ARTICLE: AI Training and Copyright: Fair Dealing, Opt-Outs or Statutory Licensing? India Compared with the US, EU and UK, AUTHOR: Arya Chandran
Abstract There are many responses to the question of whether developers can legally copy protected works in order to train their AIs. This article will look at how the issue has been dealt with in India such as the Delhi High Court's interim ruling in ANI v OpenAI that an injunction could not be granted and the DPIIT's proposed compulsory blanket licensing. This article will look at how we should move forward in India and I will suggest a combination of fair dealing, limited licensing, transparency and opt-out through contractual agreements will be the most beneficial
ARTICLE
Abstract
There are many responses to the question of whether developers can legally copy protected works in order to train their AIs. This article will look at how the issue has been dealt with in India such as the Delhi High Court's interim ruling in ANI v OpenAI that an injunction could not be granted and the DPIIT's proposed compulsory blanket licensing. This article will look at how we should move forward in India and I will suggest a combination of fair dealing, limited licensing, transparency and opt-out through contractual agreements will be the most beneficial.
1) Introduction
Within 8 months, India have presented two opposing views. Firstly, a government committee suggested in December 2025 that a compulsory blanket licence will be issued for AI Developers to legally train on works which have been legally accessed by paying statutory royalties. Secondly, the Delhi High Court ruled in July 2026 that they could not restrain OpenAI and found that using the articles from the news agency to train fell under fair dealing. There are two clear responses; creators should be compensated and at least for now they cannot prevent the copying.
This article will look at the best way to balance innovation and fair compensation. I will look at how the US, EU & UK have dealt with this issue and argue that India should move forward without basing the policy on the interim fair dealing findings. A limited statutory compensation scheme will be beneficial to both parties with transparency and an opt-out.
2) India – Lenient Judgement & Compulsory Licensing
There are no specific exceptions to text and data mining in India. Copying is an exclusive right and the closest thing is fair dealing for private or personal use including research. The DPIIT committee recognised that statutory reform was needed.
In ANI Media Pvt. Ltd v OpenAI OpCo LLC, Justice Amit Bansal recognised that copying during training would infringe on copying rights however looked to fair dealing. Training was found to be an internal research process and did not replace works from ANI. Public interest was also considered. There was no evidence to suggest ANI lost any subscribers or advertising and did not utilise the opt-outs available. Overall ANI won on the jurisdiction point as the overseas servers did not stop the laws in India applying.
Overall the ruling was quite limited. The order was interim and heavily reliant on the lack of evidence. The output claim failed as copyright protects expression and not facts. The examples given were likely to be retrieval and not memorisation. It will be interesting to see if training will be found to be lawful in a full trial.
The DPIIT working paper proposed a compulsory blanket licence. The paper was published on 8 December 2025 after an 8 member committee produced a report. It suggests a compulsory blanket licence, statutory compensation and a single non-profit collection society to be used. Royalties would be paid when the works are commercialised. The rate would be decided by a committee appointed by the Government which can be reviewed. The owners cannot opt-out. This would also impact current systems.
This has received mixed feedback. Some practitioners feel this will create a level playing field for large and small developers. However, there is no reasoning as to why a specific exception cannot be applied. This results in new rights with little analysis.
3) United States – Fair Use
The US use fair use in the courts. So far it has been very fact dependant. Two judges in California ruled in 2025 that training on books was fair use. However, Judge Alsup ruled that keeping copies of pirated works in a centralised library was not fair use. Judge Chhabria also warned that 'market dilution' could affect the outcome in a full trial.
In the first appeal the Third Circuit ruled on 29 September 2026 that training a legal research tool on Westlaw headnotes was not a fair use. They stated it was "minimally transformative, at best". They did not overturn the decisions made. There are two things to consider. The method of obtaining the data and whether the system trained on competes with the source. Again this was fact dependant like in ANI.
4) European Union – Opt-Outs
The EU have legislated opt-outs. Article 4 of the DSM Directive outlines that text and data mining can be carried out for commercial purposes on legally accessible works unless the rights holder has opted-out by reserving their rights in an appropriate machine readable format. Article 3 outlines this for research. Problems with this are;
Firstly validity. In December 2025 the Higher Regional Court of Hamburg ruled that the reserve must be in a machine readable format. The photographer failed to prove this.
Secondly extent. In GEMA v OpenAI the Munich court found that the lyrics memorised by the model both inside the model and output were infringing. They did not fall within the exception. This is currently being appealed. The Commission have also consulted on standard opt-out processes and a central register as there are issues enforcing their rights and gaining compensation as the owners are opting out.
5) United Kingdom – Default Licensing
The UK tried to implement opt-outs. Section 29A of the Copyright, Designs and Patents Act, 1988 states that only non-commercial research can be copied. However, a report was published on 18th March 2026 under the Data (Use and Access) Act, 2025 which found that the previous governments preference of exception with opt-outs was not suitable. Only 3% of people responded to the consultation in favour of this and 88% wanted compulsory licensing. Transparency is being considered.
