One line could tell every AI company: don’t train on my site

In short: today, saying “do not train AI on my site” means listing OpenAI’s, Google’s, Anthropic’s and other crawlers one by one in robots.txt. A new draft aims to replace that with one shared entry.

What changed

• On 30 Oct 2026 the IETF published the draft “Machine-Readable Training Preference Signal”. Author: Kevin Kelley of Anthropic. The draft expires on 3 May 2027.
• It adds a shared token, __usage_ai-training. You put it in a User-Agent: line, followed by Disallow: / (I do not agree) or Allow: for chosen folders.
• The draft’s example is Disallow: / plus Allow: /press-kit/, meaning “do not train on anything except my press kit”.
• No such group means no preference is expressed. A general User-Agent: * rule does not count.
• According to AI Weekly, citing The Verge, OpenAI, Google, Microsoft and Anthropic backed the shared sign. Anthropic has reportedly started implementing it, Microsoft names 1 Feb 2027, Google 2027 and OpenAI “the coming months”. I have not verified that split with each company directly.
• Google reportedly says that a site opting out with this token without naming Google-Extended will be treated as if Google-Extended were disallowed.

What it does not do

• It is not a ban on visits. A crawler may still fetch your pages; the sign only covers training.
• The draft explicitly lets private agreements override what you write.
• Dishonest scrapers can ignore it, and content collected earlier will not disappear from existing models.

How to read this

For now this is an individual draft, not an adopted standard, and it rests on company promises rather than enforcement. Still, adding three lines to robots.txt costs nothing and leaves a clear record of your choice. Keep your existing rules for specific crawlers. I write more about law and AI in the AI Act and law hub.

Sources

Machine-Readable Training Preference Signal (draft-kelley-aipref-training-preference-signal-00)IETF · Nov 2, 2026 · Primary sourcedraft-kelley-aipref-training-preference-signalIETF Datatracker · Nov 2, 2026 · Primary sourceOpenAI, Google, Microsoft, Anthropic back one AI opt-out tokenAI Weekly · Nov 2, 2026