# Copyright and AI

> Source: https://aiwiki.ai/wiki/copyright
> Updated: 2026-07-24
> Fact-checked: 2026-07-24
> Categories: AI Ethics, AI Policy & Regulation, Generative AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Copyright and AI." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/copyright
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Copyright and AI is the body of law, agency practice, and litigation that governs how copyright applies to systems that learn from existing works and produce new material. Two questions dominate the field. The output question asks whether anything an AI system generates can be copyrighted at all, and if so who counts as the author. The input question asks whether copying protected works to assemble a training corpus and fit a model infringes the rights of the people who made those works. A third question is narrower and more familiar: when a model reproduces a particular work closely enough, ordinary infringement analysis applies regardless of how the model was built.

The output question has largely settled in the United States and settled the opposite way in China. The input question has settled nowhere. A widely used docket tracker counted 128 US copyright lawsuits against AI developers as of July 19, 2026 [17], and the first federal appellate ruling on whether model training is [fair use](https://aiwiki.ai/wiki/fair_use) had been argued but not decided [11]. Because national legislatures and courts have reached conflicting conclusions about identical conduct, a training run that is plainly lawful in Tokyo may be actionable in Munich.

This article covers copyright doctrine as applied to [generative AI](https://aiwiki.ai/wiki/generative_ai). It is not a general introduction to copyright.

## Copyrightability of AI outputs

### The human authorship requirement

US copyright protects "original works of authorship," and both the courts and the Copyright Office read "authorship" to require a human. The Office's 2025 report traces the requirement through Burrow-Giles Lithographic Co. v. Sarony, which upheld copyright in a photograph because a human made the expressive choices, and through Feist Publications v. Rural Telephone Service Co., which rejected "sweat of the brow" as a substitute for originality [4]. Lower courts have applied the same rule to non-human sources, refusing protection for text attributed to spiritual beings in Urantia Foundation v. Maaherra and for photographs a macaque took of itself in Naruto v. Slater [4].

Nothing in that line of cases bars copyright in works made with machine assistance. The distinction the Office draws is between using AI as a tool and letting it stand in for human creative choice [4].

### Zarya of the Dawn and the 2023 registration guidance

The first concrete test came through the registration system. Kristina Kashtanova registered an eighteen-page comic book, Zarya of the Dawn, on September 15, 2022, without disclosing that its images came from [Midjourney](https://aiwiki.ai/wiki/midjourney). After learning of the omission from her social media posts, the Office reopened the file and, in a letter dated February 21, 2023, cancelled the original certificate (registration VAu001480196) and issued a narrower one. Kashtanova kept copyright in the text and in the selection, coordination, and arrangement of the written and visual elements; the Midjourney images themselves were held not to be products of human authorship [3].

Three weeks later, on March 16, 2023, the Office issued a statement of policy at 88 Fed. Reg. 16190 requiring applicants to disclose AI-generated material and to disclaim any portion that is more than de minimis [2].

The line is not a blanket refusal. In January 2025 the Office registered "A Single Piece of American Cheese," a digital artwork by Invoke chief executive Kent Keirsey produced through roughly 35 rounds of prompt-driven inpainting, on the theory that the human selection and arrangement of the AI-generated elements was itself protectable [8].

### Thaler v. Perlmutter

Stephen Thaler applied to register "A Recent Entrance to Paradise," an image his Creativity Machine system produced, and named the machine as author. The Office refused. The US District Court for the District of Columbia affirmed in 2023, holding that "human authorship is a bedrock requirement of copyright" and that copyright "has never stretched so far" as to cover works generated "absent any guiding human hand" (Thaler v. Perlmutter, 687 F. Supp. 3d 140) [4].

