One token could express a training preference across providers

OpenAI announced on October 30 that it plans to recognize a shared policy token in robots.txt for preferences about generative-AI training. The current Internet-Draft assigns the token __usage_ai-training. Website operators can associate it with ordinary Allow and Disallow path rules, so a site could express one preference for most pages and another for a specified directory.

The proposal separates content-use preferences from crawler access. The draft says the token is not itself a command to crawl or stop crawling. Existing product-specific user-agent rules continue to control automated retrieval, while a participating provider interprets the shared token as a statement about training use.

OpenAI has announced support, not completed deployment

OpenAI says implementation is planned for the coming months and that it will publish more details later. Until then, the company's current documentation continues to use GPTBot for training-related preferences and OAI-SearchBot for search visibility. Those controls are independent, and the announcement says they will remain available after the shared signal is added.

Website operators therefore should not replace a working OpenAI-specific configuration on the strength of this announcement alone. A useful rollout check will be whether OpenAI updates its crawler documentation, states an effective date and explains how a shared rule interacts with a more specific GPTBot rule.

The proposed mechanism can still change

The machine-readable signal is version 00 of an Internet-Draft dated October 30. It is marked as standards-track work and expires May 3, 2027 unless it is revised or advanced. It is not an RFC, and the related vocabulary draft says its text does not represent working-group consensus.

The reviewed materials provide no deployment geography or pricing because the change concerns website policy signaling rather than a paid product. They also do not establish universal provider participation or legal effect. CGH has not tested an OpenAI implementation because none was documented as live.