Copying, analysing, sharing: where does the TDM exception draw the line?

AI is now woven into the way we get our information every day. Like it or not, we all use it, directly or indirectly: from news stories in your daily apps written with the help of generative AI to the smart thermostat that keeps your living room warm as efficiently as possible.

How AI learns, and why that rubs copyright the wrong way

The AI models behind all this don’t appear out of thin air. They have to be trained on existing information. By finding patterns in that information, a model learns to make predictions about new information. Language models, such as ChatGPT, Claude, Gemini and Microsoft Copilot, predict a string of words in response to your prompt or question. Classification models, by contrast, sort a piece of text or an image into categories they have been trained on in advance, think of the filters on platforms such as YouTube or Instagram that automatically detect whether a music clip or image contains copyrighted material.

Training these models takes computers, or rather, extremely powerful servers, that work through all that information in search of (statistical) correlations. And anything you do on a computer involves making copies of the source data, often many times over: think of data stored on a hard disk or SSD, or held temporarily in RAM or the CPU cache. Those copies matter in particular when models are trained on copyrighted material, or ‘works’, as the Dutch Copyright Act calls them, such as text, music or (moving) images. And because making copies of a copyrighted work is something the author or rightsholder has the exclusive right to prohibit, this can quickly become a problem.

The TDM exception: are computers allowed to read as well?

The EU legislature recognised this problem around ten years ago. Nobody had heard of generative AI yet, but ‘big data’ and ‘text and data mining’ (TDM) were hot topics. TDM mainly referred to the techniques of the day for trawling through large volumes of raw data or text in search of patterns or relationships, or for analysing their meaning or sentiment. With access to vast amounts of data and increasingly advanced machine learning, new insights and knowledge were suddenly within reach. Copyright, however, stood in the way: being able to access data did not automatically mean you were allowed to ‘mine’ it.

Under slogans such as ‘the right to read is the right to mine’, the call for a copyright exception for TDM grew ever louder. The EU legislature eventually responded with a directive that, subject to certain conditions, exempts copies made for text and data mining from the rightsholder’s exclusive right to prohibit reproduction. In a nutshell: computers, like humans, may read protected works and extract insights, analyses and patterns from them, provided there is lawful access to the source data. And, just as for humans, simply parroting, reusing or distributing that source data is not automatically allowed. One small caveat: TDM is also off-limits where the rightsholder has expressly reserved that right, unless you qualify for the separate exception for research organisations and cultural heritage institutions.

The scope of the exception has been hotly debated, including in court. Courts at various levels in other EU member states have already ruled on it. In the Netherlands, too, the TDM exception came up in a dispute between news publishers and HowardsHome.

Publishers v. HowardsHome

Since 2001, HowardsHome had been running a digital news-monitoring service for around sixty business clients. Each day it processed some 30,000 items from public RSS feeds, each consisting of a hyperlink, the headline, a snippet of up to 150 characters and, where available, a thumbnail. It had never obtained explicit permission from publishers DPG, Mediahuis and NRC, who argued that HowardsHome was therefore infringing their press publishers’ right, copyright and database right. The publishers themselves serve that same business market through licensing deals with LexisNexis and ArtikelPro. In 2024, the Amsterdam District Court dismissed all their claims. Strikingly, one of the defences HowardsHome successfully relied on was the TDM exception.

The Amsterdam Court of Appeal has now overturned that judgment and ruled that HowardsHome infringed both copyright and the press publishers’ right. It rejected the TDM defence on two grounds. First, the exception only covers reproduction, making (digital) copies, and not communication to the public, such as publishing content on a platform or emailing it to clients. Second, selecting news items and forwarding them to clients does not qualify as an analytical technique within the meaning of Article 25a(3) of the Dutch Copyright Act (the Dutch TDM exception).

Read and analyse, but don’t regurgitate

The ruling confirms the principle outlined above: you may read and analyse copyrighted texts, but reproducing them verbatim and passing them on requires the rightsholder’s permission (or another statutory exception). The TDM exception extends the reading-and-analysing part to computers, under certain conditions, but it ends where further copying and distribution begin. That was exactly what the case between the publishers and HowardsHome was about, so it is hardly surprising that the TDM defence failed.

Meanwhile, the Court of Justice of the European Union, the highest authority on the interpretation of EU law, is also looking into the scope of the TDM exception. In Like Company v Google (C-250/25), a Hungarian court has asked the Court of Justice for guidance in a dispute over the Gemini chatbot. Like Company, a Hungarian news publisher, claims that Google used its articles without permission, both to train Gemini and to generate chatbot answers in which the content of those articles resurfaces. Among other things, the Court will therefore have to decide whether training a language model is covered by the TDM exception.

The Court has not ruled yet. First, an Advocate General, an independent adviser to the Court, delivers a reasoned, non-binding opinion, which the Court often (though not always) follows. Advocate General Szpunar delivered his on 3 November 2026. Although he phrased it slightly differently, he endorsed the same principle: protected content may be read and analysed, but may not resurface verbatim at a later stage without permission. In his view, it also matters how easily a model can reproduce (parts of) its training data, whether by design or through carelessness in the development process. He therefore considers that AI developers can be expected to put appropriate safeguards in place to stop language models from reproducing protected content verbatim in their output.

We will have to wait for the Court’s final judgment, but the contours of the TDM exception are gradually becoming clearer. One thing already seems certain: AI developers, and the businesses and individuals who use AI, need to be careful about reusing, distributing or publishing AI output.

Would you like to learn more about intellectual property and its implications for your organization?