The D.C. Circuit affirmed on March 18, 2025 in No. 23-5233, reading the Copyright Act's provisions as collectively identifying an author as a human being. The panel was careful to limit its holding to works with no human author, noting that disagreements over how much a machine contributed to a human's work were "neither here nor there in this case" [5]. Thaler petitioned for certiorari on October 9, 2025, presenting the question whether works output by an AI system "without a direct, traditional authorial contribution by a natural person" can be copyrighted. The Supreme Court denied the petition on March 2, 2026 (No. 25-449) [6][7].

### The Copyright Office's Part 2 report

The Office's Notice of Inquiry of August 2023 drew more than 10,000 comments from 50 states and 67 countries [4]. Part 2 of the resulting report, on copyrightability, was published on January 29, 2025 [1]. Its conclusions were narrow and, on the central point, restrictive:

| Conclusion | Effect |
| --- | --- |
| Existing law is adequate; no legislation needed | The Office declined to ask Congress for a new regime |
| Assistive use of AI does not affect protection | Tool use is unaffected |
| Human expression in a work survives even if the work also contains AI material | Mixed works are registrable in part |
| Purely AI-generated material, or material with insufficient human control over expressive elements, is not protected | The core exclusion |
| Sufficiency of human contribution is decided case by case | No bright-line test |
| "Based on the functioning of current generally available technology, prompts do not alone provide sufficient control" | Prompting alone does not create authorship |
| Human-authored material perceptible in the output, and creative selection, coordination, arrangement, or modification of outputs, are protectable | The practical route to registration |
| No case made for sui generis protection of AI-generated content | Rejects proposals for a separate right |

Source: US Copyright Office, Part 2 [4].

## Training data and the input question

### The acts alleged to infringe

Plaintiffs generally do not allege that "learning" infringes. They point to specific reproductions: downloading or scraping works to build a corpus, often from shadow libraries or [Common Crawl](https://aiwiki.ai/wiki/common_crawl) derivatives; making working copies during preprocessing and training; in some theories, the trained weights themselves as a copy or a derivative work; and the outputs when they resemble the inputs. Claims under section 1202 of the Digital Millennium Copyright Act, which prohibits removing copyright management information, are frequently pleaded alongside. Because prima facie copying at some stage is hard to deny, almost every case turns on the fair use defense rather than on whether copying occurred [9].

### The Copyright Office's Part 3 report

A pre-publication version of Part 3, on generative AI training, was released on May 9, 2025 and remained the operative version as of July 2026 [1]. It concluded that several stages of model development implicate the exclusive rights, making fair use the decisive question, and offered a set of general observations rather than a rule:

> Various uses of copyrighted works in AI training are likely to be transformative. The extent to which they are fair, however, will depend on what works were used, from what source, for what purpose, and with what controls on the outputs [9].

Each of those variables, the Office added, can affect the market. The report treated deployment for analysis or research as unlikely to substitute for the training works, but said 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." It also introduced market dilution, the theory that a flood of machine-generated substitutes can harm a market without any single output being a copy, as a fourth-factor consideration. On licensing it recommended against legislation, arguing that voluntary markets were growing and that extended collective licensing should be considered only where gaps persist [9].

The report's release was followed immediately by an institutional fight. Register of Copyrights Shira Perlmutter was fired by White House email on May 10, 2025, the day after the pre-publication version appeared. She sued; the D.C. Circuit voted 2-1 to order her reinstated; and on November 26, 2025 the Supreme Court deferred acting on the administration's stay application until it resolved two other removal-power cases [32].

