What is TDM rights reservation (TDMRep)?

TDM stands for text and data mining, the automated process of analyzing large volumes of content to extract patterns, train statistical models, or build AI systems. Directive (EU) 2019/790, known as the DSM Directive, grants in Article 4 an exception that allows this kind of mining without prior permission from the rightsholder, unless that rightsholder has expressly reserved their rights. TDM rights reservation is exactly that exclusion statement: a rightsholder who does not want their content used for mining, and in particular for training AI models, has to say so explicitly. The TDM Reservation Protocol (TDMRep), published as a final report by a W3C community group, offers a machine-readable way to do it: two fields, tdm-reservation and tdm-policy, placed in an HTTP header, in the HTML, or in an extended robots.txt file. The protocol should not be confused with the law: TDMRep is a technical proposal for expressing a right the directive already grants, not the only possible form and not a legal requirement in itself.

The mechanism rests on two separate pieces. The first is the right: Article 4 of the DSM Directive establishes that text and data mining is permitted by default, as long as the rightsholder has not expressly excluded it in an appropriate format, and that for content made publicly available online, that exclusion should preferably be expressed through machine-readable means. The directive does not specify what that technical means should look like; it leaves that to industry to define. The second piece is the implementation: TDMRep fills that gap with two signals. tdm-reservation can take the value 1 (rights reserved) or 0 (rights not reserved), and can be placed in three different locations: an HTTP header called TDM-Reservation, a meta tag in the HTML, or an entry inside an extended robots.txt under the tdm-reservation prefix. tdm-policy, optional, points to a URL where the rightsholder explains in detail the conditions under which they would allow mining, for example in exchange for a license. An AI crawler that respects the protocol must read that signal before adding the content to its training set and stop if it finds the reservation active. The weak point, and it needs to be said just as clearly, is that the protocol depends on the crawler choosing to comply: there is no technical mechanism that prevents an AI system from ignoring the signal, in the same way nothing prevents a crawler from ignoring robots.txt. The legal difference is that whoever places the reservation can point to Article 4 as grounds for a claim if it is ignored. The reservation can also apply at the level of an entire domain or just a single page, depending on where the technical signal is placed.

The decision that depends on this field is whether a publisher, a media outlet, or any site with its own content formally records its objection to AI models training on its material, in a form a European court can recognize as a valid reservation under Article 4. robots.txt, with its Disallow directive, is a technical request with no legal backing: a crawler that ignores it does not break any law by that fact alone, except in cases of unauthorized access to computer systems. llms.txt likewise has no legal standing and no recognized technical status either; no declared AI engine is verified to consume it. TDMRep occupies a different place: it does not technically stop crawling, but it does build the documentary evidence that an express reservation existed, which is exactly what Article 4 requires for the mining exception to stop applying. For a client with editorial content, research material, or any asset they do not want reproduced in a language model, the question is not only technical but a matter of legal risk management: who can prove, with a date, that they reserved their rights before a given model was trained. This level distinction also matters in practice: a publisher that only wants to protect its historical archive can reserve rights solely for those sections without blocking normal indexing of the rest of the site, a distinction a blanket reservation would not allow, and one that should be part of the technical planning rather than an afterthought.

• Publish the tdm-reservation signal on at least two of the three channels (HTTP header and HTML), because not every crawler reads all three.
• Combine the TDM reservation with robots.txt: each protects a different scope, one is a technical crawling request, the other a legal mining reservation.
• Document the publication date of the reservation, because Article 4 requires the exclusion to be express and prior to use.
• If there is intent to license the content for training under conditions, fill in tdm-policy with a real URL rather than leaving it empty.
• Periodically review which AI crawlers actually declare they comply with TDMRep, the list changes, instead of assuming universal coverage.
• Consult legal counsel before basing a claim on the technical signal alone: the reservation is the evidence, not the lawsuit.

• Confusing TDMRep with a law: the protocol expresses a right, it does not create it or enforce it by itself.
• Setting only the extended robots.txt and assuming it covers the same ground as the header or the HTML; different crawlers read different locations.
• Leaving tdm-policy empty when there is actually interest in licensing, missing the chance to monetize access.
• Treating the reservation as if it technically blocks access: it does not, it only determines the legal basis if a crawler ignores the signal.
• Applying the reservation across an entire domain without reviewing what content should actually stay open to mining, for example press content that wants to be cited in AI answers.