text and data mining exception (TDM)

advancedwidespread

A copyright rule that allows copies to be made from legally accessible works for machine analysis without permission; except for scientific research, the rights holder can prohibit this.

In plain terms

The main rule of copyright is that you may only copy someone else's work with permission. For machine analysis, however, text, images, or data must first be downloaded and stored, meaning a copy is made. The text and data mining exception permits this copying if the goal is the machine-based discovery of patterns, relationships, and trends.

It is like a library: anyone can enter, read, and take notes to learn from the books, but cannot photocopy them for sale. The exception allows analysis but not republication of the works.

In the European Union, there are two versions. Research institutions and cultural heritage organizations may mine anything they have lawful access to for scientific purposes, and copyright holders cannot prohibit this. Everyone else, including AI companies, may only use the exception if the copyright holder has not reserved their rights. This prohibition is called an opt-out.

In the AI era, this exception has become a key issue: the European debate on whether it is allowed to train language models on content collected from the web largely depends on the interpretation of this exception.

How it works

Legitimate access is required: only content to which the user has legally obtained access can be mined, such as freely available websites or subscribed databases. Bypassing paywalls is not permitted.
The machine makes a copy of the content and runs the analysis on it: counting words, searching for patterns, or training an AI model on it.
If you are not mining for scientific purposes as a research institution or public collection, you must first check whether the rights holder has reserved their rights. For online content, the rights holder must indicate this in a machine-readable way, for example in the website's robots.txt file or metadata. It is disputed whether a ban mentioned only in the terms of use is sufficient.
If such a ban exists, the content must be excluded, or permission (a license) must be requested. The ban does not apply to scientific research at research institutions and public collections.
Copies can be kept only as long as they are needed for analysis. For research purposes, they can be kept longer with secure storage, for example to verify results.

An everyday example

If you run a blog or webshop and do not want AI companies to train their models on your texts, you can block AI data-collecting bots in your website's robots.txt file and assert your rights in the terms of use. In the EU, AI development companies can no longer rely on the exception in such cases. Scientific research cannot be excluded this way. If you take no action, your publicly available content remains analyzable.

Who is affected

entrepreneur, executiveIf your company publishes content, you must decide and signal in a machine-readable format whether it may be used for AI training.

developerIf you collect data for training a model or operating an agent, you must pay attention to rights-holder notices and lawful access. → Code

parent, student, teacherAs a researcher at a university or research institute, you may analyze large text collections for scientific purposes even if the publisher prohibits it, provided you have lawful access. → Education & Family

lawyer, citizenIn the EU, this rule determines whether your writings and images shared online can be included in AI model training materials without permission. → Law

A common misconception

MythAnything posted on the internet can be freely used by AI companies, because the exception permits everything.

In realityThe exception applies only to copying legally accessible content for analysis purposes, and outside scientific research, the rights holder's prohibition overrides this. Furthermore, it does not allow the model to return the protected work or its recognizable parts.

Limits and risks

There is no single, universally accepted technical standard for reserving rights (opting out), so it can be disputed whether a given signal counts as a valid prohibition.
A ban only protects going forward: what has already been collected and trained on is hard to remove from a model afterwards.
The rule varies by country: in the USA the principle of fair use applies, in Japan a broader exception is in force, so the legal assessment of international services is uncertain.
Several key questions await judicial interpretation, such as whether training generative models always falls under the exception.

Similar, but not the same

fair use – US fair use is a flexible principle that courts weigh case by case, whereas the TDM exception is a precisely defined European rule set out in law.
web scraping – Web scraping is the technique for automatically collecting content, while the TDM exception is the legal rule specifying when this is permitted under copyright law.
licensing agreement – A license is a contractual permission from the rights holder, typically for a fee, whereas an exception allows use by law without permission or fee.

Related concepts

Broader

Narrower

opt-out (rights reservation), scientific text and data mining

On the edge

training data, web scraping, robots.txt, fair use, EU AI-rendelet

A brief history

The exception was introduced at EU level by the Directive on Copyright in the Digital Single Market (2019/790) in 2019, and Hungary transposed it into copyright law in 2021. The United Kingdom has had a similar rule for non-commercial research since 2014. The EU AI Act, adopted in 2024, requires providers of general-purpose AI models to respect rights reservations. This obligation has been applicable since August 2025.

Where things stand

The rule is in force in every EU member state, and the AI Act builds on it. The details, such as what counts as a valid machine-readable prohibition and how far the exception for generative AI extends, are still being shaped by courts and standardization negotiations.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Petition for Hungarian authors calls for labelling AI-written text and debate turns to the right to say noSep 1, 2026 · Jog

Also known as

TDM · TDM exemption · text and data mining · text mining exception · text and data mining · TDM exception · TDM opt-out · text and data mining exception

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.