### US litigation

| Case | Court, date | Outcome |
| --- | --- | --- |
| Thomson Reuters v. Ross Intelligence | D. Del., Feb. 11, 2025 | Judge Bibas found 2,243 of 2,830 Westlaw headnotes infringed and rejected fair use, with factors one and four favoring Thomson Reuters; a revised opinion replacing his 2023 ruling [10] |
| Bartz v. [Anthropic](https://aiwiki.ai/wiki/anthropic) | N.D. Cal., June 23, 2025 | Judge Alsup held training on lawfully acquired books "quintessentially transformative" and fair use, and that scanning purchased print copies was a permissible format change, but that retaining a central library of pirated copies was not [12] |
| Kadrey v. Meta | N.D. Cal., June 25, 2025 | Judge Chhabria granted summary judgment to Meta on training [Llama](https://aiwiki.ai/wiki/llama) models, while writing that the ruling "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful," only that "these plaintiffs made the wrong arguments and failed to develop a record in support of the right one" [15] |
| [New York Times v. OpenAI](https://aiwiki.ai/wiki/new_york_times_v_openai) | S.D.N.Y., filed Dec. 27, 2023 | Judge Stein let direct and contributory infringement claims proceed on April 4, 2025; on January 5, 2026 he affirmed an order compelling [OpenAI](https://aiwiki.ai/wiki/openai) to produce a 20 million log sample of de-identified [ChatGPT](https://aiwiki.ai/wiki/chatgpt) conversations [16] |
| Disney and Universal v. [Midjourney](https://aiwiki.ai/wiki/midjourney) | C.D. Cal., filed June 11, 2025 | Studios allege the service generates unauthorized renditions of their characters; Warner Bros. Discovery filed a parallel suit and the litigation was still active in July 2026 [17] |
| [Getty Images v. Stability AI](https://aiwiki.ai/wiki/getty_v_stability_ai) | D. Del., filed 2023 | Getty filed a notice of voluntary dismissal on August 14, 2025 to refile in the Northern District of California [35] |

The Bartz litigation produced the largest settlement so far. The parties reached agreement in September 2025, after Alsup certified a class of authors whose books had been downloaded from shadow libraries [13]. Anthropic agreed to pay $1.5 billion covering roughly 500,000 works, about $3,000 each, and Judge Araceli Martinez-Olguin granted final approval on July 20, 2026 [14]. Alsup had taken inactive status at the end of 2025, sending the case to a new judge [36], who cut the requested attorney fees from $187.5 million to about $101.6 million [37].

The Third Circuit heard argument in the Ross appeal on June 11, 2026, the first time a federal appellate court has confronted the training question directly [11].

## Output similarity and memorization

Whether models retain their training data is an empirical question with direct legal consequences, and the answer is that it depends heavily on the model and the work. Carlini and colleagues showed in a January 2023 paper that [diffusion models](https://aiwiki.ai/wiki/diffusion_model) memorize individual training images and re-emit them at generation time, recovering 109 near-copies of training images from [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion) and 23 memorized examples from Google's Imagen, and found in controlled experiments on CIFAR-10 that diffusion models leak more than twice as much training data as the GANs that preceded them [29].

A larger study by Cooper, Lemley, and co-authors ran more than 3,000 experiments across 200 books and 14 open-weight models. Most models memorized little of most books, but the exceptions were stark: [Llama 3.1](https://aiwiki.ai/wiki/llama_3_1) 70B reproduced Harry Potter and the Sorcerer's Stone almost verbatim from the opening words alone. The authors argued the pattern cuts against clean stories on either side of the litigation [30].

Courts have split on what memorization means legally. The Regional Court of Munich I held on November 11, 2025 in GEMA v. OpenAI (42 O 14139/24), over nine German song lyrics, that reproducible output of complex protected content supports a presumption that the model memorized the work, that representing the work through probability parameters is irrelevant, and that the text and data mining exception therefore did not cover the conduct. The judgment is under appeal [23]. The English High Court reached a structurally opposite conclusion in Getty Images v. Stability AI on November 4, 2025, holding that Stable Diffusion's weights "are not themselves an infringing copy and they do not store an infringing copy," which defeated Getty's secondary infringement claim under sections 22 and 23 of the Copyright, Designs and Patents Act 1988 [18].

