AI Training Is Always Legal Because It's Public Online: Training Data and Copyright Disputes

A persistent claim circulates in AI circles: if music is publicly available online, AI companies can freely use it for training. The reasoning is simple. Publicly accessible content is, by definition, public. If you did not want your music used for AI training, you should not have posted it online.

This argument is appealing but legally unsound. Public availability and permission to copy are fundamentally different concepts in copyright law. The distinction matters enormously for indie musicians whose work has likely already been used to train AI systems without their knowledge or consent. Whether the music those systems produce is yours to own is a separate question, covered in Is AI music legal? Who owns AI-generated music.

Public Availability Is Not Permission

Copyright protection attaches automatically when a work is created and fixed in a tangible medium. Posting music on a website, streaming platform, or social media does not waive copyright. It does not place the work in the public domain. It does not grant the world permission to copy, distribute, or create derivative works.

This principle is foundational to copyright law. A book in a public library is publicly accessible, but that does not mean anyone can photocopy it. A painting displayed in a gallery is visible to all, but that does not mean anyone can reproduce it. Public availability and copyright protection coexist without contradiction.

AI companies that scrape publicly available music for training data are making copies of copyrighted works. Whether those copies are legally permissible depends on other factors, not on the mere fact of public availability.

The Fair Use Question in the United States

In the U.S., the primary legal defense for AI training on copyrighted works is fair use. Section 107 of the Copyright Act allows copying for purposes such as criticism, comment, news reporting, teaching, scholarship, or research. Whether AI training qualifies as fair use is one of the most contested legal questions in technology law today.

Fair use analysis considers four factors:

Purpose and character of the use. AI training is typically commercial in nature, which weighs against fair use. However, courts also consider whether the use is “transformative.” AI companies argue that training is transformative because it creates a new tool rather than reproducing the original works. Critics argue that the purpose is to create a product that competes with the works it was trained on.

Nature of the copyrighted work. Music is a creative work, which typically weighs against fair use. However, this factor is less significant when the use is transformative.

Amount and substantiality of the portion used. AI training typically involves copying entire works, which weighs against fair use. However, some courts have found that copying entire works can be fair use when necessary for a transformative purpose.

Effect on the market. This is often the most important factor. If AI-generated music competes with the music it was trained on, it could harm the market for the original works. This is a particular concern for indie musicians who compete in the same markets as AI-generated content.

The fair use question has not been definitively resolved. Multiple lawsuits are pending, and different courts may reach different conclusions. Until the Supreme Court or Congress provides clarity, the legal status of AI training remains uncertain. Notably, the Munich Regional Court’s July 2026 ruling in GEMA v. Suno directly addressed U.S. fair use doctrine for the first time in a European court. The German court rejected Suno’s fair use defense on all four factors, distinguishing the case from Bartz v. Anthropic and Kadrey v. Meta because Suno’s model produced outputs substantially similar to the training data — something those U.S. cases did not involve. The court found Suno’s use commercial and non-transformative, the works highly creative, the amount copied entire, and the market effect severe because AI-generated outputs could substitute for the originals. While a German court’s fair use analysis does not bind U.S. judges, it signals that courts globally are scrutinizing the transformative-use argument skeptically when AI outputs closely resemble training data.

The first U.S. appellate data point arrived on September 29, 2026, when the Third Circuit affirmed summary judgment in Thomson Reuters v. Ross Intelligence, holding that copying copyrighted material to train an AI tool was not fair use. The panel found the use “minimally transformative at best,” the amount taken qualitatively significant, and the market harm material because the copying usurped a developing market for licensing content as AI training data. The decision does not resolve generative AI training: ROSS built a non-generative research tool that competed directly with the plaintiff’s product, and the court did not address models that generate new works. Its practical significance for music rightsholders lies in the fourth factor — the treatment of a growing training-data licensing market as evidence of harm, which is the same argument the labels press against Suno and Udio.

The EU Approach: Transparency and Opt-Outs

The European Union has taken a different approach to AI training data through a combination of copyright law and AI-specific regulation.

The EU Copyright Directive (2019) includes a text and data mining exception that permits copying for research purposes by research organizations and cultural heritage institutions. For commercial entities, the exception applies only if the rights holder has not opted out.

This opt-out mechanism is significant. It means that in the EU, rights holders can potentially prevent their works from being used for AI training by expressing their objection. However, the practical implementation of opt-outs remains challenging. How does a musician opt out? What constitutes an effective opt-out? These questions are still being worked out.

The EU AI Act (2024) adds another layer of regulation. General-purpose AI model providers must:

• Provide detailed summaries of the content used for training
• Comply with EU copyright law, including respecting opt-outs
• Maintain transparency about their training data

These requirements create a framework where rights holders have more information and potentially more control. As of August 2, 2026, the EU AI Act’s Article 50 transparency obligations became fully enforceable, requiring AI-generated content to be marked in a machine-readable format and compelling providers to publish detailed training data summaries.

The UK Situation

The United Kingdom presents an interesting case study in how policy can shift. In 2022, the UK government proposed a broad text and data mining exception that would have allowed AI training on copyrighted works without permission. This proposal faced strong opposition from creative industries and was ultimately shelved.

