Meta won a case over training its LLaMA model on books downloaded from pirate libraries. A federal court in California held that the use was fair use. The internet reacted enthusiastically: “AI can learn from anything!”
But this ruling does not protect you at all if you operate in the EU.
There is no fair use doctrine in European law. Different rules apply: a closed list of exceptions, an opt-out mechanism and the AI Act, which expressly imposes copyright obligations on AI model providers. Kadrey v. Meta is a decision of one federal court in the United States, based on rules that do not apply in Poland or the EU.
This article explains what the California court actually decided, why Meta won and what would have to change for it to lose, how the law works in the EU and Poland, and what to consider if your business builds, deploys or uses AI models trained on third-party content.
Key terms before we begin
Before looking at the decision and its implications, it helps to clarify a few terms used throughout this article.
| Term | Explanation |
|---|---|
| Fair use | A US legal doctrine that lets a court assess flexibly whether using another person’s work without permission is permissible. Courts consider four factors: the purpose of the use, the nature of the work, the amount used and the effect on the market. It does not exist in EU law. |
| TDM (text and data mining) | Automated analysis of digital text and data to identify patterns, trends and correlations. Training AI models falls within this definition. |
| Opt-out | A mechanism under Article 4(3) of the DSM Directive. A rights holder may expressly reserve the right to prevent commercial TDM of their content. For online content, the reservation must be in a machine-readable format, such as metadata or HTTP headers. |
| GPAI (general-purpose AI model) | An AI model capable of performing many different tasks, such as generating text, images or code. Examples include GPT, LLaMA, Claude and Mistral. The AI Act imposes specific obligations on GPAI providers. |
| DSM Directive 2019/790 | The Directive of the European Parliament and of the Council on copyright and related rights in the Digital Single Market. Articles 3 and 4 establish TDM exceptions for research and commercial purposes, respectively. |
| Regurgitation | When an AI model reproduces parts of its training works in a form close to the original, word for word or with only minor changes. |
What the court decided in Kadrey v. Meta
The case was heard in the United States District Court for the Northern District of California. The plaintiffs, writers Richard Kadrey, Christopher Golden and other authors in a class action, alleged that Meta used their books without permission or payment to train LLaMA.
Meta admitted using Books3, a collection of approximately 195,000 e-books downloaded from the pirate site Bibliotik. During the proceedings, internal Meta communications came to light, including concerns about “torrenting on a company laptop” and alleged authorization from Mark Zuckerberg to use pirated materials.
Nevertheless, in late June 2025 the court held that training LLaMA on the plaintiffs’ books was fair use. Its reasoning rested on three points:
- Transformative use. The court found that AI does not “read” books for their plots. It learns statistical language patterns. This is a different purpose from the original work, making the use “transformative” under fair use.
- Copying entire works was considered reasonable. The fact that Meta copied entire books did not itself establish infringement. The court found full copying necessary for effective model training.
- No evidence of market harm. This was decisive. The authors presented no market analysis, data showing a decline in sales, or evidence that LLaMA generated content competing with their books. Meta, in turn, presented its own data indicating no effect on the plaintiffs’ sales.
Why Meta won, and what would have to change
Meta did not win because training AI on pirated data is legal. It won because the authors did not provide sufficient evidence.
The court expressly said its decision was fact-specific and concerned the circumstances of this case. It identified “market flooding” as the plaintiffs’ most promising, but unproven, argument: AI models may generate millions of works competing with originals, potentially reducing creators’ sales and incentives.
The court’s message is clear: if authors better document actual market harm in the future, the outcome could be different.
Other cases shaping the developing line of decisions
Kadrey v. Meta is not an isolated case. Other proceedings are underway, and their outcomes may shape market standards in the coming years.
| Case | Jurisdiction | Status or outcome | What it changes |
|---|---|---|---|
| NYT v. OpenAI | United States (S.D.N.Y.) | Pending; 13 May 2025 order preserving evidence, including ChatGPT chat histories | Shows that courts treat a model’s ability to reproduce content as a serious allegation. The judge prioritized the NYT claims over privacy rights, creating a conflict with the GDPR. |
| Bartz v. Anthropic | United States (N.D. Cal.) | Decision: June 2025 | Training on lawfully acquired data was fair use, but mass downloading from pirate libraries was “ordinary piracy.” The key distinction is that the legality of the data source matters. |
| Getty v. Stability AI | United Kingdom (High Court) | Decision: 4 November 2025 | The model weights were not an “infringing copy”; approximately 220 TB of training data were compressed into a file of about 3.44 GB. The court did not decide whether mass downloading the data was lawful. |
Each case confirms one point: the line of decisions is still developing. There is no settled standard, and outcomes may differ substantially depending on the evidence, jurisdiction and specific circumstances.
Why the Meta ruling does not protect you in the EU
If your business operates in Poland or elsewhere in the EU, Kadrey v. Meta has no direct legal effect for you. There are three reasons.
There is no fair use doctrine in European law
The EU has no open-ended general clause that lets a court assess flexibly whether using another person’s work is permissible. Instead, it has a closed list of exceptions and limitations. Member States cannot create new exceptions outside the list in the directives. A new form of use requires legislative action at EU level.