## Licensing

Alongside the litigation, a licensing market has formed. Commenters told the Copyright Office that individual and collective licenses for AI training already existed or were under development in music, vocal recordings, and news reports, and pointed to public deals between OpenAI and the Associated Press for news and Shutterstock for images, plus Getty Images collaborations with [NVIDIA](https://aiwiki.ai/wiki/nvidia) and Bria [9]. Terms are usually confidential, and reported values come from journalists rather than the parties.

| Deal | Announced | Terms |
| --- | --- | --- |
| Shutterstock and OpenAI | July 11, 2023 | Six-year agreement licensing Shutterstock's image, video, and music libraries and their metadata; Shutterstock pays contributors from a fund tied to licensing activity [34] |
| News Corp and OpenAI | May 22, 2024 | Multi-year global agreement covering The Wall Street Journal, New York Post, The Times, The Sunday Times, and other mastheads; the WSJ reported a value above $250 million over five years in cash and OpenAI credits [31] |

The Office concluded that voluntary licensing was "fast emerging in certain sectors, although their availability so far is inconsistent," which is why it advised Congress against intervening and suggested extended collective licensing only for gaps unlikely to be filled [9]. The agreements cut both ways in court. Defendants cite them as evidence that innovation and licensing coexist; plaintiffs cite them as proof that a functioning market exists whose value unlicensed training destroys, which is the fourth fair use factor.

## International divergence

### European Union

The EU never relied on fair use. Articles 3 and 4 of the 2019 Digital Single Market Directive create text and data mining exceptions, the second of which permits mining for any purpose, including commercial purposes, unless the rightsholder has expressly reserved the use in an appropriate manner, such as machine-readable means for content made publicly available online. The [EU AI Act](https://aiwiki.ai/wiki/eu_ai_act) layers disclosure on top: Article 53(1)(c) requires providers of general-purpose models to maintain a policy to comply with EU copyright law and to identify and honor reservations expressed under Article 4(3) of the Directive, and Article 53(1)(d) requires a "sufficiently detailed summary about the content used for training," following a template from the AI Office. Both obligations applied from August 2, 2025 [20]. The Commission published the General-Purpose AI Code of Practice on July 10, 2025 and the training content summary template on July 24, 2025 [21].

German courts have applied the exceptions in both directions. In Kneschke v. [LAION](https://aiwiki.ai/wiki/laion) (310 O 227/23), the Hamburg Regional Court held on September 27, 2024 that a non-profit's scraping for dataset construction fell within the scientific research exception; the Higher Regional Court of Hamburg dismissed the photographer's appeal on December 10, 2025 (5 U 104/24) [22]. GEMA v. OpenAI went the other way on commercial training [23].

### United Kingdom

The UK government opened a consultation on copyright and AI in December 2024 whose preferred option was an EU-style exception with a rights reservation. It drew 11,520 responses. Under sections 135 and 136 of the Data (Use and Access) Act 2025 the government was required to report back, and did so on March 18, 2026 [24]. The report abandoned the opt-out approach: on the published analysis 81 percent of respondents favored mandatory licensing and 3 percent supported the opt-out, and the government committed only to further evidence gathering and to exploring narrow exceptions for uses such as non-commercial research. Over 90 percent of respondents backed training data disclosure, but no transparency legislation was proposed [25].

The Getty judgment sharpened the pressure. Getty abandoned its primary training and output copyright claims on June 25, 2025, citing an inability to prove that the relevant acts happened in the UK, leaving secondary infringement, trade mark, and passing off [19]. Mrs Justice Joanna Smith rejected the secondary claim and found only limited historical trade mark infringement over watermarks appearing in outputs [18]. The judgment therefore left the central question, whether UK-based scraping and training infringes, undecided.

