On October 30, 2026, four of the largest Western AI laboratories — Google, Microsoft, OpenAI, and Anthropic — announced they will honour a single shared signal that lets a website opt out of AI model training. Instead of writing separate rules for each company's crawler, a publisher will be able to write one entry in its robots.txt file and have it understood by all four. It is the first coordinated attempt of its kind, and it arrives at a moment when publishers, regulators, and AI companies are still arguing over who controls the content that trains generative models.
The signal is a policy token called __usage_ai-training. A site owner adds a group to the robots.txt file at the root of the site, with User-Agent: __usage_ai-training followed by Disallow: / to opt out the whole site, or a narrower path to opt out only part of it. Exceptions can be carved out with Allow: lines for specific folders. The token names a use — training — rather than a crawler, which is what makes it different from the familiar per-bot rules such as GPTBot or Google-Extended that publishers use today.
Where the token comes from
The technical write-up behind the announcement is an individual Internet-Draft titled "Machine-Readable Training Preference Signal," authored by Anthropic's chief standards engineer Kevin Kelley and submitted to the Internet Engineering Task Force's AI Preferences working group on October 30, 2026. That status matters. An individual draft is not an official IETF standard, carries no formal standing in the IETF process, and is not endorsed by the IETF. The announcement is therefore a voluntary industry commitment, not a regulation and not a finished standard.
What the token does, and what it does not do
The draft is explicit on its limits, and publishers who misunderstand them could make costly mistakes. First, the token is a statement of preference about how content is used, not a directive about crawling. As the draft puts it, the policy token "is not a directive for crawling behavior." Opting out of training does not stop a crawler from visiting or caching a page, and it does not remove content already collected into datasets or already trained into model weights. It applies to future training.
Second, the token covers training only. It does not block search crawling, and it does not keep a page out of AI-written answers that quote or summarise it. For many publishers, that second use — retrieval and grounding in live answers — is now the more immediate commercial concern, because it can divert traffic even when training is off the table. A publisher that opts out of training but wants to control AI answers needs separate controls, and those controls differ by provider.
Timelines are staggered, and Microsoft goes furthest
Nothing changes immediately. Anthropic says it has already begun implementing support for the token. OpenAI says it aims to implement the signal in the coming months, without committing to a specific date. Microsoft will honour the signal from February 1, 2027, and Google says it aims to support it in 2027. Google's John Mueller confirmed that if a site uses the shared token without explicitly naming Google-Extended, Google will interpret it as if the publisher had disallowed Google-Extended.
Microsoft went further than the shared token alone. It is adding a page-level training setting and a separate "AI Grounding" setting that decides whether a page can feed AI answers in Bing, Copilot, and Microsoft 365, with a DisallowExtended option that allows grounding in Microsoft's own products but not in the grounding services it sells to other companies. Microsoft is also retiring the training and grounding meaning of its NoArchive and NoCache tags after March 1, 2027, which means publishers that currently rely on those tags to limit Copilot will need to act: according to reporting, those still using the old tags must record a temporary preference in Bing Webmaster Tools, a setting that opens December 15 and is open only to sites still using the tags on that date.
Why it matters for founders and investors
For founders, this announcement is a practical prompt to audit how their content is used, not just how it ranks. If your business depends on proprietary content, documentation, community contributions, or licensed material, the new token gives you a simpler way to state a training preference across four major labs — but only for future training, only if you implement it correctly, and only on each lab's timeline. Treat it as one layer of a content strategy, alongside licensing, paywalls, terms of use, and the separate grounding controls that actually affect AI answers and traffic.
For investors, the pact is a sign that the AI data economy is moving from confrontation toward negotiated plumbing. Voluntary standards rarely settle the underlying questions of compensation and copyright, and this one does not try to. But shared technical signals reduce friction, make publisher preferences machine-readable, and create the infrastructure on which future licensing markets could be built. The companies that build tools to manage, audit, and monetise those preferences — for publishers, enterprises, and creators — are operating in a space that just got a little more standardised, and therefore a little more investable.
Angeled.com connects founders with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



