Mayank Chaudhary
Fairfield Institute of Management and Technology
Abstract
Generative AI systems learn from enormous collections of books, articles, images and code, much of it protected by copyright. Whether that learning is “fair use” is now the central question in AI copyright law. This article sets out the doctrine, follows the main American decisions from 2025 to 2026, and compares India’s narrower fair dealing regime, including the Delhi High Court’s interim order in ANI v. OpenAI and the government’s licensing proposal. It argues that fair use is being stretched from a doctrine about copying into one about markets.
Keywords: generative AI, fair use, fair dealing, copyright, training data, transformative use
1. Introduction
Fair use was designed for the critic quoting a book, the teacher photocopying a page or the parodist borrowing a tune. Generative AI asks it to answer a different question: may a company copy millions of works, in full, to build a commercial system that then produces new text and images?
Courts have started to answer, but not in one voice. Some judges have called training highly transformative; others have focused on lost licensing income and on how the data was obtained. The law is changing quickly, so the position below is stated as at October 2026, and readers should check the primary sources at the end.
2. The Fair Use Doctrine
2.1 The statutory test
Section 107 of the US Copyright Act allows some unlicensed uses, such as criticism, teaching and research, after weighing four factors. The Supreme Court has treated the test as a flexible, case-by-case inquiry, not a checklist. In Campbell v. Acuff-Rose Music (1994) it asked whether a new use adds something with a further purpose or different character, and in Harper & Row v. Nation Enterprises (1985) it described the effect on the market as an especially important element.
2.2 Earlier technology cases
Courts had already faced mass copying for new technologies. Sony v. Universal (1984) protected home recording of broadcasts. In Authors Guild v. Google (2015), the Second Circuit held that scanning millions of books to build a searchable index was fair use because users saw only snippets. The Supreme Court tightened the first factor in Andy Warhol Foundation v. Goldsmith (2023), warning that a new use which serves the same purpose as the original, and is commercial, is less likely to be transformative. Both strands are now cited in AI disputes.
3. Why Generative AI Strains the Doctrine
Two stages. Litigation separates the input stage, where works are copied into training sets, from the output stage, where a model responds to prompts. A model may be lawful at one stage and not the other, for example if it can be prompted to reproduce passages.
Scale and substitution. Classic fair use cases involved a single borrowed work. Here the copying is total, but any single work contributes little to the result. Rights holders reply that a system trained on their works can flood the market with competing content.
Source of the data. Whether copies came from lawful purchases, licensed access or pirate repositories has emerged as a distinct issue, separate from what the model later does.
4. The US Case Law, 2025 to 2026
4.1 Thomson Reuters v. Ross
In February 2025, a federal court in Delaware rejected fair use where a start-up used Westlaw headnotes to train a competing legal search tool. The system was not generative. On 29 September 2026, the Third Circuit affirmed, reportedly describing the use as commercial and only minimally transformative, and noting that generative systems might present a different analysis. It is the first appellate ruling on fair use in AI training, but its reach is narrow.
4.2 Bartz v. Anthropic
In June 2025, Judge Alsup held that training a language model on lawfully acquired books was “exceedingly transformative” and a fair use. He refused the same protection for a central library built from pirated copies, which went forward to trial. That dispute later ended in a widely reported settlement of about US$1.5 billion. The decision therefore separates how a model learns from how the data was obtained.
4.3 Kadrey v. Meta
Two days later, Judge Chhabria also ruled for Meta, but on a narrower footing. He found the authors had not shown market harm on the record before him, and warned that in many circumstances training would not be fair use. He identified “market dilution”, meaning a flood of competing AI works that undercuts human authors, as a possibly decisive argument in a better-prepared case.
4.4 The Copyright Office view
In May 2025, the US Copyright Office published a pre-publication report on generative AI training. It concluded that some training uses will be fair and others will not, depending on the data, the purpose and the outputs, and it took seriously the harm caused by market dilution and by lost licensing.
5. The Indian Position
India has no open-ended fair use. Section 52 of the Copyright Act, 1957 lists permitted acts, and fair dealing under section 52(1)(a) is confined to named purposes such as private use including research, criticism or review, and reporting current events. Copying a work without permission is otherwise infringement under section 51.