The Courts have not provided much clarity. In Getty Images v Stability AI the High Court ruled that if the model weights did not contain the works then it was not an 'infringing copy'. The model was trained overseas and Getty have permission to appeal.
6) Best Model?
It is useful to consider 4 factors. Certainty, Compensation, New Entrants and Enforceability. Fair dealing is quite flexible however does not do well on the other factors. It is very fact dependant and puts the onus on the creators to prove market harm. This was largely why ANI lost. It also relies on there being no opt-outs which goes against the Berne principle of no formalities.
Opt-outs do not work well. As seen in Europe there is confusion around what constitutes a machine readable format. The UK have completely dismissed this method. It also assumes you agree if you do not opt-out. This could cause issues for smaller creators as they would have to deal with the technicalities.
The DPIIT model does overcome many of these issues. Developers will know they can train without issues, creators gain compensation and there is no need for millions of negotiations. It could also remove barriers to entry for smaller developers. Statutory licences are nothing new to India, there is currently one in place for broadcasters (§ 31D). However, there are issues with the draft. There is no opt-out, the rates are set by the state, it is retroactive and there is only one regime in place for vastly differently valued works. If it overrides all the owners decisions it may need defending using the three step test from Berne and TRIPS.
Overall I would support the DPIIT model as it compensates creators for commercial training. However, I would make 4 changes. Firstly, allow fair dealing for non-commercial research and non substitution. This means the Courts have a role to play. Secondly, the licence should only apply to copying for training. Memorisation and regurgitation should still be infringement (as found in Munich). Thirdly, there should be an obligation for developers to be transparent about where they are training from. This was seen as vital in the UK debate. Fourthly, there should be an opt-out for rights holders by entering into direct licences with a minimum rate. This will ensure there is a licensing market. The rates should vary between sectors and be independently reviewed.
7) Counter Arguments & Limitations
Developers will argue that any levy increases costs, discourages start-ups and encourages training overseas. Whilst the Delhi court ruled that India have jurisdiction over services aimed at Indian users, it is unclear how enforceable this will be. Creators will argue they lose control by being compensated without their consent. The rates are set by the state and there is only one collection society which encourages corruption and argument. These are valid points.
There are also limitations to my research. I am relying on a summary of the ANI order as it is 135 pages long. The stats from the UK consultation were taken from secondary sources. They should be cross referenced with primary sources.
8) Conclusion
India have started with fair dealing. This is a weak basis to build a national policy on. Whilst the DPIIT proposal is ambitious there are improvements to be made. Overall there is one clear winner. The US is fact dependant and relies on competition. The opt-outs in Europe cause confusion with what constitutes a machine readable format. The UK are still waiting on legislation for default licensing. By implementing a balanced compensation system, with transparency, opt-out and fair dealing, developers can train with certainty knowing that the creators will benefit from the value. The ANI trial and second part of the DPIIT paper will show how far India will go in this direction.
References
Cases
ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, I.A. 45300/2024 in CS(COMM) 1028/2024, 2026:DHC:5900 (Delhi H.C. July 24, 2026) (India).
Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025).
GEMA v. OpenAI, Landgericht [LG] [Regional Court] München I, Nov. 11, 2025, 42 O 14139/24 (Ger.).
Getty Images (US) Inc. v. Stability AI Ltd. [2025] EWHC 2863 (Ch) (Eng.).
Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025).
Kneschke v. LAION, Landgericht [LG] [Regional Court] Hamburg, Sept. 27, 2024, 310 O 227/23 (Ger.), aff'd, Oberlandesgericht [OLG] [Higher Regional Court] Hamburg, Dec. 10, 2025 (Ger.).
Thomson Reuters Enter. Ctr. GmbH v. ROSS Intel. Inc., No. 25-2153 (3d Cir. Sept. 29, 2026), aff'g 765 F. Supp. 3d 382 (D. Del. 2025).
Legislation and treaties
Agreement on Trade-Related Aspects of Intellectual Property Rights art. 13, Apr. 15, 1994, 1869 U.N.T.S. 299.
Berne Convention for the Protection of Literary and Artistic Works arts. 5(2), 9(2), Sept. 9, 1886, as revised July 24, 1971, 828 U.N.T.S. 221.
Copyright Act, 1957, No. 14 of 1957, §§ 14(a)(i), 31D, 52(1)(a)(i), India Code (1957).
Copyright, Designs and Patents Act 1988, c. 48, § 29A (UK).
Data (Use and Access) Act 2025, c. 18, §§ 135–136 (UK).
Directive 2019/790, of the European Parliament and of the Council of 17 April 2019 on Copyright and Related Rights in the Digital Single Market and Amending Directives 96/9/EC and 2001/29/EC, arts. 3–4, 2019 O.J. (L 130) 92.
17 U.S.C. § 107 (2018).
Government sources
Dep't for Promotion of Indus. & Internal Trade, Gov't of India, Working Paper on Generative AI and Copyright, Part 1: One Nation One Licence One Payment (Dec. 8, 2025), https://openfuture.eu/wp-content/uploads/2026/02/251208one_nation_one_license_one_payment.pdf.
