A rights reservation only counts if it's machine-readable

The process wall

An opt-out from AI training is not one button; the documents that create it operate at different layers, from statute to platform setting. The oldest of the three is Directive (EU) 2019/790, dated 17 April 2019, which created the underlying legal hook: an exception letting AI developers mine text and data unless “the rightholders have not reserved in an appropriate manner the rights” to do so. For content already public online, the directive states that reservation “should only be considered appropriate” if made by “machine-readable means, including metadata and terms and conditions of a website or a service” — a form requirement, not just an intention.

What the documents show

Five years later, the General-Purpose AI Code of Practice, published 10 July 2025, builds a “Copyright” chapter on top of that same rule, offering providers of general-purpose AI models “practical solutions to meet the AI Act's obligation to put in place a policy to comply with EU copyright law” — meaning a provider's compliance is measured against whether it respects reservations made exactly the way the 2019 directive specifies. Outside the EU, the mechanism looks different again: OpenAI's own help article states that “website owners can manage whether publicly available content from their sites may be accessed for use in training by using standard web controls such as robots.txt to disallow GPTBot” — a single crawler-level control, not a rights reservation under any statute.

Authorship and limitation

The mechanism worth separating is scope: a robots.txt block against GPTBot only reaches OpenAI's own crawler and says nothing about any other company's training practice or about material already collected before the block was added. The EU's machine-readable reservation, by contrast, is a legal opt-out tied to a directive with cross-border effect among EU member states, enforced through the copyright compliance obligations the Code of Practice describes rather than through a single company's goodwill. Neither route retroactively removes material already used in a completed training run; both documents describe forward-looking controls, not deletion.

What survives

This is an editorial guide built from three separate documents, not a single opt-out standard: a maker checking whether their own site or portfolio is protected has to check the robots.txt-level block against named crawlers, separately from whether their site's terms or metadata carry a machine-readable reservation of the kind the EU directive requires, since one blocks a named company's crawler and the other invokes a broader legal right.

• Check a site's own robots.txt for named AI crawler tokens, and confirm which specific companies' documentation says they honour that block.
• Look for a machine-readable rights reservation in a site's terms or metadata, which is the form the EU directive specifies for an opt-out to count.
• Ask whether a claimed opt-out route is described as retroactive or forward-looking only, since none of these documents claims to undo a finished training run.

An opt-out is only as strong as the document that defines it: a crawler block reaches one company's future crawling, a machine-readable reservation invokes a broader statute, and neither reaches backward into a model already trained.