AI training opt-outs: the new shared signal still needs a rollout check

A publisher can choose to stay discoverable while questioning whether its work should train a generative model. Announcements on 30 October offer a shared way to express that training preference across participating providers. The immediate job for a website owner is to understand which provider will read the signal, when, and how it sits beside existing settings.

The proposed signal uses a policy token, _usageai-training, inside robots.txt. It could reduce the need to repeat the same choice for separate companies. But the document behind it is an individual Internet-Draft, with no formal standing or endorsement in the IETF standards process. Treat it as work in progress, rather than a finished internet standard. Read the draft and its status.

Three questions before changing a setting

• What content and use do you intend to allow or withhold?
• Which providers currently implement the preference, and which have only announced plans?
• Which existing crawler or page-level rules must continue to do a separate job?

A shared announcement has several rollout dates

OpenAI says it aims to implement the signal in the coming months and will publish implementation details. It says its existing crawler controls will continue to apply. That gives publishers a reason to keep the existing provider-specific instructions in view while they assess the new option. OpenAI announcement.

Google aims to support the new control in 2027. Its announcement says that where the shared token opts out of training and Google-Extended is not explicitly specified, it plans to interpret that as a Google-Extended disallow. Existing company-specific controls will continue alongside the shared signal. A broad preference and an explicit provider setting therefore need to be read together. Google's explanation.

Microsoft describes separate training and grounding preferences. Its new preferences begin taking effect on 1 February 2027 as it crawls pages. It also sets out a transition for publishers using NoArchive or NoCache, including actions before 1 March. Anyone relying on those older tags should read the full transition instructions rather than assume that a new announcement has already preserved every setting. Microsoft's implementation and transition details.

Training, grounding and crawl access are separate

Training concerns changing a model's learned parameters; grounding uses content to help inform generated responses. Microsoft says its grounding preference is page-level, while the shared robots.txt preference concerns training at a site or path. It says these choices do not change eligibility for traditional Bing results, subject to normal crawl and indexing requirements. They do not create a licence or payment arrangement.

There is another boundary in the draft itself: the policy token is not a crawling directive. Its content-use preference is independent of rules for a crawler's own product token. A publisher wanting to control retrieval must still consider its crawler rules. The presence of a training-preference line is consequently not a measurement of who fetched an article, or proof that an unwanted use stopped.

Make the editorial choice inspectable

Our suggested starting point is a short record of the site or section, the intended use, the current rules and the person accountable for changing them. Ask whoever maintains the site to compare that record against each provider's current documentation. Preserve the existing configuration and agree how to check a change. This is a proposed working method, not a crawler audit PING has carried out.

A Lagos newsroom, a Nairobi creator site and a London publisher can ask the same operational questions. The announcement does not establish identical contracts, legal treatment or enforcement in their countries. Separate those questions from the narrower technical choice about expressing a preference, and keep provider implementation dates beside the editorial decision.

The useful outcome is a documented choice with an owner and a recheck date. The common signal could simplify maintenance as support arrives; it is not a reason to treat all existing controls as redundant today. Explore more source-led explanations on PING's Innovation desk.