Cruz v. Anthropic: 28 Authors Who Opted Out of the $1.5B Settlement File New Copyright Suit Against Claude's Developer
On 13 May 2026 – one day before the final approval hearing in Bartz v. Anthropic, the record-breaking $1.5 billion class-action settlement – a group of 28 authors who had previously opted out of that settlement filed a separate lawsuit against Anthropic in the U.S. District Court for the Northern District of California
The plaintiffs are demanding a jury trial and allege that Anthropic illegally downloaded, reproduced, and distributed their works from pirated libraries (LibGen and similar repositories) to train the Claude family of large language models.
Key Facts
Cruz v. Anthropic (N.D. Cal.), filed 13 May 2026. The 28 authors opted out of the Bartz v. Anthropic claims before the deadline (early February 2026), thereby preserving their individual claims – which they have now exercised.
Legal grounds are the same as in Bartz: copyright infringement through mass downloading of books from pirated sources (LibGen, PiLiMi) and their use as the training corpus for Claude. Per Judge Alsup's June 2025 ruling, fair use doctrine does not cover the use of pirated copies.
Jury trial demand is a strategic choice. Juries have historically been harsher on large defendants in copyright cases. Statutory damages under U.S. law can reach up to $150,000 per willfully infringed work.
Who Is Most Affected
- M&A tech targets with proprietary foundation or fine-tuned models – especially those trained on 'open' internet datasets without source audits.
- Polish and EU SaaS / GenAI startups using open models (Claude, GPT, Llama, Mistral) via API – the training data indemnification question surfaces in every deal.
- Publishers, media groups, authors, and rights holders in the EU/Poland – the practical value of the text-and-data-mining (TDM) opt-out under Article 4 of the DSM Directive (2019/790) and machine-readable rights reservations has increased significantly.
- Venture / PE investors in AI model companies – the 'one settlement closes everything' rule no longer holds; opt-out classes require separate risk assessment.
| Corporate deployers of Claude and similar models in the EU – AI Act obligations under Article 53 (GPAI training content summary) increase pressure on vendors and developers. |
What This Means for Business
The Bartz settlement set a pricing benchmark; the new lawsuit demonstrates that every opt-out must be treated as a separate case with its own statutory damages exposure. For buyers of AI targets, this means that standard 'AI compliance' representations and indemnities from 2024 are no longer sufficient – dedicated provisions on training data provenance and open litigation/opt-out claims are now required.
For EU businesses using Claude / other LLMs via API a direct claim is unlikely, but indirect risks exist:
- Changes to vendor T&C and pricing – Anthropic has already revised its policies;
- Strengthened AI Act obligations – training content disclosure for GPAI providers (Article 53(1)(d)) applies from August 2025;
- Growing confidence among EU rights holders in asserting TDM opt-outs under Article 4 of the DSM Directive – this is shifting the negotiating position in licensing discussions.
| For Polish tech startups training their own models: use of any pirated or 'grey' corpora constitutes a direct risk – even if operations are based outside the U.S. – provided the model is distributed on the American market. |
Action list
M&A: in SPAs for AI/tech targets, we recommend updating the IP/AI representations block to include separate warranties on:
- Training data sources;
- Absence of pirated corpora (LibGen, PiLiMi, Anna's Archive, etc.);
- List of opt-out reservations received from rights holders;
- Current list of pending/threatened copyright claims.
Consider expanding special indemnities to cover statutory damages and litigation costs.
Due Diligence: request a data sheet / model card / training data audit; verify the existence of publisher licensing agreements and corpus audit reports; assess the geographic distribution of the model.
Vendor contracts with AI providers (Anthropic, OpenAI, Google, Mistral, etc.): review IP indemnification – caps, 'training data' carve-outs, notice/defence procedure, and termination rights in the event of a material copyright claim against the vendor. Many standard templates currently exclude training data claims.
For EU and Polish rights holders: implement TDM opt-out in accordance with Article 4 of DSM Directive 2019/790 – robots.txt, TDM Reservation Protocol, explicit ToS clauses. Without a machine-readable reservation, the TDM exception applies by default.
GPAI deployers under the AI Act: verify that your vendor has provided a 'sufficiently detailed summary' of training content (Article 53(1)(d) AI Act) – this document is becoming a key piece of evidence in due diligence and in responding to rights-holder enquiries.
| REVERA keeps tracking the outcome of the Bartz hearing (postponed) and the progress of Cruz v. Anthropic – the first U.S. jury trial on AI training data will set both the pricing and doctrinal benchmark for the EU. |
Write to our lawyer to learn more details
Contact a lawyer