The boundaries of the exceptions are set by the three-step test in Article 5(5) of the InfoSoc Directive 2001/29/EC: the use must be a special case, must not replace normal exploitation of the work, and must not unreasonably prejudice the creator’s legitimate interests. The CJEU consistently rejects broad interpretations of exceptions.
TDM exceptions follow different rules from fair use
The DSM Directive 2019/790 introduces two exceptions for text and data mining, implemented in Polish law by the Act of 26 July 2024:
- Article 26² of the Polish Copyright Act, corresponding to Article 3 of the DSM Directive: TDM for scientific research. It applies only to cultural heritage institutions and entities in the higher education system. The use must not be for financial gain. Rights holders cannot contractually exclude this exception.
- Article 26³ of the Polish Copyright Act, corresponding to Article 4 of the DSM Directive: TDM for other purposes, including commercial purposes. It applies to all entities, provided they have lawful access to the content and there is no effective opt-out.
The opt-out mechanism is a fundamental difference from fair use. A rights holder may expressly reserve the right to prevent commercial TDM, and that reservation is binding. For online content, it must be in a machine-readable format, such as metadata, HTTP headers or a robots.txt file. ZAiKS used this mechanism on 10 December 2024.
Importantly, using pirated databases such as Books3 does not meet the lawful-access condition required by the TDM exceptions. Training on such data cannot rely on the EU copyright exceptions, regardless of what a court in California decided.
The AI Act imposes additional obligations
The AI Act (2024/1689), applicable from 1 August 2024, expressly addresses copyright in the context of AI model training. GPAI providers must:
- Have a copyright-compliance policy, including identifying and respecting opt-outs under Article 4(3) of the DSM Directive.
- Publish summaries of training data from August 2025, with enough detail to let copyright holders assess and pursue claims. The European AI Office is developing a template.
- Apply these obligations regardless of where training takes place. No provider may gain an advantage in the EU market “by applying lower copyright standards than those applicable in the Union” (Recital 106 of the AI Act).
Penalties for breaches of GPAI obligations may reach EUR 7.5 million or 1.5% of global annual turnover. For prohibited practices, penalties may reach EUR 35 million or 7% of turnover.
Contact us to discuss whether your AI product meets EU requirements.
Who is responsible for copyright infringement: the model provider or the business deploying it?
This is one of the most common questions we hear from founders and CTOs. The answer depends on your company’s role in the AI value chain.
Roles and responsibilities under the AI Act
| Role | Who it is | Copyright responsibilities |
|---|---|---|
| GPAI provider | An entity developing a model and placing it on the EU market, such as Meta, OpenAI or Anthropic | Lawfulness of training data, copyright-compliance policy, respecting opt-outs and publishing training-data summaries |
| Deployer | A business using an existing model in its product, such as a startup integrating LLaMA | How AI is used, lawfulness of generated content and compliance of the deployment with the AI Act and GDPR |
| Distributor or importer | An entity making an AI system available further down the chain | Conformity of labeling and documentation |
What you do not need to worry about
If you obtain LLaMA from an official source under Meta’s license, you are not responsible for copyright infringements committed by Meta during the original training. The end user is not “obtaining pirated software”; they are using a lawfully licensed model.
Where your risks lie
You are responsible for:
- Your own copyright infringements involving generated content. If a model reproduces part of someone else’s work and you publish it, your company may be liable.
- Unlawful data used for fine-tuning. If you fine-tune a model on other people’s materials without a license, you may directly infringe their rights.
- Compliance of your deployment with the AI Act, GDPR and other regulations.
- Moving from deployer to provider. This may happen if you brand the model as your own, substantially modify its operation or change its intended purpose. An indicative criterion is training compute exceeding one-third of the original training compute. In that case, provider obligations may apply, including responsibility for the lawfulness of training data.
There is an additional risk: a court may in the future prohibit the use of a particular model version if it finds an infringement that cannot be remedied. You cannot control that scenario, but you can prepare for it.
What you should do: practical steps
The actions below are ordered by priority. Not all apply to every business. It depends on whether you build your own models, integrate existing ones or use AI tools only to generate content.
Required actions
- Audit the origin of training data. Identify which AI models your business uses or builds. Check whether their training sets may contain copyrighted works obtained without a license. If you use open-source models such as LLaMA or Mistral, review the license terms and known training-data sources.
- Add indemnity clauses to agreements with AI model providers. Check whether the provider accepts the risk of copyright claims arising from training data and whether the license addresses liability for IP infringement.
- Check whether your AI uses are subject to EU regulations. The DSM Directive, AI Act and GDPR impose different obligations. The fair use doctrine in Kadrey v. Meta does not determine your legal position in Poland or the EU.
Recommended actions
- Prepare an internal policy for generative AI. Specify what content may be generated, for what purposes and how it must be checked before publication. This is particularly important in e-commerce, including product descriptions, and content marketing, including articles and graphics.
- Monitor legal decisions. The outcomes of NYT v. OpenAI, Bartz v. Anthropic and Getty v. Stability AI may shape market standards. Assign someone to track developments and share recommendations.
