A TDM opt-out is the reservation of rights by which a copyright holder excludes a work from the general text and data mining exception in Article 4 of the EU DSM Copyright Directive, so that copying the work for automated analysis, including training of AI models, again requires a license. In the field of Copyright & AI, it is the main lever that rightholders in the European Union have over the commercial use of their published content as training data.
The opt-out matters because the default is permissive. Without a valid reservation, anyone with lawful access to a work may reproduce it for mining. With a reservation expressed in an appropriate manner, which for content online means machine-readable means, the exception no longer applies to that work.
How the TDM opt-out works in EU law
The DSM Copyright Directive contains two separate mining exceptions, and only one of them can be switched off. Article 3 covers research organizations and cultural heritage institutions mining for scientific research and cannot be overridden by contract or reservation. Article 4 covers everyone else, including commercial AI developers, and is conditional on the absence of an express reservation.
| Feature | Article 3 exception | Article 4 exception |
|---|---|---|
| Beneficiaries | Research organizations, heritage institutions | Any person or company |
| Purpose | Scientific research only | Any purpose, including commercial |
| Can be opted out | No | Yes, by express reservation |
| Retention of copies | With appropriate security, for research | As long as necessary for mining |
| Lawful access required | Yes | Yes |
The AI Act connects to this mechanism. Providers of general-purpose AI models must adopt a policy to comply with EU copyright law and, in particular, to identify and respect reservations made under Article 4(3), using state-of-the-art technologies. The voluntary code of practice for these providers adds concrete commitments, such as following robots.txt instructions and not circumventing paywalls.
Technical ways to express a reservation
No single format is mandated, so rightholders usually combine several signals. The legal test is whether a crawler operating at scale can reliably detect the reservation without human reading.
| Signal | Where it lives | Strength | Limitation |
|---|---|---|---|
| robots.txt rules | Site root, per user agent | Widely honored by major crawlers | Blocks crawling, not later reuse of copies |
| TDM Reservation Protocol | HTTP header or well-known JSON file | Expresses the legal reservation directly | Adoption still uneven |
| HTML meta tags | Page head | Granular per page | No single agreed vocabulary |
| Embedded file metadata | Image or document fields | Travels with the file | Often stripped by platforms |
| Terms of use text | Legal page of the site | Clear to humans | Machine readability disputed |
Robots.txt is the most practical tool today because large crawlers publish distinct user-agent tokens for AI training. A site can allow a search crawler while disallowing a training crawler, but only if the operator keeps the list of tokens current.
Comparison with related Copyright & AI mechanisms
The TDM opt-out is one of several legal models for training data, and jurisdictions differ sharply. Comparing them clarifies why a reservation that is effective in the EU may have no effect elsewhere.
| Model | Where used | Default position | Role of an opt-out |
|---|---|---|---|
| Exception with opt-out | European Union | Mining allowed | Reservation restores exclusivity |
| Fair use | United States | Case-by-case defense | No statutory opt-out; courts weigh factors |
| Broad mining exception | Japan | Mining largely allowed | Limited where interests are unreasonably harmed |
| Licensing | Any system | Permission by contract | Opt-out creates the incentive to license |
The opt-out also interacts with training data transparency. Under the AI Act, model providers must publish a sufficiently detailed summary of training content, which gives rightholders a way to check whether their reservations were respected.
Legal questions that remain open
Several points are still disputed and depend on future case law. A 2024 decision of the Hamburg Regional Court in a dispute over the LAION image dataset suggested that a reservation written in natural language on a website may count as machine-readable, but that view was not necessary for the outcome and is not settled.
- Timing: a reservation is generally understood to act from the moment it is made, so copies made earlier are not covered.
- Territoriality: whether EU rules bind models trained outside the Union is contested, although the AI Act applies its copyright policy duty to any provider placing a model on the EU market.
- Who may reserve: platforms and publishers can express reservations for content they host, but authorship and licensing chains can make the authority to reserve unclear.
- Downstream datasets: a reservation on the original site does not automatically follow copies in third-party datasets.
Implementing a TDM opt-out in practice
An effective opt-out is a layered, maintained configuration rather than a single file. The decision table below links typical goals to suitable measures.
| Goal | Recommended measure | Reason |
|---|---|---|
| Stay in search, leave training | Disallow training user agents only | Search visibility is preserved |
| Strong legal evidence | TDM protocol plus terms of use | Documents an express reservation |
| Protect distributed images | Embedded metadata and watermark policy | Signal travels with the file |
| Monetize training use | Reservation plus license offer | Creates a clear path to permission |
The most common mistake is assuming that an opt-out deletes existing copies. It does not; it only prevents reliance on the exception from that point on. A second mistake is blocking every crawler, which also removes the site from search results.
In everyday language
In Europe, companies may copy texts and images that are freely available online to analyze them with software, for example to train an AI model. The owner of the content can say no, but has to say it in a way that machines can read, such as a rule in a file that crawlers check first.
Think of it as a sign on a shop window that robots understand. Human visitors still come in, but the automatic collector is told not to take anything for its training archive. The sign only works from the day it is posted.
Seen in the real world
A regional news publisher with about 40,000 archived articles decides that it will license its archive for AI training rather than let it be used for free. Its technical team adds rules to robots.txt that disallow known training crawlers while keeping search crawlers allowed, publishes a TDM reservation file and updates the terms of use with an explicit reservation under Article 4.
Six months later, a model provider approaches the publisher for a license, citing the reservation detected by its crawler. The opt-out turned free use into a negotiated agreement, while the archive kept its visibility in search results.