AI: how to stop AIs from learning from your site, in a single line?

Do you have a blog, an online store, a photo site? If you don't want AIs to learn from what you publish there, today you have to draw up a list. GPTBot for OpenAI, Google-Extended for Google, ClaudeBot for Anthropic, CCBot, Bytespider, and a new one shows up every three months. On October 30, OpenAI, Google, Microsoft and Anthropic announced that they would all respect one single line.

One line for all four, finally! Except that it isn't a lock, and nobody starts on the same day.

The same principle as the “No junk mail” sticker, but for AI robots

The sign at the entrance to your site

Everything happens in a small file that has existed since 1994: robots.txt. It's a simple text file placed at the root of a site, which search engine robots read before coming in. It says who is allowed to visit what. Exactly like the “No junk mail” sticker on your mailbox: it doesn't physically block anything, it asks. Serious distributors respect it, the others don't care.

The problem is that every AI company sends its own robots, with their own names. To say no to training, you therefore have to list them one by one, and come back to update the list every time a new one shows up. You might as well stick a sticker on for every pizza brand.

Today's list, and tomorrow's promised card

The new line, as it is written

The signal is called __usage_ai-training. Yes, the name was chosen by engineers, you can tell: nobody will remember it the first time, copy it. It doesn't refer to a particular robot, it refers to a use: training a generative AI, that is, the moment when your text or photos are used to make the model itself. To say no everywhere on your site, it looks like this:

User-agent: __usage_ai-training
Disallow: /

And if you want to leave part of it open, your articles for example, you add an exception:

User-agent: __usage_ai-training
Disallow: /
Allow: /blog/

Watch out for a trap: the general rule User-agent: * that many sites already have doesn't count. The signal only counts if it is written out in full.

The idea comes from Anthropic. Kevin Kelley, from Anthropic, submitted it on October 30 in the form of a draft to the IETF, the organization that writes the technical rules of the Internet. A detail that matters: it is a personal draft, not yet a standard, and it expires on May 3, 2027 if it doesn't move forward.

Who will respect it, and starting when

This is where the enthusiasm drops a little. Nobody starts on the same day:

• Anthropic, which wrote the draft, supports the signal, but I didn't find an official launch date;
• OpenAI talks about “the coming months”, with no date;
• Microsoft will start reading it on February 1, 2027;
• Google, “in 2027”, will treat it as if you had closed the door to Google-Extended, its robot reserved for AI;
• Meta “supports the idea”, but hasn't committed to anything.

In short, if you add the line tonight, it will be almost useless until next year. It costs nothing to add it anyway.

What it really brings

The good news, and it's not a small thing: saying no to training won't make you disappear from Google or Bing. Google specifies that its setting for AI affects neither inclusion in its results nor ranking, and Microsoft writes that these preferences change nothing about your presence in Bing. Until now, a lot of people didn't dare block it, for fear of shooting themselves in the foot when it came to SEO.

And apparently, that's what people were waiting for. In mid-September, Cloudflare, which protects a good part of the world's sites, was already counting 17% of site owners limiting AI training, almost one in six, compared with less than 1% blocking search engines, less than one in a hundred. People want to be found. They don't necessarily want to serve as free raw material.

What it doesn't do

Three limitations, and they matter.

First, it only concerns training. When you ask Copilot or ChatGPT a question and it goes and reads your page live to answer you, that's not training, that's what Microsoft calls "grounding": the AI consults your page like you consult a dictionary, without learning it by heart. Microsoft is preparing a separate setting for that, page by page, and warns bluntly that closing that door will make you less visible in AI answers.

Next, it only applies to the future. What was already scraped last year is already in the models, and no robots.txt line will get it back out.

The label is firmly stuck on, the cookies have already been eaten

Finally, it's a polite request, not a lock. The signal relies on the goodwill of those sending the robots. The four big ones have promised. The hundreds of small anonymous scrapers, on the other hand, haven't promised anything at all. And nobody talks about money: the signal says "no", it doesn't let you say "yes, but you pay".

So, are you adding it or not?

It depends on what you sell. A photographer, illustrator or writer who makes a living from their creations has every reason to say no to training. A shop or craftsperson, on the contrary, may have an interest in making sure the AIs know their products well. I told you last month that AI agents are already getting lost on badly made sites: the day your customer asks their assistant where to buy, it's better if your site is readable to it.

For me, this is genuinely good news, the first time the four biggest players have agreed on something this simple. It doesn't solve everything, far from it, but it's a clear sticker on the mailbox, and that's better than a list of twenty names to keep updated all by yourself. And to Meta, which "supports the idea" without promising anything: we're waiting for you to put the sticker up too!

Sources

• SiliconANGLE, October 30, 2026: Google, Anthropic and others will support a new opt-out signal
• IETF: Kevin Kelley's draft, "Machine-Readable Training Preference Signal"
• Microsoft Bing, October 2026: training and grounding settings, and their dates
• Digidai, October 31, 2026: what the signal changes, and doesn't change, for publishers
• Cloudflare, September 15, 2026: the share of sites limiting AI training

Article written with the help of Claude Code, proofread and corrected by me.