Articles and online sources
Laura Adde, Getty Images v Stability AI: UK High Court Finds No Secondary Copyright Infringement in Landmark GenAI Ruling, AIPPI (Jan. 23, 2026), https://www.aippi.org/news/getty-images-v-stability-ai-uk-high-court-finds-no-secondary-copyright-infringement-in-landmark-genai-ruling/.
AI Sued by Image Library for Intellectual Property Infringement in Training Models, UK Human Rights Blog (Nov. 7, 2025), https://ukhumanrightsblog.com/2025/11/07/ai-sued-by-image-library-for-intellectual-property-infringement-in-training-models/.
AI Training and Fair Use After the First Appellate Ruling: Thomson Reuters v. ROSS (3d Cir. 2026), Legal 500 (Nov. 2, 2026), https://www.legal500.com/intelligence/united-states/intellectual-property/ai-training-and-fair-use-after-the-first-appellate-ruling-thomson-reuters-v-ross-3d-cir-2026.
ANI v. Open AI: Delhi High Court Refuses Interim Injunction in Landmark AI Copyright Dispute, Cyril Amarchand Mangaldas Blogs (July 2026), https://corporate.cyrilamarchandblogs.com/2026/07/ani-v-open-ai-delhi-high-court-refuses-interim-injunction-in-landmark-ai-copyright-dispute/ (last visited Nov. 10, 2026).
Centre Proposes Mandatory Blanket Licence for AI Training on Copyrighted Works, Royalties for Creators, LiveLaw (Dec. 16, 2025), https://www.livelaw.in/amp/ipr/centre-mandatory-license-for-ai-training-ai-training-copyrighted-works-royalties-for-creators-513425.
Espie Angelica A. de Leon, India Proposal on AI Developers' Use of Copyrighted Content Needs Clarity, Says Lawyer, Asia IP, https://asiaiplaw.com/sector/licensing-franchising/india-proposal-on-ai-developers-use-of-copyrighted-content-needs-clarity-says-lawyer (last visited Nov. 10, 2026).
EU Consults on Copyright Landscape in the Age of Generative AI, Legal 500, https://www.legal500.com/intelligence/european-union/intellectual-property/eu-consults-on-copyright-landscape-in-the-age-of-generative-ai (last visited Nov. 10, 2026).
Gen AI, Copyrights, and Hybrid Licensing in India: Why the Assumptions May Not Sustain the Model, BananaIP Intellepedia, https://www.bananaip.com/intellepedia/?p=147195 (last visited Nov. 10, 2026).
Hogan Lovells, AI Litigation Case Law Tracker: GEMA v. OpenAI, https://digital-client-solutions.hoganlovells.com/resources/ai-litigation-case-law-tracker/case/gema-v-openai (last visited Nov. 10, 2026) (noting appeal pending before the Higher Regional Court of Munich).
Christina Meese, [DE] Munich Regional Court Upholds GEMA's Claim Against OpenAI for Unauthorised Reproduction of Song Lyrics, IRIS 2025-10:1/6 (2025), https://merlin.obs.coe.int/article/10428 (last visited Nov. 10, 2026).
Alishan Naqvee, Saher Gandhioke & Amir Shejeed, India's First AI Copyright Ruling: The Delhi High Court in ANI v. OpenAI, Mondaq (Aug. 12, 2026), https://www.mondaq.com/india/copyright/1829216/indias-first-ai-copyright-ruling-the-delhi-high-court-in-ani-v-openai.
Emilio B. Nicolas, Training Day: Third Circuit Affirms Westlaw Headnotes Decision in AI Copyright Fair Use Case, Jones Walker LLP (Oct. 5, 2026), https://www.jw.com/news/insights-ai-copyright-fair-use-case/.
One Nation, Forced Licenses, Multiple Payments: (Un)Balancing AI Innovation & Copyright, SpicyIP (Jan. 2026), https://spicyip.com/2026/01/one-nation-forced-licenses-multiple-payments-unbalancing-ai-innovation-copyright.html (last visited Nov. 10, 2026).
Modhura Roy, ANI v OpenAI Ruling: Reading the Tea Leaves, Law.asia (2026), https://law.asia/ai-training-and-copyright/ (last visited Nov. 10, 2026).
Text and Data Mining Copyright Exception: The UK's 2026 Decision, CASRAI (July 3, 2026), https://casrai.org/news/text-data-mining-exception.
Text and Data Mining for AI Training, Hannes Snellman, https://www.hannessnellman.com/news-and-views/blog/text-and-data-mining-for-ai-training/ (last visited Nov. 10, 2026).
What the March 2026 Report Actually Says, Algeria Tech News (2026), https://algeriatech.news/uk-ai-copyright-data-use-access-act-2026/ (last visited Nov. 10, 2026).
Author Note
Arya Chandran is a BA LL.B. student at Lord’s Universal College of Law, Mumbai. She is particularly interested in Intellectual Property, Data Protection, and the way the law responds to new technology. Alongside her studies, she enjoys Legal Research and Writing, and following current Legal Developments.