Dutch Court Confirms Narrow Interpretation of TDM Exception for AI Training

In its appeal ruling on October 20, 2026, the court of Amsterdam (Gerechtshof Amsterdam) strictly limits the scope of the Text and Data Mining (TDM) exception under EU law, signaling heightened copyright enforcement for digital reproduction and AI training. The court ruled that HowardsHome infringed the copyright and press publisher rights of Dutch media companies DPG Media, Mediahuis, and NRC. HowardsHome operated a commercial digital news monitoring service that collected information from RSS feeds. It reproduced article headlines, short snippets, and thumbnails without a license.

While the lower court had previously accepted HowardsHome’s TDM defense, the Gerechtshof Amsterdam overturned this decision. The appeal court emphasized that the TDM exception of the Dutch Copyright Act applies strictly to reproduction, not to the communication to the public (Dutch “openbaarmaking”) of copyrighted works. Furthermore, the court clarified that HowardsHome’s news curation did not meet the statutory definition of TDM.

What is the TDM exception?

The TDM exception is a legal rule that allows computers to copy and analyze large amounts of copyrighted material to find patterns, trends, and connections without breaking copyright laws. Think of it as giving a software program permission to read millions of books, articles, or images to learn how language works or to spot scientific trends. Normally, making those digital copies during the scanning process would be illegal. The TDM exception acts as a safe harbor. It recognizes that the software is extracting underlying facts and statistics rather than stealing the creative expression to republish it.

However, this legal protection comes with strict boundaries to protect original creators. First, the user must have lawful access to the original content, such as through a paid subscription or a publicly accessible website. Second, in regions like the European Union, rightsholders can “opt out” of commercial data mining by placing a machine-readable “do not scrape” code on their websites to block AI companies from harvesting their work. For example through robots.txt instructions. Finally, the exception strictly covers the backend: automated analysis of the data. It does not give a company permission to share, display, or sell those copied texts or images to the public. And that’s exactly what happened in the HowardHome’s infringement.

EU Context and AI Lawsuits

This narrow interpretation of the TDM exception by Gerechtshof Amsterdam mirrors a restrictive European judicial trend directly affecting generative AI developers. In the EU, AI companies have relied on Article 4 of the Digital Single Market Directive to justify web scraping for model training. However, recent litigation demonstrates that EU courts treat copyright as a strict economic right. In the Munich Regional Court rulings of GEMA v. OpenAI (November 2025) and GEMA v. Suno (July 2026), courts held that training AI on copyrighted song lyrics and music without a license constituted unlawful reproduction. These rulings effectively rejected broad TDM safe harbors for generative AI. The recent Dutch appeal court’s refusal to stretch the TDM exception beyond its technical definition reinforces the legal vulnerability of unconsented data scraping across the EU.

The US Approach: Fair Use Over TDM

Conversely, US AI litigation centers on the fair use doctrine rather than a statutory TDM exception. The legal focus remains on whether AI training is transformative and whether the outputs harm the original market.

Anthropic agreed to a historic $1.5 billion settlement in August 2025 in the Bartz v. Anthropic class action after authors proved the company’s Claude model was trained on pirated books. Meanwhile, major media publishers are aggressively pursuing OpenAI and Microsoft in the US. For example, The New York Times seeks damages for copyright infringement based on chatbot outputs that mimic its articles.

While the US framework offers more flexibility for technological innovation through fair use, the EU approach, as cemented by the Amsterdam appeal ruling, demands strict compliance with narrow statutory exceptions and explicit licensing agreements. Nevertheless, the courts in both economic blocs seem to converge on holding AIs accountable for the economical damage they cause to copyright owners.