### Japan

Article 30-4 of Japan's Copyright Act, added in a 2018 revision, permits exploiting a work "in any way and to the extent considered necessary" where it is not the user's purpose to enjoy, or to cause another to enjoy, the thoughts or sentiments expressed in it, with data analysis named as an example. A proviso withdraws the exception where the act "would unreasonably prejudice the interests of the copyright owner in light of the nature or purpose of the work or the circumstances of its exploitation." The "General Understanding on AI and Copyright in Japan," published by the Legal Subcommittee under the Copyright Subdivision of the Cultural Council and summarized in a May 2024 overview from the Japan Copyright Office, reads the proviso to exclude, among other things, reproducing a database work for training where analysis licenses are available in the market, and reads the exception as inapplicable where an enjoyment purpose is also present [26]. The document is explicitly not legally binding.

### China and India

China took the opposite turn on outputs. In Li v. Liu, case (2023) Jing 0491 Min Chu No. 11279, decided by the Beijing Internet Court on November 27, 2023, the plaintiff generated an image in Stable Diffusion with long positive and negative prompt strings, changed the sampler steps, resolution, CFG scale, seed and LoRA weight, and then added further prompts before settling on the result. The court agreed that the model, not the plaintiff, drew the lines and colors, but held that the prompts and parameters reflected "the plaintiff's choice and arrangement" and that the successive adjustments showed "aesthetic choice and personal judgment," so the image was not a "mechanical intellectual achievement" and met the originality requirement. It ruled that neither the model nor its designer could be the author, and that the plaintiff owned the copyright, while noting he should disclose the AI technology used. The defendant, who had reposted the image with the watermark removed, was ordered to apologize and pay 500 yuan [33]. The judgment is a first-instance decision and is not precedential in the common law sense, but it has been widely read as China's position on AI outputs [27]. On these facts a US examiner would refuse registration.

India's first substantial ruling came on July 24, 2026, when Justice Amit Bansal of the Delhi High Court refused Asian News International an interim injunction against OpenAI, holding that training use was prima facie covered by the fair dealing exception for research in section 52(1)(a) of the Copyright Act while rejecting OpenAI's jurisdictional challenge. The main suit continues [28].

| Jurisdiction | AI-only outputs | Training on protected works |
| --- | --- | --- |
| United States | Not copyrightable; human contribution assessed case by case [4] | Fair use, unresolved; district rulings split [9] |
| European Union | Most member states told a 2024 Council questionnaire that existing principles suffice and that protection requires significant human input [4] | TDM exceptions with a machine-readable opt-out, plus AI Act disclosure duties [20] |
| United Kingdom | Section 9(3) CDPA 1988 assigns authorship of computer-generated works to whoever made the arrangements, but no case law applies it to generative AI [4] | No AI-specific exception; the opt-out proposal was dropped in 2026 [25] |
| Japan | Case by case, weighing prompt detail, number of attempts, selection among outputs, and later human edits [4] | Broad non-enjoyment exception in Article 30-4, subject to a proviso [26] |
| China | Copyrightable where prompting and refinement show personal judgment (Li v. Liu) [33] | Not addressed in Li v. Liu, which concerned an output |
| India | Not directly addressed | Fair dealing for research, prima facie, at the interim stage [28] |

## See also

- [AI copyright](https://aiwiki.ai/wiki/ai_copyright)
- [Fair use](https://aiwiki.ai/wiki/fair_use)
- [New York Times v. OpenAI](https://aiwiki.ai/wiki/new_york_times_v_openai)
- [Getty Images v. Stability AI](https://aiwiki.ai/wiki/getty_v_stability_ai)
- [Perplexity copyright lawsuits](https://aiwiki.ai/wiki/perplexity_copyright_lawsuits)
- [NO FAKES Act](https://aiwiki.ai/wiki/no_fakes_act)

