counts vary. Because it settled, it sets no binding precedent. The final approval order itself was not read; the approval date and fee figure are as reported by those sources, which differ on minor details.
• Kadrey v. Meta Platforms (N.D. Cal., No. 23-cv-03417-VC, Judge Chhabria, order of 25 June 2025; primary, order read): summary judgment for Meta on fair use for these thirteen authors on the record before the court. The court said plaintiffs had presented no meaningful evidence of market dilution, and wrote that 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", not that Meta's use of copyrighted material to train its models is lawful. A separate claim about distribution during torrenting continued; a summary judgment hearing on it was reported as set for 25 February 2027 (secondary).
• New York Times v. Microsoft and OpenAI (S.D.N.Y., No. 1:23-cv-11195, Judge Stein; coordinated with other cases in MDL No. 25-md-03143, In re OpenAI, Inc. Copyright Infringement Litigation): no ruling on fair use reported. Cross-motions for summary judgment were filed on 4 September 2026 (secondary: litigation trackers and press); no trial date is set. The US Department of Justice has filed a statement of interest supporting OpenAI's fair use position (secondary). The dockets were not opened, and secondary sources give conflicting dates for opposition and reply briefs, so no briefing schedule is stated here.
• Thaler v. Perlmutter (D.C. Cir. No. 23-5233, 18 March 2025): the Copyright Act requires human authorship; rehearing was denied on 12 May 2025 and the Supreme Court denied certiorari on 2 March 2026 (No. 25-449) (primary: Supreme Court docket page). Certiorari denial is not a merits ruling. The U.S. Copyright Office report Copyright and Artificial Intelligence has three parts: Digital Replicas (31 July 2024), Copyrightability (29 January 2025) and Generative AI Training (pre-publication version of 9 May 2025; the Office's index page, read 2026-10-18, still describes a final version as to come) (U.S. Copyright Office, 2025).
• Thaler v. Vidal (Fed. Cir., No. 2021-2347, 5 August 2022; 43 F.4th 1207): inventors under the Patent Act must be natural persons (secondary: case reporter listing and commentary).

United Kingdom

• Getty Images v Stability AI [2025] EWHC 2863 (Ch) (High Court, Mrs Justice Joanna Smith, 4 November 2025; primary, judgment text read via the National Archives). Getty's training claim was abandoned because there was no evidence that training took place in the UK, and its outputs claim was abandoned because the relevant prompts were blocked (para 9). The court dismissed the secondary-infringement claim: an "article" may be intangible, but an AI model that does not store or reproduce any copyright works is not an "infringing copy" under ss. 22-23 CDPA, and Stable Diffusion is such a model (para 758(viii)). Getty succeeded in part on trade mark infringement (iStock watermarks under s. 10(1) and 10(2) for v1.x; Getty Images watermarks under s. 10(2) for v2.x), failed under s. 10(3), and the court declined to decide passing off (para 758). At a consequentials hearing in December 2025 Getty was granted permission to appeal on the secondary-infringement point (sources give 16 or 17 December 2025; secondary: CMS, Burges Salmon). I found no listed Court of Appeal hearing date or decision. One law-firm report says Stability was refused permission to appeal the trade mark findings by the High Court and would need to seek it from the Court of Appeal; I did not find whether it did (secondary).
• Thaler v Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49 (UK Supreme Court, 20 December 2023): an inventor under the Patents Act 1977 must be a natural person (neutral citation and date: Supreme Court case page; holding: secondary summary).
• Government policy. On 18 March 2026 the UK Government published its report on copyright and AI under section 136 of the Data (Use and Access) Act 2025, with an economic impact assessment under section 135. It states that a broad copyright exception with an opt-out "is no longer the government's preferred way forward", and sets out further evidence-gathering and monitoring rather than legislation; it also proposes removing protection for wholly computer-generated works while keeping protection for AI-assisted works (primary, gov.uk report page; first 100,000 characters read).

Germany

• GEMA v OpenAI (Regional Court Munich I, 42 O 14139/24, 11 November 2025): per commentary, memorisation of song lyrics in model parameters was held to be reproduction, the TDM exception did not cover it, and outputs of lyrics infringed; OpenAI as model operator was held responsible. First instance. OpenAI's appeal is reported as pending at the Higher Regional Court Munich (reported docket 6 U 3662/25 e) (secondary: Morgan Lewis, William Fry, Bird & Bird). I did not find a hearing date or an appeal decision.
• Kneschke v LAION (Regional Court Hamburg, 27 September 2024; Higher Regional Court Hamburg, 10 December 2025, 5 U 104/24): both courts ruled for LAION. The appeal court reportedly relied on the TDM exception in section 44b of the German Copyright Act for the automated image-description matching and, additionally, on the scientific-research exception in section 60d for a non-commercial research body; it held that a rights holder's reservation must be machine-readable and that the plaintiff's natural-language reservation was not, and it limited its reasoning to dataset creation, not later training (secondary: Morgan Lewis, ELBKANZLEI; judgment text not read). It admitted a further appeal; that appeal is pending at the Federal Court of Justice (I ZR 281/25), which heard it on 3 September 2026, is reported to have signalled a referral of questions to the CJEU, and has a pronouncement date reported as 17 December 2026 (secondary: BFF report). I found no report of the outcome or of a CJEU referral having been made.
• Like Company v Google Ireland (CJEU, C-250/25): a Hungarian referral from April 2025 asks whether LLM training and chatbot outputs involve reproduction and communication to the public and whether the TDM exception applies. A public hearing was reported for 10 March 2026; I found no Advocate General opinion or judgment in secondary sources; the Court's own case page was not checked (secondary).

