Opting Out of AI Training in One Line? Major Tech Companies Adopt a Shared Standard
機械翻訳 / Machine-translated
A
AI Friends
3日前・18分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
__usage_ai-training — is enough to express your preferenceIf you've ever wanted to prevent your blog from being used to train AI but found the per-company configuration a hassle, you're not alone. On October 30, 2026, major AI companies announced support for a shared mechanism designed to reduce that burden.
On October 30, 2026, Google's official blog, OpenAI, and Microsoft all made announcements on the same day. The common thread: support for a shared signal that site operators can use to opt out of AI training use.
The signal was proposed by Kevin Kelley of Anthropic and published as an IETF draft (the IETF being the body that discusses technical standards for the internet).
Previously, opting out required separate configurations for each company — "Google-Extended" for Google, "GPTBot" for OpenAI, and so on. Opting out from five companies meant writing five separate entries.
With a shared signal, a single entry can convey your preference to all of them.
robots.txt — a plain text file that tells automated web crawlers which parts of your site they may access — is already in use on most websites. This new mechanism makes use of that existing file.
No new syntax is required. According to the draft, it uses the same formatting rules as the existing standard (RFC 9309).
The policy token (a special name representing a type of use) is __usage_ai-training. Here is an example that opts the entire site out of training:
User-Agent: __usage_ai-training
Disallow: /
You can also allow specific sections. The draft gives an example where the entire site is disallowed, but a press kit page at /press-kit/ is explicitly permitted:
User-Agent: __usage_ai-training
Disallow: /
Allow: /press-kit/
This signal is an expression of preference — a request not to use the content for training — and is not a command that stops crawling itself. To restrict crawling, you will still need per-crawler directives as before.
Also note that the wildcard User-Agent: * does not apply to this token. Unless you explicitly write the dedicated token name, your preference will not be registered.
The pace of adoption varies by company.
In other words, writing the signal today will not immediately take effect across all companies. For the time being, it is best to maintain both your existing settings and the new shared setting in parallel.
Microsoft's official blog explains that in addition to training preferences, it will also offer a setting for "grounding." Grounding refers to the practice of an AI reading your page as a source when generating a response to a query.
Training reshapes the AI's underlying model. Grounding involves referencing up-to-date pages each time a question is answered. They are related but distinct.
For grounding, you can choose from three options on a per-page basis:
To opt out of training on a per-page basis, add the following to your HTML:
<meta name="Bingbot" content="ai-training:disallow">
A schedule has also been set. A temporary configuration period in Bing Webmaster Tools begins December 15, 2026, with a deadline of March 1, 2027. After that date, the previously used NoArchive and NoCache directives will no longer be interpreted as preferences regarding training or response generation.
Microsoft warns that opting out of grounding will significantly reduce your visibility in AI-generated answers — though it notes this will not affect your Bing search rankings.
Existing options like Google-Extended and GPTBot were company-specific identifiers. The new signal is a use-case identifier that is not tied to any particular company. That is the key distinction.
However, this is not a replacement. Both the draft and each company's announcement confirm that company-specific settings remain valid. If you want granular control over individual companies, those settings are still available.
This development comes against a backdrop of increasing pressure from publishers. In July 2026, Cloudflare announced a policy that AI crawlers should either pay or be blocked.
This is relevant to operators of news sites, blogs, and corporate media in Japan as well. Major Japanese newspaper websites have long used robots.txt to signal their refusal of unauthorized training use. For those who have been managing multiple per-company configurations, this simplifies the workload.
Consider, for example, a company running an in-house blog with ten contributors. Previously, the person in charge had to verify Google and OpenAI settings separately — and add new entries whenever a new AI company appeared. A single shared token now communicates the basic policy in one place.
For individual bloggers, it's even simpler. Adding two lines to your WordPress robots.txt is enough to signal "please don't use this for training" to all participating companies.
Creators who publish images or illustrations will likely use path-specific configurations — for instance, targeting only the folder where their work is stored. Keep in mind that data already used for training will not be erased.
As for search rankings: the signal applies only to training, and it has been stated that it will not cause your site to disappear from Google or Bing search results.
Here is the recommended approach:
User-Agent: __usage_ai-training block on top of those existing entries.Conversely, if you want your site to appear in AI-generated answers, no action is needed. The default state is "allow."
No. Microsoft's support begins February 1, 2027; Google's in 2027; and OpenAI's within the next few months. Some companies will not act on the signal immediately.
Please do not. Each company has confirmed that existing settings remain valid. The recommended approach is to add the shared setting on top of your existing ones.
No. This signal targets training only and has been confirmed to have no effect on Google or Bing search results. However, opting out of Microsoft's grounding feature will reduce your visibility in AI-generated answers.
No. This mechanism is a preference signal for future use. It does not guarantee the removal of data that has already been used.
Not yet. It is still at the IETF draft stage and has not been approved by the IETF. The content may change going forward.
__usage_ai-training block to robots.txt is all it takes to express your preferenceStart by opening your site's robots.txt and reviewing your current settings.
This article is a cross-post from AI Friends.