## References

1. US Copyright Office, "Copyright and Artificial Intelligence" (initiative page, reports and publication dates). https://www.copyright.gov/ai/
2. US Copyright Office, "Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence," 88 Fed. Reg. 16190, March 16, 2023. https://www.copyright.gov/ai/ai_policy_guidance.pdf
3. US Copyright Office, letter to Van Lindberg re Zarya of the Dawn (Registration VAu001480196), February 21, 2023. https://www.copyright.gov/docs/zarya-of-the-dawn.pdf
4. US Copyright Office, "Copyright and Artificial Intelligence, Part 2: Copyrightability," January 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
5. Authors Alliance, "Thaler v. Perlmutter: D.C. Court of Appeals Confirms That a Non-Human Machine Cannot Be an Author Under the U.S. Copyright Act," March 19, 2025. https://www.authorsalliance.org/2025/03/19/thaler-v-perlmutter-d-c-court-of-appeals-confirms-that-a-non-human-machine-cannot-be-an-author-under-the-u-s-copyright-act/
6. SCOTUSblog, case page for Thaler v. Perlmutter, No. 25-449. https://www.scotusblog.com/cases/thaler-v-perlmutter/
7. Baker Donelson, "Supreme Court Denies Certiorari in Thaler v. Perlmutter: AI Cannot Be an Author Under the Copyright Act," 2026. https://www.bakerdonelson.com/supreme-court-denies-certiorari-in-thaler-v-perlmutter-ai-cannot-be-an-author-under-the-copyright-act
8. Fennemore, "A Single Slice of Legal History: What the Cheese Copyright Means for AI and IP Law." https://www.fennemorelaw.com/a-single-slice-of-legal-history-what-the-cheese-copyright-means-for-ai-and-ip-law/
9. US Copyright Office, "Copyright and Artificial Intelligence, Part 3: Generative AI Training," pre-publication version, May 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
10. Loeb & Loeb, "Thomson Reuters v. Ross Intelligence Inc.," February 2025. https://www.loeb.com/en/insights/publications/2025/02/thomson-reuters-v-ross-intelligence-inc
11. Baker Botts, "Third Circuit Hears Oral Argument in Ross v. Reuters AI Training," July 2026. https://www.bakerbotts.com/thought-leadership/publications/2026/july/third-circuit-hears-oral-argument
12. Wiggin, "Bartz v. Anthropic: First Court Decision on Fair Use Defense in LLM Training." https://www.wiggin.com/publication/bartz-v-anthropic-first-court-decision-on-fair-use-defense-in-llm-training/
13. Inside Tech Law, "Bartz v. Anthropic: Settlement Reached After Landmark Summary Judgment and Class Certification," September 2025. https://www.insidetechlaw.com/blog/2025/09/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification
14. TechCrunch, "Anthropic's landmark $1.5B copyright settlement is approved," July 20, 2026. https://techcrunch.com/2026/07/20/anthropics-landmark-1-5b-copyright-settlement-is-approved/
15. Copyright Alliance, "Kadrey v. Meta Decision," 2025. https://copyrightalliance.org/kadrey-v-meta-decision/
16. National Law Review, "OpenAI Loses Privacy Gambit: 20 Million ChatGPT Logs Likely Headed to Copyright Plaintiffs," January 5, 2026. https://natlawreview.com/article/openai-loses-privacy-gambit-20-million-chatgpt-logs-likely-headed-copyright
17. ChatGPT Is Eating the World, AI copyright case tracker, update of July 19, 2026. https://chatgptiseatingtheworld.com/
18. Latham & Watkins, "Getty Images v Stability AI: English High Court Rejects Secondary Copyright Claim," November 4, 2025. https://www.lw.com/en/insights/getty-images-v-stability-ai-english-high-court-rejects-secondary-copyright-claim
19. Pinsent Masons Out-Law, "Getty Images drops main copyright claims against Stability AI," June 25, 2025. https://www.pinsentmasons.com/out-law/news/getty-images-stability-ai-copyright-claims-dropped
