OpenAI and Google Plan Shared AI Training Preference Control

Website operators would be able to express a common preference about generative AI training through a robots.txt signal that OpenAI and Google plan to support, according to separate company announcements on October 30, 2026. OpenAI said participation is voluntary and put its planned implementation in the coming months.

The proposed approach is intended to give operators one way to communicate their preference to multiple participating providers, OpenAI said. The company described the previous process as requiring operators to identify each provider’s crawler and keep separate instructions for those crawlers.

An individual Internet-Draft available through the Internet Engineering Task Force Datatracker sets out the mechanism. Under the proposal, standard groups within robots.txt would use a common policy token to identify training preferences. Operators could apply those preferences to specific website paths rather than using the same setting throughout a site. The draft illustrates this with a configuration that opposes training use for most paths but permits it for one designated path.

For this purpose, the draft defines AI training as using content to alter a generative model’s learned parameters. Work devoted exclusively to non-generative tasks falls outside that definition, including when the model involved also has the ability to generate content.

The draft also distinguishes a preference about content use from an instruction governing crawler access. It says the proposed signal is not itself an access rule, with instructions for particular crawlers remaining separate. According to the draft, a private agreement may take precedence over a preference communicated through the shared mechanism.

Google’s planned treatment includes a condition for sites without an explicit Google-Extended setting. In those cases, the company said it would interpret an opt-out from training under the shared control as a disallow instruction for Google-Extended. Google presented that interpretation as part of its intended support, rather than an already implemented change.

The specification remains an individual Internet-Draft, listed by the IETF Datatracker as published on October 30, 2026. The Datatracker describes documents in that category as working material with no formal status within the IETF’s standards process.

* This article is based on publicly available information at the time of writing.