AI Copyright and Training Data in 2026: Who Is Winning the Lawsuits, and What It Means for Your Company

For three years the AI industry trained first and asked later. In 2026 the bill arrived. Anthropic paid $1.5 billion to settle the authors' lawsuit over pirated books, OpenAI is a defendant in 34 cases, a US appeals court has for the first time rejected a fair use defence for AI training, and a court in Munich has twice ruled that AI models can infringe copyright simply by memorising songs. At the same time the EU AI Office started enforcing the AI Act's copyright rules on 2 August 2026. This guide brings the numbers together and explains what they mean for companies in Croatia and the DACH region that use Claude, ChatGPT, Mistral or Manus.

The short answer: training is often allowed, piracy and memorisation are not

Courts on both sides of the Atlantic are converging on a pattern that matters more than any single verdict. Using lawfully obtained works to train a model has so far mostly been tolerated: as fair use in the US, under the text and data mining (TDM) exception in the EU. What gets AI companies into trouble is how they obtained the data (pirate libraries, stream-ripping, ignoring opt-outs) and what the model gives back (memorised lyrics, articles or images that come out nearly verbatim). For businesses that use AI rather than build it, the practical risk sits almost entirely in the second category: outputs.

The litigation wave in numbers

The AI Lawsuit Tracker counted 145 AI copyright cases worldwide as of 8 October 2026, 31 of them outside the US. The pace keeps rising: 15 cases were filed in 2023, 28 in 2024, 50 in 2025 and already 49 in 2026 before the end of the third quarter. News publishers, book authors, music publishers, record labels, photo agencies and visual artists are all suing.

OpenAI is by far the most-sued company, named in 34 cases, more than twice as many as Meta or Microsoft (14 each). Anthropic follows with 13, music generator Suno with 9 and Google with 8. Most cases are still far from a verdict: 86 are active, 19 stayed or consolidated and 9 waiting for a ruling. Only three have settled, and one of them, Anthropic's, is the largest copyright settlement in US history.

United States: fair use for training, not for piracy

Bartz v. Anthropic: the $1.5 billion lesson

In June 2025 Judge William Alsup ruled that training Claude on books was fair use because it was highly transformative, but that Anthropic's downloading of millions of books from pirate sites such as Library Genesis and Pirate Library Mirror was not. Facing a jury trial over statutory damages, Anthropic settled for $1.5 billion, covering an estimated 500,000 works at roughly $3,000 per book. Judge Araceli Martinez-Olguin gave final approval on 20 July 2026. Because the case settled, Alsup's fair use ruling will never be reviewed on appeal and binds no other court.

In a parallel ruling in Kadrey v. Meta (June 2025), Judge Vince Chhabria also found fair use, but stressed that the authors had simply failed to prove market harm and that AI output flooding the market could weigh against fair use in future cases.

Thomson Reuters v. Ross: the first appeal goes against AI

On 29 September 2026 the US Court of Appeals for the Third Circuit became the first appellate court to decide fair use in AI training, and it ruled for Thomson Reuters. ROSS Intelligence had used more than 2,000 Westlaw headnotes to build a competing legal research tool; the court held the headnotes were protected and the commercial, competing use was not fair. The ruling has limits: ROSS was not a generative model, and the court expressly set generative AI aside. ROSS has said it will take the case to the Supreme Court. Still, the message to AI developers is clear: using someone's content to build a product that competes with them is the weakest position to defend.

OpenAI: litigate and license

The New York Times case against OpenAI and Microsoft, consolidated with authors' suits in a multidistrict litigation in New York, has not settled and is in the summary judgment phase. In parallel, OpenAI has signed more publisher licensing agreements than any other lab (24 publicly announced, according to one tracker), including News Corp (reportedly $250 million over five years). In December 2025 Disney agreed to invest $1 billion in OpenAI and licensed more than 200 Disney, Marvel, Pixar and Star Wars characters for Sora for three years. The strategy is clear: pay the largest rightsholders and fight the rest in court.

Anthropic: books settled, music still open

Anthropic's remaining exposure is in music. Universal Music Publishing, Concord and ABKCO have sued twice: in 2023 over about 500 song lyrics (potential damages around $75 million), and again in January 2026 over more than 20,000 works, seeking over $3 billion, after a judge refused to add piracy claims to the first case. The new complaint alleges that songbooks were among the pirated books and that newer Claude models were trained without permission.

Europe: Munich becomes the AI copyright capital

While US courts debate fair use, European courts apply narrower exceptions, and Germany is leading.