The UK Intellectual Property Office has since conducted further consultations. In March 2026, Parliament’s Communications and Digital Committee concluded that UK copyright law is not fit to protect creative works from AI, and in May 2026 the Government confirmed it had dropped its earlier preference for a broad training exception with an opt-out, choosing instead to encourage a licensing market alongside voluntary transparency commitments. As of September 2026, no legislation has been enacted, so the current legal position is unchanged: AI training on copyrighted works requires permission unless an existing exception applies.

The UK’s existing text and data mining exception is narrower than what AI companies would prefer. It applies primarily to non-commercial research, and its applicability to commercial AI training is disputed.

For indie artists, the UK situation illustrates how quickly the legal landscape can change. A government proposal that would have legalized AI training was withdrawn due to creative industry pressure. Future policy changes could go in either direction.

Ongoing Lawsuits and Their Implications

Several high-profile lawsuits are testing the legality of AI training on copyrighted works. While most focus on visual art and text, the outcomes will likely affect music as well.

Andersen v. Stability AI: Visual artists sued an AI company for using their copyrighted works in training data. The case raises fundamental questions about whether AI training constitutes fair use.

Getty Images v. Stability AI: A stock photo company sued for the unauthorized use of millions of images in AI training. This case tests whether commercial AI training requires a license.

The New York Times v. OpenAI: A major media company sued an AI developer for using its articles in training data. The case explores the boundaries of fair use in the AI context.

Thomson Reuters v. Ross Intelligence (3d Cir., Sept. 29, 2026): The Third Circuit held that copying copyrighted material to train a competing AI tool was not fair use — the first U.S. appeals court decision on the question. The tool at issue was non-generative, so the ruling does not decide whether generative AI training qualifies as fair use.

Music industry lawsuits: Several cases have been filed against AI music generators, alleging that they copied copyrighted music without permission. These cases directly address the legality of AI training on musical works:

• RIAA v. Suno and RIAA v. Udio (filed June 2024): Major record labels sued both AI music generators for allegedly copying copyrighted recordings to train their models. These are the leading U.S. cases on AI music training.
• GEMA v. Suno (Germany, July 2026): The Munich Regional Court ruled that Suno infringed the copyrights of GEMA members by training on their protected musical works without permission. This is the first court ruling in any jurisdiction to find that an AI music generator violated copyright through its training process. The court ordered Suno to stop using GEMA-protected music. While this ruling applies under German law and does not directly set U.S. precedent, it signals that courts in other jurisdictions may reach similar conclusions.
• UMG v. Suno II (filed September 18, 2026): Universal Music Group and Sony Music filed a second lawsuit against Suno, this time targeting its v6 model suite launched on September 9, 2026. The complaint alleges that v6 was built using “knowledge distillation” from earlier models trained on unlicensed recordings — meaning the v6 models embody the “accumulated learnings” of infringing predecessors. The labels argue that “training a ‘new’ model on the outputs of an infringing model does not eliminate the infringement; it launders it.” The suit seeks damages for 60,202 specific recordings at up to $150,000 per work for willful infringement.
• Sony Music v. Udio (filed July 2026): Sony Music filed a new lawsuit alleging that Udio copied approximately 30,000 songs to train its AI model, adding to the legal pressure on AI music companies.
• Justice v. Suno (D. Mass., decided Aug. 20, 2026): A putative class action by independent artists survived most of Suno’s motion to dismiss. The court held that plaintiffs need not identify a specific infringing output at the pleading stage, and that their allegation that Suno circumvented YouTube’s rolling cipher to obtain training data states a claim under the DMCA. The order lets the case proceed; it makes no finding that Suno infringed.
• Bartz v. Anthropic (settlement approved July 2026): A federal judge gave final approval to a $1.5 billion settlement — the largest copyright settlement in history. The case involved Anthropic’s use of copyrighted books to train its AI model. While the case involved literary works rather than music, the settlement amount and the court’s willingness to treat AI training as actionable copyright infringement may influence how music-related AI cases proceed. The settlement works out to approximately $3,000 per work for an estimated 500,000 works.
• Sony Music Publishing & Warner Chappell v. Anthropic (filed August 28, 2026): The publishing arms of Sony and Warner, joined by other music publishers, sued Anthropic and its co-founders in the U.S. District Court for the Northern District of California, alleging that the company torrented, scraped, and downloaded copyrighted compositions — including lyrics and sheet music — to train its Claude models. The complaint claims “tens of thousands” of compositions and seeks up to $150,000 per work for willful infringement, plus up to $25,000 for each removal of copyright management information.

These lawsuits are in various stages of litigation, and their outcomes are uncertain. Different courts may reach different conclusions, creating a patchwork of legal precedent. Appeals are likely, and final resolution may take years. However, the German ruling in GEMA v. Suno marks a significant development: for the first time, a court has found that training an AI model on copyrighted music without permission constitutes infringement. The Bartz v. Anthropic settlement further signals that the financial exposure for companies that train on copyrighted material without permission can be enormous.

For indie artists, the practical reality is that their music has almost certainly been used in AI training without their consent. Whether they have legal recourse depends on the outcomes of these cases and any legislation that may follow.

