AI Training and Copyright: Why the Origin of Training Data Determines the Liability Risk
Bartz v. Anthropic: The USD 1.5 billion settlement and its lessons for AI providers and companies in Germany
In Bartz v. Anthropic, a US federal court has given final approval to a USD 1.5 billion settlement – according to available reports, the largest settlement reached to date in a copyright case. The case matters in practice above all because of one distinction: the court assessed the training of an AI model on copyrighted works separately from the acquisition and permanent storage of the training data. Training was held to be permissible (“fair use”); acquisition from piracy sources was not. This article puts the decision in context and shows what it means for companies that develop, train, deploy or acquire AI models.
1. The case: what the court decided and what the parties settled
The case was heard before the United States District Court for the Northern District of California (No. 3:24-cv-05417). The plaintiffs were authors acting on behalf of a group of rights holders. In a partial summary judgment order of 23 June 2025 (Judge Alsup), the court assessed three uses separately:
Training on the books: permissible as fair use under 17 U.S.C. § 107; the court considered the use “exceedingly transformative”.
Scanning lawfully purchased print books: also permissible as fair use, as it merely changed the format.
Building a permanent central library from pirated copies: not covered by fair use. According to the court, a later intended use does not justify downloading works from piracy sources and retaining them permanently where they could have been purchased lawfully. Summary judgment was denied in this respect; infringement and damages were to be determined at a jury trial.
The trial did not take place. The parties settled, and the court gave final approval on 20 July 2026 (Judge Martínez-Olguín). According to available reports, the settlement comprises:
a non-reversionary fund of at least USD 1.5 billion plus interest, payable in four instalments;
approximately 482,000 listed works at roughly USD 3,000 per work before costs and fees (authors and publishers generally split equally under the default allocation);
a release limited to past conduct relating to the acquisition and copying of the listed works; claims relating to model outputs and future conduct are not covered;
an obligation to destroy the original files downloaded from the piracy sources and copies made from them.
Two qualifications are essential. First, the settlement was reached without a finding of liability; its amount reflects litigation risk, not a final judgment on infringement and damages. Second, the order of 23 June 2025, as a district court decision, does not bind other courts. According to available reports, appeals to the Ninth Circuit are pending against parts of the settlement implementation, in particular the fee order.
2. Why the amount is so high: statutory damages per work
US copyright law allows a rights holder to claim, instead of proven damages, a statutory amount per infringed work (“statutory damages”, 17 U.S.C. § 504(c)). The range is generally USD 750 to 30,000 per work, and up to USD 150,000 for willful infringement. For a library of several hundred thousand works, this method produces exposure that can threaten a company’s existence. This explains why even a provider that wins the central dispute over training carries substantial residual risk in respect of how the data was acquired.
3. Transferability to German law
German copyright law has no general “fair use” clause; it provides exhaustively defined limitations and exceptions (Sections 44a et seq. of the Copyright Act, UrhG). For AI training, Section 44b UrhG (text and data mining), which implements Article 4 of Directive (EU) 2019/790, is the key provision. Under Section 44b(2) sentence 1 UrhG, reproductions of “lawfully accessible works” are permitted for text and data mining; under sentence 2, they must be deleted once no longer required for that purpose. For works made available online, a rights reservation is effective only if declared in machine-readable form (Section 44b(3) UrhG). Section 60d UrhG contains a separate exception for scientific research organisations.
Lawfully accessible: The parallel to Bartz lies in the requirement of “lawful access”. It seems plausible that works obtained from sources that themselves make them available unlawfully are not “lawfully accessible”. To our knowledge, no supreme court ruling has settled this question; it must be treated as open.
Case law: The scope of Section 44b UrhG in AI training is currently occupied German courts intensively:
Regional Court of Hamburg, judgment of 27 September 2024 (Kneschke v. LAION), and subsequently Higher Regional Court of Hamburg, judgment of 10 December 2025 – 5 U 104/24: A photographer’s claim was dismissed. According to available reports, the Higher Regional Court held that creating the dataset was covered by the text and data mining exceptions and that a rights reservation expressed only in natural language does not meet the machine-readability requirement; leave to appeal on points of law was granted.
