Google, OpenAI, and Microsoft want to simplify AI training opt-outs

Google, OpenAI, and Microsoft support a common mechanism for sites to more easily refuse AI training on their content. The signal will rely on robots.txt, but its effectiveness depends on company compliance.

Google, OpenAI, and Microsoft support a common mechanism that allows website publishers to more easily refuse the use of their content for training AI models. The idea, presented as part of the IETF's work, is based on the robots.txt file already used to guide web crawlers. The key point is a single instruction instead of a series of settings for each robot. One important detail remains: this signal does not technically block anything; it mainly asks companies to play along.

Un signal unique pour dire non aux robots d’entraînement

The principle is simple. A publisher can signal once, via the robots.txt file, that a site should not be used to train the AI of participating companies. The system introduces a shared identifier, __usage_ai-training, to centralize this preference. Exceptions will still be possible, leaving room for specific agreements or targeted uses.

Until now, the task was more tedious. Each publisher had to define separate rules for the different robots that collect web content, which complicated daily management. The new framework mainly aims to reduce administrative friction, not to completely reshuffle the rules of content access rights. In the spirit of “please do not enter,” the sign finally becomes readable.

Pourquoi les éditeurs réclamaient ce type de mécanisme

Complaints from publishers have multiplied since AI models have been extensively using articles, images, and other content published online. The issue is sensitive because this data feeds the training of systems without rights holders always having given explicit consent. According to the source, the new mechanism is meant to spare them from repeating the same instruction to each provider.

Google announced on October 30 its intention to support this common process. Microsoft says it will apply the system starting February 1, 2027, OpenAI plans to do so in the coming months, and Google announces implementation in early 2027. Anthropic has also begun applying it, while Meta has not yet confirmed adoption [to be verified].

Un accord utile, mais pas une barrière technique

The central point is that this mechanism does not technically prevent AI robots from retrieving content. It works as a preference expressed by the site, not as a lock. Its effectiveness therefore depends on the discipline of the companies that accept it. If they respect the signal, publishers gain simplicity; if they ignore it, the situation does not really change.

Google also specifies that this new signal will be taken into account for certain uses of content by Gemini. However, refusing training does not automatically mean refusing the use of content to generate responses. The two issues remain linked but are not the same. This is where legal and commercial discussions are likely to continue making noise, even if the robots.txt file tries to speak more calmly.

Ce que cela change pour le marché de l’IA

For AI players, the agreement marks an attempt at standardization. For publishers, it offers a simpler way to express their choice and keep a technical record of this refusal. But the real impact will depend on the commitment of major providers, their timeline, and their interpretation of the signal.

At this stage, the issue is not only technical. It also touches on the distribution of value between AI platforms and content producers, a topic that now comes up in almost every discussion about web monetization. The new mechanism does not solve everything, but it at least offers a common language to say no. And on the internet, that is already a small step forward.

By the numbers

  • October 30 — date of Google's announcement.
  • February 1, 2027 — deadline announced by Microsoft.
  • Early 2027 — timeline announced by Google.
  • 3 companies — Google, OpenAI, and Microsoft support the mechanism.

To read