Your best article took a week to write — and a chatbot can now summarize it in seconds, possibly trained on it without asking. Until last week, saying no meant juggling a different opt-out rule for every AI company. That just changed.
On October 30, 2026, OpenAI, Google, Microsoft and Anthropic backed a single shared robots.txt policy token that lets any site owner opt out of AI training across all four companies at once. This guide shows you exactly how to opt out of AI training — and, just as importantly, what this new rule does not do.
Key takeaway: One two-line robots.txt block now covers four AI giants — but it only covers training, not search results or AI answers.
How to Opt Out of AI Training in Two Lines
Here is the whole trick. Add this block to your robots.txt file:
User-Agent: __usage_ai-training
Disallow: /
That is it. The token __usage_ai-training is a new shared policy token from an IETF Internet-Draft written by Anthropic’s Kevin Kelley and dated October 30, 2026. Every participating company reads the same token, so you no longer need a separate block for GPTBot, Google-Extended, ClaudeBot and the rest. The announcement was covered in detail by Skimwire.
Two details matter enormously:
- The token must match exactly, apart from letter case. A generic
User-Agent: *group does not count as naming the token. - You can carve out exceptions. The draft’s own example keeps one folder open:
User-Agent: __usage_ai-training
Disallow: /
Allow: /press-kit/
That version says: no AI training on anything except your press kit. A photographer could allow a portfolio folder while protecting client galleries; a news site could open press releases while keeping reporting off-limits.
What the __usage_ai-training robots.txt Token Actually Does
The token is a preference about use, not a crawl rule. The draft states plainly that “this policy token is not a directive for crawling behavior.” Bots may still visit your pages, and your search rankings are untouched.
What it covers:
- Using your content to train generative AI models — the process that changes a model’s learned parameters.
- All four participating companies, through one shared signal.
What it does not cover:
- Content already collected into existing datasets or trained into model weights. Updating robots.txt today cannot un-train yesterday’s models.
- Non-generative models used only for tasks like spam detection or ranking.
- Anyone who simply ignores it. The token is voluntary — a request, not a law — and binds only the companies that choose to follow it.
Who Honors It, and When
Rollout is staggered:
- Anthropic: Already begun implementing support, according to the draft’s author.
- OpenAI: Aims to implement the signal “in the coming months,” with details to follow.
- Microsoft/Bing: Starts honoring the preferences on February 1, 2027, as it crawls.
- Google: Aims to support the control in 2027, without a more specific date yet.
- Meta: Expressed support for the idea but has not committed to implementing it.
The practical move: Add the token now, and keep your existing company-specific rules until the shared token is live everywhere. The old rules protect you today; the new token protects you tomorrow.
Google-Extended Opt Out: What Changes for Google
Google has said that if a site uses the shared token without naming Google-Extended separately, Google will read it as if the publisher had disallowed Google-Extended. In plain English: the new token doubles as a Google-Extended opt out.
What to do: If you already disallow Google-Extended, add the shared token alongside it. If you have never touched Google-Extended, the shared token alone now handles Google’s training opt-out once its 2027 support lands. Either way, do not remove your Google-Extended block yet — until Google’s implementation is live, it remains the rule that actually works.
See also our guide on how to stop Google AI from using your content.
Bing AI Training Opt Out: February 2027 and the NoArchive NoCache March 2027 Deadline
Microsoft went further than the other three companies — and set the firmest dates. This is the part most site owners will miss if they only skim the headlines.
Bing AI training opt out, February 2027: Microsoft starts honoring the shared preferences on February 1, 2027, as it crawls. Note the catch: a page’s setting only updates when Microsoft recrawls it, so changes take effect gradually, not overnight.
NoArchive NoCache March 2027 deadline: Microsoft currently reads NoArchive and NoCache as AI-related preferences. After March 1, 2027, it stops. If your site relies on those older signals for Bing, switch to the new token before that deadline — or your preference will silently stop working. See RR Tech Pulse’s timeline breakdown for the full Microsoft details.
Microsoft also added a separate control for AI grounding — the step where live web content is pulled into chat answers. Its DisallowExtended option lets content inform Microsoft’s own AI experiences while blocking the grounding services Microsoft sells to other companies.
AI Training vs AI Answers Grounding: Don’t Confuse Them
This is the single biggest misunderstanding about the new token, and it could cost you traffic if you get it wrong.
| Control | What it stops | What it does NOT stop |
|---|---|---|
__usage_ai-training token | Your content being used to train AI models | Crawling, search rankings, or AI answers quoting your pages |
| Google-Extended | Google’s AI training (and Gemini app use, per Google’s docs) | Your pages appearing in Google Search or AI Overviews |
| Search Console “Search generative AI” setting | Your content appearing in AI Overviews and AI Mode | AI training |
| Bing grounding controls | Your content grounding Bing/Copilot answers | AI training or regular Bing results |
| Standard robots.txt crawl blocks | Crawlers visiting blocked pages | Anything on pages you leave crawlable |
A concrete example: Imagine you run a small cooking blog and add the __usage_ai-training token. Next year, a chatbot could still answer “how do I make sourdough?” by quoting your recipe — because grounding is separate. If you also want out of AI answers, you need the separate grounding controls: Google’s Search Console setting and Bing’s grounding controls.
Key takeaway: Opting out of training is not opting out of AI answers. Decide which one you actually object to — most site owners object to both, and that means flipping two switches, not one.
Step-by-Step: Add the Opt-Out to Your Site
- Back up your current robots.txt. Download a copy before editing, so you can roll back if anything breaks.
- Add the shared token block at the end of the file, exactly as shown above. Keep your existing GPTBot, Google-Extended, ClaudeBot and other blocks for now.
- Add carve-outs if you want them. Use
Allow:lines for any folder you are happy to share, like a press kit. - Check Microsoft’s deadline. If you use
NoArchiveorNoCachefor AI preferences, switch before March 1, 2027. - Handle AI answers separately. If you also want out of AI Overviews and AI Mode, use Google Search Console’s “Search generative AI” setting, and check Bing’s grounding controls.
- Test the file. Use Google Search Console’s robots.txt tester or a validator to confirm the file parses cleanly.
If you would rather have an expert handle the technical side, our team can audit your robots.txt and AI opt-out setup — see our services or contact us for help.
Frequently Asked Questions
Does opting out of AI training hurt my Google rankings?
No. The token is a use preference, not a crawl directive. It does not block Googlebot, remove pages from search results, or affect rankings. Microsoft has said the same about Bing: opting out of training does not affect eligibility for regular Bing results.
Is the __usage_ai-training token a law or an official standard?
Neither, yet. It is an individual IETF Internet-Draft — not endorsed by the IETF and with no formal standing in its standards process — and the mechanism is voluntary. It works because the biggest AI companies agreed to honor it, not because anyone is forced to. Emergent’s report has more background on the draft’s status.
What if I only want to opt out some of my content?
Use Allow: lines to carve out folders, or apply the token block to specific paths instead of the whole site. You can also keep company-specific blocks for finer control while the shared token rolls out.
The Bottom Line on How to Opt Out of AI Training
For years, protecting your content from AI training meant a messy patchwork of per-company rules — and hoping you had not missed one. The shared __usage_ai-training token finally collapses that patchwork into a single line, backed by the four biggest Western AI labs.
Add the token today, keep your old rules until the new one is live everywhere, and remember the two dates that matter: February 1, 2027 for Bing, and March 1, 2027 for the NoArchive and NoCache cutoff. And if AI answers quoting your work bother you as much as AI training on it, flip the grounding switches too — because the token that stops training will never, on its own, stop the answers.