In ANI Media v. OpenAI the Delhi High Court, on 24 July 2026, refused an interim injunction. On the output claim, it held that copyright protects expression and not facts, and the sample articles postdated training, so the responses likely came from retrieval and not memorised data. On the training claim, it accepted that storing works for training falls within the Act but weighed it against fair dealing, calling training an internal research process that did not substitute for ANI’s works. It also noted that ANI had not used available crawler opt-outs. The order is interim, and the trial will decide the merits.
Policy is also moving. In December 2025, the Department for Promotion of Industry and Internal Trade issued a working paper proposing a mandatory blanket licence, under which developers may train on lawfully accessed works and pay statutory royalties through a collective body. It is a proposal open to comment and not yet law.
6. Critical Analysis
• Transformation is doing too much work. If building a model counts as transformative in itself, the first factor stops constraining anything. After Warhol, courts must ask whether the use serves the same purpose as the original.
• The licensing circle. If a licensing market for training data counts as a market harm, a growing market makes fair use harder to claim, even though the market exists only if fair use fails. The Third Circuit’s treatment of this point, and its application to generative models, remains open.
• Lawful access is the weak link. Decisions so far reward developers who bought or licensed their data and penalise those who took pirated copies, so provenance now matters more than model design.
• Outputs are mostly untested. Courts have said little about regurgitation, style imitation and retrieval features that reproduce news content.
• Different roads in different countries. The US relies on case-by-case fair use, the EU on text and data mining exceptions with rights-holder opt-outs, and India is considering statutory licensing. Developers training across borders face conflicting rules.
7. The Way Forward
1. Maintain records of training data sources, so that lawful acquisition can be shown.
2. Develop collective or statutory licensing, so that small creators and small developers are both served.
3. Adopt clear opt-out standards for crawlers, and honour them.
4. Treat outputs separately from training, with duties to prevent verbatim reproduction.
5. Let appellate courts and legislatures settle the generative AI question, rather than extending narrow precedents by analogy.
8. Conclusion
Fair use is not collapsing under generative AI, but it is being redirected. The early decisions suggest that learning from lawfully obtained works may be fair, while piracy, close competition and market substitution weigh heavily against it. The next phase will be decided by appeals, by the outputs that models actually produce, and, in India, by the choice between fair dealing and statutory licensing. Until then, uncertainty remains the working rule.
9. Frequently Asked Questions
Q1. Is training an AI model on copyrighted works always fair use?
No. Decisions so far turn on the facts: the purpose, how the data was obtained, whether the system competes with the works and what the outputs contain.
Q2. What did the Third Circuit decide in Thomson Reuters v. Ross?
On 29 September 2026, it affirmed that using Westlaw headnotes to train a competing non-generative search tool was not fair use. Its reach to generative AI is limited.
Q3. Does India recognise fair use?
No. Indian law uses fair dealing for listed purposes under section 52 of the Copyright Act, 1957, although courts may interpret those purposes broadly.
Q4. What did the Delhi High Court decide in ANI v. OpenAI?
It refused an interim injunction on 24 July 2026, finding no prima facie case on the output claim and leaning towards fair dealing on training. The suit continues.
Q5. What is the DPIIT licensing proposal?
A December 2025 working paper suggesting a mandatory blanket licence for training on lawfully accessed works, paired with statutory royalties for rights holders.
References
Cases
• Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994).
• Harper & Row v. Nation Enterprises, 471 U.S. 539 (1985).
• Sony Corp. v. Universal City Studios, 464 U.S. 417 (1984).
• Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015).
• Andy Warhol Foundation v. Goldsmith, 598 U.S. 508 (2023).
• Thomson Reuters v. Ross Intelligence (D. Del., 11 February 2025; 3d Cir., 29 September 2026).
• Bartz v. Anthropic PBC (N.D. Cal., 23 June 2025).
• Kadrey v. Meta Platforms, Inc. (N.D. Cal., 25 June 2025).
• ANI Media Pvt. Ltd. v. OpenAI OpCo LLC (Delhi High Court, interim order, 24 July 2026).
Statutes and reports
• 17 U.S.C. section 107; Copyright Act, 1957 (India), sections 51 and 52.
• US Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version, May 2025).
• DPIIT, Working Paper on Generative AI and Copyright, Part 1 (December 2025).
Note: Citations, figures and dates are given to the best of the author’s knowledge, and several decisions are very recent or interim. Please check them against official sources before use in academic or professional work.
This article is for academic and educational purposes and is not legal advice.
