Shared AI training preference promises simpler controls, with different rollouts

A shared website preference could reduce the number of separate AI-training instructions publishers maintain, but the announcements made on 30 October are commitments to implement it, rather than proof that one new rule already works everywhere. OpenAI says it aims to add support in the coming months while retaining its existing crawler controls. [1]

The proposed signal gives participating providers a common name to look for in robots.txt, the website file used to communicate crawler preferences. Kevin Kelley’s Internet-Draft calls that name __usage_ai-training and reuses existing Allow and Disallow path rules. Its definition concerns using content to change a generative model’s learned parameters. The draft explicitly separates this content-use preference from rules governing crawler access; it remains work in progress. [3]

Independent specialist outlet PPC Land’s comparison highlights why a shared format does not produce a shared start date. Google says it aims to support the control in 2027. Microsoft says it will begin honoring its new preferences as it crawls from 1 February 2027. OpenAI has given no more specific implementation date in its announcement. [1][2][4][5]

The purpose of the content also matters. Microsoft distinguishes training from grounding: using web content to inform a generated response when an answer is created. It plans a separate page-level grounding preference, allowing publishers to make those choices independently. Google, meanwhile, says it plans to treat a shared training opt-out as a Google-Extended block when the website has not explicitly specified Google-Extended. Its announcement also links a specific Google-Extended allowance to eligibility for grounding in Gemini Apps. [4][5]

Taken together, the records describe a simpler way to express a broad training choice, while leaving implementation timing and product-specific behavior with the providers. They do not establish identical treatment of generated answers across services. The next practical milestones are the companies’ implementations and further updates; OpenAI says it will publish implementation details when support arrives. [1][3][4][5]