What This Means for Indie Artists

The uncertainty around AI training data has several practical implications for musicians:

Your music has likely been used. If your music was publicly available online, it has probably been scraped and used in AI training datasets. This includes music on streaming platforms, social media, personal websites, and digital storefronts. The scale of AI-generated music production underscores the problem: as of July 2026, Deezer reported receiving nearly 90,000 fully AI-generated tracks per day, representing over 50% of all new uploads on peak days. This flood of synthetic content competes directly with human-created music for listener attention and streaming royalties.

You may have limited recourse. Even if AI training on your music was illegal, proving it and obtaining compensation is difficult and expensive. Class action lawsuits may eventually provide some relief, but individual enforcement is impractical for most indie artists.

Terms of service matter. The platforms where you host your music may have terms of service that affect your rights regarding AI training. Some platforms have explicitly allowed or prohibited AI scraping.

The landscape is evolving. Legislation, court decisions, and industry practices are all in flux. What is permissible today may be prohibited tomorrow, and vice versa.

Collective action may be more effective than individual action. Industry organizations, collecting societies, and advocacy groups are working to establish frameworks that protect creators’ rights in the AI era.

Common Misconception

“If it’s online, it’s free to use for AI training” — Public availability does not equal permission. Copyright law protects works regardless of where they are published. The legality of AI training depends on fair use analysis, licensing agreements, and jurisdiction-specific rules, not merely on public accessibility.

Practical Checklist

• Assume your publicly available music has been used in AI training
• Review the terms of service on platforms where you host your music
• Consider whether opt-out mechanisms are available in your jurisdiction
• Monitor ongoing lawsuits and legislative developments
• Join or support industry organizations advocating for creator rights
• Consult a lawyer if you believe your rights have been significantly violated

When to Ask a Lawyer

• You have evidence that a specific AI company copied your music for training
• You are considering joining a class action lawsuit related to AI training
• You need to understand your rights under specific platform terms of service
• You are entering a licensing agreement with an AI company

Sources

1. U.S. Copyright Office, “Copyright and Artificial Intelligence, Part 3: Generative AI Training” (May 2025), https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
2. European Parliament, “EU AI Act” (2024)
3. European Parliament, “Directive on Copyright in the Digital Single Market” (2019)
4. UK Intellectual Property Office, “Consultation on Artificial Intelligence and Copyright” (2023)
5. U.S. Copyright Act, Section 107 (Fair Use)
6. Andersen v. Stability AI Ltd., Case No. 3:23-cv-00201 (N.D. Cal.)
7. Getty Images v. Stability AI Ltd., Case No. 1:23-cv-00135 (D. Del.)
8. RIAA v. Suno Inc., Case No. 1:24-cv-11611 (D. Mass.)
9. RIAA v. Uncharted Labs, Inc. (Udio), Case No. 1:24-cv-11612 (D. Mass.)
10. GEMA v. Suno, Munich Regional Court, Germany (July 2026)
11. Sony Music Entertainment v. Uncharted Labs, Inc. (Udio) (filed July 2026)
12. Bartz v. Anthropic PBC, Case No. 3:24-cv-05417 (N.D. Cal.) (settlement approved July 2026)
13. UMG Recordings, Inc. v. Suno, Inc. (II), filed Sept. 18, 2026, D. Mass. — Music Business Worldwide coverage
14. Deezer, “AI Music Tops 50% of Daily Uploads on Deezer” (July 21, 2026) — Deezer Newsroom
15. Sony Music Publishing LLC et al. v. Anthropic PBC et al., Case No. 5:26-cv-09217 (N.D. Cal., filed Aug. 28, 2026) — Music Business Worldwide coverage
16. UK Parliament, Communications and Digital Committee, “Committee publishes Government response to AI and copyright report” (15 May 2026) — committees.parliament.uk
17. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 25-2153 (3d Cir., judgment entered Sept. 29, 2026) — opinion
18. Justice v. Suno, Inc., No. 1:25-cv-11739 (D. Mass., memorandum and order on motion to dismiss, Aug. 20, 2026) — Loeb & Loeb case summary

Frequently Asked Questions

Is training AI on copyrighted music legal?

It remains contested. The Munich Regional Court ruled in GEMA's favor against Suno in July 2026, and in September 2026 the Third Circuit became the first U.S. appeals court to reject a fair use defense for AI training — though that case involved a non-generative AI tool. No U.S. court has yet ruled on fair use for AI music training, with dispositive motions in the U.S. label cases not expected before mid-2027.

What does the EU AI Act require for AI music?

The Act's Article 50 transparency obligations, enforceable since August 2, 2026, require providers to mark synthetic audio content as machine-generated and to publish summaries of copyrighted material used for training, subject to penalties of up to EUR 15 million or 3 percent of global turnover.

Can artists opt out of having their music used for AI training?

In the EU, the DSM Directive's text-and-data-mining provision allows rights holders to reserve (opt out of) machine-readable uses, and the AI Act requires compliance with such reservations. The U.S. has no statutory opt-out; whether unlicensed training is fair use is being litigated.