AI Opt-Out : Four Labs Wrote in One Line the Rule Brussels Hasn’t Finished

AI Opt-Out : Four Labs Wrote in One Line the Rule Brussels Hasn’t Finished
3 min read

On October 30, OpenAI, Google, Microsoft, and Anthropic adopted a common signal in robots.txt to refuse AI model training. This mechanism, which responds to a European expectation, provides neither acknowledgment of receipt nor verification of compliance.

One line in robots.txt

The robots.txt file allows a site to indicate to robots what they can read. The signal described by Kevin Kelley of Anthropic in a draft submitted to the IETF, the Internet standards body, involves adding a group whose header bears the token __usage_ai-training, followed by a Disallow. No new directive is necessary: existing reading tools ignore this group, as they do any unknown agent. The preference can be set by folder, with the draft example blocking the entire site except the press folder. Without this group, no preference is expressed, which Microsoft reads by default as permission.

The token concerns only usage, not access to pages. It covers training, meaning the modification of a generative model’s parameters, but not grounding, which is consulting a page when the assistant responds. OpenAI mentions the coming months, Microsoft February 1,2027, and Google 2027 for the application of this signal. Meta welcomed the idea without committing.

What European law requested

The 2019 copyright directive allows rights holders to reserve their rights against text and data mining, provided they do so “in an appropriate manner, ” including through machine-readable means. The AI Act, in its Article 53, requires providers of general-purpose AI models to respect these reservations “including through state-of-the-art technologies.” The code of practice, signed by 22 companies including the four authors of the announcement, cites robots.txt “and any subsequent version, ” before other protocols to be identified.

On November 3, Advocate General Maciej Szpunar, First Advocate General at the Court of Justice of the European Union, delivered his opinion in the case of Like Company v. Google. His opinion does not bind the judges. He considers that training on lawfully accessible content generally falls under the text and data mining exception, and that a reservation is appropriate if state-of-the-art robots can understand it with sufficient reliability. He cites robots.txt, Google-Extended, and file metadata, stating that standardization is, in his view, underway and should be completed “in all probability.”

For its part, the European Commission consulted, from December 1,2025 to January 23,2026, on seven protocols other than robots.txt. As of November 4, no list had been published. Its report notes the concern of model providers that the consultation “prematurely endorses immature solutions.” Google states in its blog post that the proposal is “consistent with the EU’s code of practice.”

What the line does not resolve

The token remains, in the words of the draft, a preference. The Commission’s report notes that “it is not possible to verify that a reservation has been respected.” The PEReN, the state’s digital regulation expertise hub, described in July 2025 a “trust-based” system where the robot self-declares. The Advocate General observes that rights holders have no means of ensuring their refusals are respected, and that evidence shows some developers ignore them.

In May 2024, OpenAI promised a tool allowing creators to refuse training “by 2025.” According to the source, it never materialized. In France, according to a written question published in June 2026, SACEM, SACD, and SCAM have notified their opposition to hundreds of providers since October 2023, “for the most part” without response.

The Advocate General’s opinion further distinguishes the case of a chatbot reading a page when responding. The proposed criterion is then illicit access, particularly circumvention of effective technical measures. He notes that robots.txt technically prevents nothing. Finally, the line only covers training; copies already made remain in the models, and Google reads it as blocking Google-Extended, which also removes the site from Gemini’s responses, whereas Microsoft separates the two choices. The Commission must conclude its consultation on copyright and AI on November 10 and publish its list of protocols.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x