- Build AI literacy across your team. Since 2 February 2025, the AI Act has required providers and deployers to ensure staff have appropriate AI skills, including awareness of copyright risks.
Actions worth considering
- Consider moving to models trained on licensed data. Alternatives include PLLuM (Apache 2.0 license for commercial use), Bielik.AI, Google’s Gemma models and Microsoft’s Phi models (MIT license). This may reduce legal risk and strengthen your negotiating position with customers and investors.
The “market flooding” theory: why this argument may return
In Kadrey v. Meta, the court called market flooding the authors’ most promising argument, but found it unproven. It is worth understanding why the theory may return in future cases.
The mechanism is straightforward: AI models can generate millions of works that compete with originals. If the market is flooded with synthetic content, creators’ sales and earnings may fall, even if no individual AI output is a “copy” of a particular work.
The AI Act expressly identifies this risk. Recital 133 states that “different AI systems can generate large amounts of synthetic content that is becoming increasingly difficult to distinguish from human-generated content,” creating “new risks of misleading information and manipulation on a large scale.”
Some reported data points have raised concerns:
- According to the South China Morning Post, about 38% of submissions on Amazon are AI-generated publications.
- A survey of about 400 writers in the UK found that 39% already felt their income had fallen, 85% expected a further decline and 59% said their work had been used without permission.
- There have been cases of AI-generated pseudo-books, including fake guides about King Charles III and books impersonating author Jane Friedman.
For your business, this means that if you generate AI content commercially, such as product descriptions, articles or marketing materials, you could become a co-defendant in future market-flooding cases. The risk may increase with the scale of generation.
Check whether your AI product is legally safe
Kadrey v. Meta does not close the issue; it opens it. The line of decisions is still developing, and European rules, including the DSM Directive and AI Act, impose obligations with no equivalent in US law.
If your business builds, integrates or deploys AI models, check three things:
- Where the training data in the models you use comes from.
- Whether your agreements with AI model providers protect you against IP claims.
- Whether your deployment meets the requirements of the AI Act, DSM Directive and GDPR.
We help technology businesses address these areas, from auditing AI data pipelines and reviewing model-provider agreements to preparing internal AI policies. Contact us to arrange an IP review of your AI product.
Summary
The decision in Kadrey v. Meta does not legalize training AI on other people’s works. It confirms only that, in the specific circumstances and without evidence of market harm, a US court may find such training to be fair use. For businesses operating in the EU, the decision changes nothing: the DSM Directive’s TDM exceptions and opt-out mechanism, the AI Act obligations and the closed list of copyright exceptions continue to apply.
Three points to remember:
- Fair use does not exist in EU law. Do not rely on the Meta decision as a defense in Europe.
- You are responsible for generated content and for the lawfulness of data you use to fine-tune a model yourself.
- Regulation is becoming stricter. The sooner you address IP issues in your AI product, the lower your exposure may be.
Frequently asked questions
Can I safely use LLaMA-type models in a commercial product because a court found training lawful?
The decision concerns US law and the specific circumstances of Kadrey v. Meta only. If you operate in the EU, the DSM Directive’s TDM exceptions and the AI Act apply, not the fair use doctrine. Using LLaMA under Meta’s official license is permitted, but you remain responsible for how you use the model, its generated content and compliance of your deployment with EU rules.
Who is liable for copyright infringement: the AI model provider or the business deploying it?
The model provider, such as Meta, is responsible for the lawfulness of training data and its copyright policy. The business deploying the model is responsible for generated content, the lawfulness of data used in fine-tuning and compliance with the AI Act. If you substantially modify the model or change its purpose, you may move from deployer to provider, significantly expanding your obligations.
Can AI-generated content, such as product descriptions, expose my business to a lawsuit?
Yes. There are two risks: (1) a model may reproduce parts of someone else’s work, and publishing them may make your business liable for infringement; and (2) mass generation of content competing with originals could be alleged to constitute market flooding, a theory the court in Kadrey v. Meta considered promising but unproven. Establish procedures for checking content before publication.
What clauses should I include in an agreement with an AI model provider?
Consider an indemnity clause under which the provider accepts the risk of copyright claims arising from training data. Also review the provider’s statements about lawful data sources, its commitment to respecting opt-outs, its duty to inform you about changes to training datasets and its process for handling third-party claims.
Should I switch to models trained exclusively on licensed data?
That depends on your business’s risk profile. Models with more transparent training-data policies include PLLuM (Apache 2.0), Bielik.AI, Gemma (Google) and Phi (Microsoft, MIT). A switch may reduce legal risk and strengthen your position with customers and investors, particularly if you plan a funding round involving IP due diligence.
How does the situation in the US differ from Poland? Does the Meta decision protect me in Europe?
No. In the US, courts assess fair use flexibly in each case. The EU has a closed list of exceptions, including the DSM Directive’s TDM exceptions implemented in Polish law in Articles 26² and 26³ of the Copyright Act. Rights holders can block commercial TDM through an opt-out. The AI Act also requires GPAI providers to respect these reservations, regardless of where model training takes place. The argument “a US court found it lawful” is not a defense before a Polish court or supervisory authority.