• Divergent first-instance outcomes. US district courts reached different results on training (Bartz, Kadrey for developers on the records before them; Ross against, on a non-generative tool, now affirmed by the Third Circuit); UK and German courts differ on the status of models. No court of final appeal has decided the central training question for generative AI.
• Source of the works matters. In Bartz, the court distinguished training use from the building of a permanent library from pirated copies, which the fair use ruling on training did not excuse; this has practical consequences for data acquisition.
• Market harm theories. Whether outputs that compete without copying (market dilution) defeat fair use is open (Kadrey, primary: the judge described the theory as potentially winning but unproven on that record).
• Opt-out mechanics. The Hamburg appeal court's reported requirement of machine-readable reservations affects how rights holders can exclude their works; the UK Government has said an opt-out exception is no longer its preferred option.
• Licensing. Reported settlements and licences do not establish legal rules.
• Authorship. Human-authorship requirements leave the status of works with substantial but partial AI contribution to case-by-case analysis; the Copyright Office's Part 2 report (29 January 2025) concludes that, given currently available technology, prompts alone do not provide sufficient human control to make users the authors of AI output, while human-authored input, and human selection, arrangement or modification of AI output, can be protected, assessed case by case (primary, report text read).
• Criminal, contract and terms-of-service issues (for example scraping terms) are outside this draft.

As of 2026-10-18:

• There is no settled rule on whether training generative models on copyrighted works is lawful. This is Contested and varies by jurisdiction.
• The Third Circuit affirmed the rejection of a fair-use defence in Ross (non-generative AI) on 29-30 September 2026; the panel reportedly distinguished generative models, so this does not directly resolve generative training, and ROSS has said it will seek rehearing and Supreme Court review. Well Supported as to the Third Circuit's ruling; its reach for generative AI is Contested (opinion text beyond the docket stamp not read).
• The proposition that US copyright requires human authorship is Well Supported by appellate authority (Thaler v. Perlmutter), with certiorari denied; this is a statement about works generated without human authorial contribution, not about AI-assisted works generally.
• The proposition that inventors must be natural persons under UK and US patent law is Well Supported by apex or appellate rulings (Thaler cases; secondary).
• Statutory transparency and opt-out duties for general-purpose AI model providers exist in EU law (AI Act Art. 53).
• In the UK, no statutory text-and-data-mining reform for AI training has been enacted; the Government's March 2026 report defers action.

This section contains no registered forecasts; outcomes of pending litigation are not predicted. Observed open points:

• Near-term (0-3 years). Appeals in Getty (UK), possible rehearing or certiorari in Ross, the appeal in GEMA and the Federal Court of Justice decision in Kneschke (Germany), the CJEU case Like Company, and district-court proceedings such as NYT (summary judgment pending) may clarify whether models contain copies and how fair use applies.
• Medium-term (3-10 years). Legislative responses (licensing, opt-out registries, transparency duties) are possible but not established.
• Long-term (10+ years). The status of increasingly autonomous AI creation under authorship rules is unresolved.

• Foundational Concepts:
  • CAP-GENAI-001 Generative AI: The systems whose training and outputs are at issue.
• Related Methods & Technologies:
  • TECH-DATA-001 Training Data, Contamination and Provenance: Data sourcing and documentation.
  • SAFE-PRIV-001 Privacy and Data Protection in AI: Parallel questions on personal data in training sets.
  • SAFE-SEC-001 AI Security: Extraction attacks and memorisation.
  • SAFE-ETHICS-001 AI Ethics, Bias and Fairness: Ethical dimensions of creative labour and consent.
  • APP-MEDIA-001 AI in Media and Creative Industries: Industry effects.
• Advanced Implications & Future Trajectories:
  • GOV-LAW-001 Global AI Governance & Regulation: Overview of legal responses.
  • GOV-EU-001 The EU AI Act: Copyright-related duties and dates.
  • GOV-US-001 United States AI Policy: US policy context.