The four big AI labs backed one switch to opt out of training: what changed on October 30
Direct answer
On October 30, 2026, OpenAI, Google, Microsoft and Anthropic backed a shared robots.txt token for expressing a no-training preference: __usage_ai-training. A site writes that line as a user agent, followed by Disallow, and can add exceptions with Allow. The practical idea is simple: instead of configuring a different rule for each crawler, a publisher states the preference once and the four companies commit to honoring it for model training.
What it is and what it is not
- It is a draft, not a final standard: the document is an individual IETF Internet-Draft titled Machine-Readable Training Preference Signal, authored by Kevin Kelley of Anthropic, dated October 30, 2026 and expiring May 3, 2027. An individual draft carries no formal IETF endorsement and is not an adopted standard.
- It governs training, within its own limits: the signal regulates use of content to train models. Crawling for search and web grounding have their own controls and sit outside this signal.
- It does not delete what was already trained: the preference looks forward. Content already included in existing training datasets is not removed by publishing this line.
- Meta is not part of the joint announcement: coverage notes Meta is on a separate track for its own terms. A business publishing on Meta platforms should review those terms separately.
The timelines, company by company
- Anthropic: it has already started implementing support for the token, per announcement coverage.
- OpenAI: it aims to implement it in the coming months, per its own October 30 blog post.
- Microsoft: it targets February 1, 2027, with more granular controls that separate training from web grounding.
- Google: it targets early 2027. Per its documentation, a site using the shared token will be treated as if it had blocked Google-Extended, unless Google-Extended is set directly in another way.
What to do this week if you publish content
- Decide separately: training and search visibility are different decisions. Before publishing a rule, define whether you want to limit training, appearance in results, or both, and check that the chosen rule expresses that decision.
- Review your current robots.txt: many sites carry per-crawler rules accumulated over years. The shared token simplifies what is new, but it does not automatically clean up old rules that may contradict it.
- Document the decision: record who decided the preference, when it was published and which exceptions were left with Allow. When third parties produce the content, that note prevents a rule from changing without a record.
- The honest caveat: until each company completes its implementation, the signal is a published commitment with different dates, not a switch that already works the same way at all four. Check each company's date before telling readers that your content is out of training.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- Privacy and data: what happens to your customers' information
- What an AI agent is and how a real business uses one
- How to choose an AI vendor: 7 questions before signing
Frequently asked questions
What did the four companies announce on October 30, 2026?
That they will back a shared robots.txt token, __usage_ai-training, so a site can state once its preference not to have its content used to train AI models.
How is the token written?
As a user-agent line in robots.txt, followed by Disallow, with optional exceptions using Allow.
Does the token block search?
No. The signal governs model training, not crawling for search or web grounding, which have their own controls.
Does it already work at all four companies?
Not equally yet. Anthropic has started implementing it, OpenAI targets the coming months, Microsoft targets February 1, 2027 and Google targets early 2027.
