AI training opt-outs: a shared signal with staggered plans

Start with the proposal’s actual status

The IETF Datatracker lists Kevin Kelley’s October 30 submission as an active individual Internet-Draft. Its notice says the document is not IETF-endorsed and has no formal standing in the standards process. The draft’s existence is a development; completion of a standard is a separate milestone.

For an editorial team, that distinction belongs in the change record. A new technical option may deserve review before it becomes a finished standard. The record should still explain why the organization considered it, which version it read and who owns the next review, rather than treating a recognisable standards-body name as approval of every proposed detail.

Source notes: 1. Original editorial examples and interpretations are identified separately.

Separate a use preference from a fetching rule

The proposed token is __usage_ai-training. It appears in an ordinary robots.txt User-Agent group with Allow or Disallow rules for paths. The draft keeps crawler-specific rules separate: this signal concerns use for training, rather than instructing all crawlers to stop fetching. A wildcard User-Agent group does not match the policy token; without an explicitly named group, no preference is expressed through this mechanism.

That creates two questions for a publisher to answer in plain language. May a provider retrieve a page, and may it use that page for the specified training purpose? A file review should preserve both answers. Replacing an established crawler rule because a new purpose-based signal exists could lose an existing choice before its replacement is effective.

Source notes: 2. Original editorial examples and interpretations are identified separately.

A shared spelling does not mean a shared launch date

OpenAI’s October 30 announcement calls its participation voluntary and targets implementation in the coming months, with details to follow. It says existing crawler controls will continue. This is an announced plan, not evidence that OpenAI already applies the new signal to every relevant collection workflow.

Google’s separate October 30 post targets 2027. It plans to interpret a shared training opt-out as a Google-Extended disallow when Google-Extended is not explicitly specified. Existing company-specific controls continue alongside the new signal. A publisher therefore needs to read its explicit Google-Extended choice before assuming what the broad preference would mean.

In an October 30 message to the IETF list, Kelley said Anthropic had started implementation, while adding that support across its services would take time and interpretation could evolve. An implementation effort and completed service-wide deployment are different states. The message does not supply a universal completion date.

Source notes: 3, 4, 5. Original editorial examples and interpretations are identified separately.

Read Microsoft’s training and grounding choices separately

Microsoft’s October 30 announcement sets February 1, 2027 as the start for honoring new preferences as it crawls. It also introduces separate page-level grounding preferences. These govern use in generated responses and are not the shared training setting. Microsoft says neither choice guarantees inclusion, traffic or payment.

Its transition guidance asks existing NoArchive or NoCache users to act before March 1, 2027. Some grounding transition steps involve Bing Webmaster Tools as well as page settings. A publisher using those older preferences should read that specific guide, rather than assume one new robots.txt group completes every migration task.

For planning, attach a provider and an effective condition to each deadline. “Published on our server” describes the publisher’s action; “discovered on recrawl” describes another system’s action. Keep an unresolved discovery step visible. That is a useful operational distinction even when the provider does not offer the publisher a direct receipt for every downstream use.

Source notes: 6. Original editorial examples and interpretations are identified separately.

Keep today’s controls legible while reviewing tomorrow’s

OpenAI’s current crawler documentation describes independent settings for OAI-SearchBot and GPTBot. Its example permits search while disallowing training. The same page distinguishes user-triggered ChatGPT-User visits, for which robots.txt rules may not apply. That guidance concerns OpenAI’s named agents, not every provider’s interpretation.

RFC 9309, the existing Robots Exclusion Protocol, says its rules are not access authorization. It also defines the group structure and path matching that the new proposal reuses. A public preference file should therefore not be treated as the place to enforce access to private material.

Our suggested review begins with a copy of the current published file and a short inventory of intentional choices. Mark the rule’s purpose beside its owner, and identify any disagreement between a broad instruction and a provider-specific one. This is a proposed editorial workflow; Lumacta has not changed a reader’s file or tested the resulting behavior.

Source notes: 7, 8. Original editorial examples and interpretations are identified separately.

Create a record that survives the announcement

A practical record has six fields: the intended content use, the relevant pages, the provider, the published setting, the implementation status and the next review date. Keep the exact source and reading date beside the provider’s statement. Do not turn a broad phrase such as “coming months” into an invented calendar deadline.

Add a second column for evidence actually available to the team. A successful fetch of the live preference file establishes what the site serves at that moment. An announcement establishes a provider’s stated intention. Neither observation, by itself, demonstrates what went into a model’s training set. Leave the latter question unresolved when no supporting evidence is available.

Source notes: 1, 3, 4, 5, 6. Original editorial examples and interpretations are identified separately.

Account for what this standards work leaves outside scope

The AI Preferences working-group charter excludes technical enforcement, client authentication or authorization, and auditing or transparency measures for training. That scope helps explain why a machine-readable preference is one part of a publishing policy rather than a complete assurance system.

Our analysis is to ask for the next observable event before calling the work complete: a documented implementation update, an applicable provider explanation or another relevant record. Preserve the unanswered question when that event has not occurred. This article does not determine copyright ownership, grant a licence to third-party material or replace advice about a particular agreement.

Source notes: 9. Original editorial examples and interpretations are identified separately.

What the evidence establishes

This explanation compares dated primary documents and proposes a review method. We performed no compliance audit, training-data inspection or crawler experiment. The archive photograph shows physical server equipment in 2012; it depicts none of these providers’ current implementations and supplies no evidence of training or a policy being honored.

Source notes: 2, 6. Original editorial examples and interpretations are identified separately.