Federal Court of Justice (BGH), I ZR 281/25: According to available reports, the First Civil Senate held an oral hearing on 3 September 2026. No decision was issued; the Senate signalled a need for clarification under EU law, so a referral to the CJEU is possible. Whether and on which questions a referral will be made was reported as open.
Regional Court of Munich I, judgment of 11 November 2025 – 42 O 14139/24 (GEMA v. OpenAI): According to available reports, the court regarded the memorisation of song lyrics in the model as a reproduction (Section 16 UrhG) not covered by Section 44b UrhG, and found infringement also through the chatbot’s outputs; responsibility lay with the provider. According to the reports, the judgment was not final.
The consequence: while the US case separates training from acquisition, the Munich court additionally asks whether the model itself contains and reproduces works. Both lines are relevant for providers and complement each other: sound compliance must cover both the origin of the data and the output behaviour of the model.
Legal consequences: German law knows neither punitive damages nor statutory damages per work as in US law. However, the rights holder may claim injunctive relief and removal (Section 97(1) UrhG), damages including calculation by way of a licence analogy (Section 97(2) UrhG), information (Section 101 UrhG) and destruction and recall (Section 98 UrhG). Intentional unauthorised exploitation is also a criminal offence under Section 106 UrhG. For a model trained on unlawfully obtained data, the claims for injunctive relief and destruction may weigh more heavily in economic terms than the claim for payment.
Regulation: Under Article 53(1)(c) of Regulation (EU) 2024/1689 (AI Act), providers of general-purpose AI models must maintain a policy to comply with Union copyright law, in particular to identify and respect rights reservations under Article 4(3) of Directive (EU) 2019/790, and under point (d) must publish a sufficiently detailed summary of the content used for training. Documenting data origin is therefore not merely a risk topic but a legal obligation.
4. Significance for transactions and contracts
In acquisitions of and investments in AI companies, the origin of training data becomes a separate subject of due diligence. The risk is not confined to the past: if there is a suspicion that a model was trained on unlawfully obtained data, claims for injunctive relief and destruction may affect the value of the target’s core asset. Buyers should reflect this in warranties, indemnities and, where appropriate, purchase price mechanisms; providers of AI services should indemnify their customers against third-party claims only to an extent that their own data provenance can substantiate.
5. Practical checklist
Document the origin of training data (source, time, legal basis for use, licence or exception).
Exclude piracy sources (“shadow libraries”) and datasets of unclear origin; review existing datasets for such sources.
Technically detect and observe machine-readable rights reservations under Section 44b(3) UrhG and log the procedure.
Implement a deletion concept for reproductions no longer required for text and data mining (Section 44b(2) sentence 2 UrhG).
Test memorisation and output behaviour of the model and provide safeguards against reproducing protected content.
Consider licensing content where the exceptions do not clearly apply.
Implement the obligations under Article 53 of the AI Act (copyright policy, training data summary).
Negotiate warranties, indemnities and liability caps on data provenance carefully in contracts and transactions.
Monitor legal developments, in particular the pending BGH decision (I ZR 281/25) and a possible referral to the CJEU.
Conclusion
Bartz v. Anthropic shows that liability risk in AI training is not determined solely by whether the training itself is permissible. How the training data was acquired and retained is equally decisive. This finding cannot be transferred directly to German law, which has no general “fair use” clause but applies the narrow requirements of Section 44b UrhG, and where key legal questions remain open pending a decision by the BGH or the CJEU. Anyone who develops, deploys or acquires AI models should therefore be able to substantiate data provenance. Please contact us if you have any questions.
As of: 17.11.2026
NIETZER Rechtsanwälte · Attorneys at Law (USA) · Heilbronn · www.unternehmensrecht.com
This article is for general information purposes only and does not replace legal advice in an individual case.