• GEMA v. OpenAI (Munich Regional Court, 11 November 2025). The court held that song lyrics memorised in a model's weights are reproduced in the model itself, that the TDM exception does not cover such memorisation and that the operator, not the user, is responsible for infringing outputs. OpenAI was ordered to stop, disclose usage and pay damages. The judgment is not final; OpenAI is appealing.
• GEMA v. Suno (Munich Regional Court, 31 July 2026). The same chamber rejected Suno's defence under Section 44b of the German Copyright Act, finding that memorising entire works goes beyond permitted analysis, and accepted evidence that Suno stream-ripped songs from YouTube by circumventing technical protection. Suno must stop using the works for training and disclose revenue; damages are still to be set. Not yet final.
• Getty Images v. Stability AI (UK High Court, 4 November 2025). The counterweight: after Getty dropped its main training claims, the court found that Stable Diffusion "does not store or reproduce" the copyrighted works and therefore is not an infringing copy. Getty won only a narrow trademark point over watermarks in some outputs.

The EU AI Act: copyright rules are now enforceable

Article 53 of the AI Act requires every provider of a general-purpose AI model placed on the EU market to maintain a copyright policy, respect machine-readable rights reservations (the TDM opt-out, for example via robots.txt) and publish a summary of the content used for training using the Commission's template. These duties have applied since 2 August 2025; since 2 August 2026 the AI Office can request information, demand model access and impose fines of up to 3% of global turnover or €15 million. Models already on the market before August 2025 have until 2 August 2027.

The voluntary GPAI Code of Practice turns this into concrete steps: crawl only lawfully accessible content, do not circumvent paywalls, exclude sites on an EU list of persistent infringers, honour robots.txt and other machine-readable opt-outs and offer a complaints channel for rightsholders. Anthropic, OpenAI and Mistral have signed it, as have Google and Microsoft; Meta refused. Politically, the pressure is rising: in March 2026 the European Parliament adopted a non-binding resolution calling for full transparency and fair remuneration and questioning whether the opt-out model works at all, and the Commission is consulting on standard opt-out protocols ahead of its review of the copyright directive.

Mistral and Manus

Mistral, as an EU company, is directly in the AI Office's sights and has positioned compliance as a selling point: it signed the Code of Practice and licenses news content, such as its AFP agreement that brings the agency's archive into Le Chat. Manus, now part of Meta, raises a different question. Autonomous agents browse, copy and assemble content on the user's behalf, so the risk shifts from training data to what the agent collects and what you publish. If an agent builds a market report from paywalled articles, the copyright question lands on the company that distributes the report.

The price of training data

The money now on the table shows how fast the market for training data is forming. Settlements and potential damages run into billions, while licensing deals give the largest rightsholders recurring income and, increasingly, visibility inside AI answers.

What this means for companies that use AI

Most companies in Croatia and the DACH region do not train foundation models. Their exposure is in outputs they publish and inputs they feed into AI. The good news: the major vendors take most of the training-data risk contractually, if you are on the right plan.

Free and consumer plans generally come without indemnity. Even on business plans, the protection usually falls away if you knowingly prompt for protected content, disable safety or citation features, or modify the output into an infringement. A practical checklist:

• Use business tiers for anything you publish or sell. Make indemnity and EU data processing part of your vendor selection.
• Do not ask for copies. Prompts like "write it in the exact words of" or "reproduce the lyrics of" are precisely what the Munich court and the indemnity exclusions target. Put that rule in your AI policy and address shadow AI, where no indemnity applies.
• Review before you publish. Run marketing texts, images and code through a human review and a similarity or plagiarism check. See our guide to AI in marketing and content.
• License what you feed in. Uploading paid studies, standards, newsletters or databases into a RAG system or a custom GPT is a reproduction too. Check your licences first.
• Do not expect to own pure AI output. The US Copyright Office (January 2025) and the D.C. Circuit in Thaler v. Perlmutter (March 2025) confirmed that works without human authorship are not protected, and EU law requires the author's own intellectual creation. Document the human contribution where ownership matters.
• Decide how AI may use your content. robots.txt lets you separate training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot) from search and user agents (OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User). Blocking the first while allowing the second protects your content without disappearing from AI answers; see our GEO guide.
• Keep an eye on the AI Act. If you fine-tune or substantially modify a general-purpose model and put it on the market, you may become a provider yourself. Our EU AI Act compliance guide explains where the line runs.

The bottom line

2026 is the year AI copyright moved from theory to invoices. The US has a $1.5 billion benchmark price for pirated training data and its first appellate defeat for fair use; Germany has twice ruled that memorising works in a model is reproduction; and the EU can now fine model providers that ignore opt-outs. For AI developers, the free-data era is over. For companies that use Claude, ChatGPT or Mistral, the takeaway is calmer but concrete: choose plans with indemnity, keep humans in the loop on what you publish, license what you feed in and set clear rules for your teams. Do that, and copyright is a manageable risk, not a reason to wait.