20. EU Artificial Intelligence Act, Article 53 (obligations for providers of general-purpose AI models). https://artificialintelligenceact.eu/article/53/
21. European Commission, "Explanatory Notice and Template for the Public Summary of Training Content for General-Purpose AI Models," July 24, 2025. https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models
22. DLA Piper, "Robert Kneschke v LAION," December 2025. https://www.dlapiper.com/en/insights/blogs/mse-today/2025/robert-kneschke-v-laion
23. Preu Bohlig, "GEMA v. OpenAI: Judgment of the Regional Court of Munich I," November 2025. https://preubohlig.de/gema-v-openai-judgment-of-the-regional-court-of-munich-i/?lang=en
24. UK Intellectual Property Office, "Report and impact assessment on copyright and artificial intelligence," March 18, 2026. https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence
25. Reed Smith, "UK copyright and AI report: the 'opt-out' is dead, but what comes next?," 2026. https://www.reedsmith.com/articles/uk-copyright-and-ai-report-the-opt-out-is-dead-but-what-comes-next/
26. Japan Copyright Office, Agency for Cultural Affairs, "General Understanding on AI and Copyright in Japan (Overview)," May 2024. https://www.bunka.go.jp/english/policy/copyright/pdf/94055801_01.pdf
27. China Justice Observer, "Landmark Ruling: China's Court Upholds Copyright for AI-Generated Images" (Li v. Liu, (2023) Jing 0491 Min Chu No. 11279, November 27, 2023). https://www.chinajusticeobserver.com/a/landmark-ruling-china's-court-upholds-copyright-for-ai-generated-images
28. LawBeat, "ANI v. OpenAI: Delhi High Court refuses interim injunction against ChatGPT training," July 24, 2026. https://lawbeat.in/news-updates/ani-v-openai-delhi-high-court-refuses-interim-injunction-against-chatgpt-training-1615482
29. Nicholas Carlini et al., "Extracting Training Data from Diffusion Models," arXiv:2301.13188, January 30, 2023. https://arxiv.org/abs/2301.13188
30. A. Feder Cooper, Mark A. Lemley et al., "Extracting Memorized Pieces of (Copyrighted) Books from Open-Weight Language Models," arXiv:2505.12546, May 2025. https://arxiv.org/abs/2505.12546
31. The Wall Street Journal, "News Corp, OpenAI Strike Content Deal," May 22, 2024. https://www.wsj.com/business/media/openai-news-corp-strike-deal-23f186ba
32. SCOTUSblog, "Supreme Court defers decision on whether Trump can fire head of U.S. Copyright Office," November 26, 2025. https://www.scotusblog.com/2025/11/supreme-court-defers-decision-on-whether-trump-can-fire-head-of-u-s-copyright-office/
33. Beijing Internet Court, Civil Judgment (2023) Jing 0491 Min Chu No. 11279 (Li v. Liu), November 27, 2023, official English translation. https://english.bjinternetcourt.gov.cn/pdf/BeijingInternetCourtCivilJudgment112792023.pdf
34. Shutterstock, "Shutterstock Expands Partnership with OpenAI, Signs New Six-Year Agreement to Provide High-Quality Training Data," PR Newswire, July 11, 2023. https://www.prnewswire.com/news-releases/shutterstock-expands-partnership-with-openai-signs-new-six-year-agreement-to-provide-high-quality-training-data-301873298.html
35. IPDE, "Getty Images Voluntarily Dismisses AI Action to Re-File in N.D. Cal.," August 15, 2025. https://ipde.com/blog/2025/08/15/getty-images-voluntarily-dismisses-ai-action-to-re-file-in-nd-cal/
36. Authors Alliance, "Bartz v. Anthropic: Updated Opt Out and Objection Dates, and a New Judge," January 12, 2026. https://www.authorsalliance.org/2026/01/12/bartz-v-anthropic-updated-opt-out-and-objection-dates-and-a-new-judge/
37. OpenClassActions, "Anthropic AI Books Copyright Settlement" (final approval and fee award). https://openclassactions.com/settlements/anthropic-ai-books-copyright-settlement.php