On 1 September 2026, the US Department of Justice filed a statement of interest in the New York Times litigation against OpenAI arguing that training AI models on copyrighted works is fair use. The next day, Commerce Secretary Howard Lutnick urged G20 ministers to embrace fair use for AI training. For developers and rights holders working between Europe, Latin America and the US, the message only reaches as far as US law: the same training can still infringe in Munich or Madrid.
Key takeaways
- The statement of interest is a government position in a pending case, not a ruling on fair use.
- It calls training “exceedingly transformative” and the “market dilution” theory of Kadrey v Meta deeply flawed.
- The G20 statement agreed in Chapel Hill did not adopt fair use: it leaves copyright exceptions for AI to each member’s own legal processes.
- In the EU, training relies on the text and data mining exception and its opt-out, and the AI Act applies to models placed on the EU market wherever they were trained.
- In July 2026, a Munich court applied US law to US training and still rejected Suno’s fair use defence.
What did the US statement of interest say about fair use and AI training?
The United States filed under 28 U.S.C. § 517, which lets the Department of Justice set out the federal interest without becoming a party. The statement was filed in In re OpenAI, Inc. Copyright Infringement Litigation (No. 25-md-3143, Southern District of New York), the consolidated case that includes the New York Times claim. Its main arguments:
- Training is a separate use from outputs. Copying works to teach a model statistical patterns serves a different purpose from the works, so the first fair use factor favours the developer.
- On the fourth factor (effect on the market), only substitution through substantially similar copies counts; outputs that merely compete in the same genre cause no cognisable harm.
- Kadrey wrongly merged training and outputs, and the Copyright Office’s May 2025 report on AI training deserves no deference on this point.
- Outputs that reproduce protected text can still be assessed case by case.
It adds that mandatory licensing would favour the largest technology companies, without taking a position on feasibility. It does not bind the court, which had not yet ruled on fair use.
What did the G20 agree in Chapel Hill?
At the G20 technology meeting in Chapel Hill, North Carolina, Lutnick told ministers to allow AI companies to train on creators’ work while protecting artists, and to embrace fair use, Reuters reported. The consensus text went less far. Pillar 4 of the G20 Innovation Ministerial Statement recognises doctrines such as prior and express consent and limitations and exceptions to copyright, and says their application to AI remains a matter for each member’s established legal processes. There is no harmonised fair use rule.
Does a US fair use defence protect training in Europe and Latin America?
No. Copyright is territorial, and courts apply the law of the country for which protection is claimed.
| Jurisdiction | Basis for training on protected works | Status in September 2026 |
|---|---|---|
| United States | Open fair use test (17 U.S.C. § 107), case by case | 2025 district court decisions (Bartz, Kadrey) found training transformative; New York Times case pending |
| European Union | Text and data mining exception (Article 4, Directive 2019/790), unless rights are reserved in machine-readable form | AI Act Article 53(1)(c) requires general-purpose model providers to respect those reservations |
| Germany | Same EU exception, applied by national courts | GEMA v Suno (Munich, July 2026): fair use rejected for US training; not final |
| Brazil | Closed list of exceptions (Article 46, Law 9,610/1998); no general fair use or mining exception | No specific exception for training: a licence is the safe route |
Recital 106 of the AI Act is explicit: providers placing general-purpose models on the EU market must meet EU copyright obligations regardless of where the training took place.
Same training, opposite results: the Suno contrast
On 31 July 2026, the Munich Regional Court I (case 42 O 763/25) largely upheld GEMA’s claims against Suno. For the training carried out in the US, it applied US law under the country-of-protection principle and found that all the Warhol fair use factors weighed against Suno, distinguishing Bartz and Kadrey because simple prompts produced outputs substantially similar to the songs. The judgment is not final, but a European court will not necessarily read fair use the way Washington does.
What this means for your business
- Map where each model is trained, fine-tuned and offered: the law of each market can apply.
- If you place models on the EU market, document how you identify and respect mining opt-outs.
- If you are a publisher or rights holder, reserve your rights in machine-readable form and keep dated evidence.
- Review AI vendor indemnities that only work under US law.
If you need training, licensing and opt-out decisions aligned across several countries, our team for AI training data, fair use and digital asset protection can review your position. AI Act compliance is covered within the same META Channel group.
Where companies get this wrong
- Exporting a US legal opinion. A memo concluding that training is fair use says nothing about Germany, Spain or Brazil.
- Reading a statement of interest as a judgment.
- Ignoring outputs, which even the US filing assesses separately.
- Choosing a forum too late; our cross-border IP strategy team plans where to act before a dispute starts.
Frequently asked questions
Is the US statement of interest a court ruling on fair use?
No. It is a filing in which the federal government sets out its view under 28 U.S.C. § 517. The court in the OpenAI litigation is free to disagree, and fair use is decided on the facts of each case. Until there is a judgment, and probably an appeal, the question remains open in the United States.
Did the G20 adopt fair use for AI training?
No. The Chapel Hill statement recognises consent and limitations and exceptions to copyright, but leaves their application to AI to each member’s own legal processes. The United States argued for fair use; the consensus text does not impose it on anyone.
If my model was trained in the US, can I still be sued in Europe?
Yes. A rights holder can sue in Europe over acts there, such as offering the model or its outputs, and in GEMA v Suno the Munich court also assessed US training under US law and rejected fair use. The AI Act adds copyright duties for models placed on the EU market wherever trained.
Can IP Global Guard coordinate an AI copyright strategy across Europe, Latin America and the US?
Yes. We review training data, opt-outs, licences and vendor contracts across Europe, Latin America and Africa from a single point of contact, coordinating qualified local correspondents. For corridor clients with US exposure, we work alongside US counsel so both sides of the strategy stay aligned.
How IP Global Guard can help
The US and the EU are moving in different directions on AI training, and Latin American laws follow neither model fully. IP Global Guard, the IP services line of META Channel Corporation Limited, aligns training, licensing and enforcement decisions across more than 25 jurisdictions in Europe, Latin America and Africa with one strategy and one billing relationship.
Tell us where your models are trained and offered, or which works you need to protect, and we will map the exposure market by market. Contact our AI and copyright team.
This article is general information, not legal advice, and reflects the position on its publication date.
Sources
- US Department of Justice, Statement of Interest, In re OpenAI, No. 25-md-3143 (1 September 2026)
- ChatGPT Is Eating the World (2 September 2026)
- G20 Innovation Ministerial Statement (2 September 2026)
- Reuters via The Standard (3 September 2026)
- Landgericht München I, press release 16/2026 (31 July 2026)
- Directive (EU) 2019/790 (17 April 2019)
- AI Act, Article 53 and Recital 106 (2024)
- Brazil, Law 9,610/1998, Article 46
