AIpediya News
All news

A Single "No" for the Whole Web

For years, telling AI companies not to train on your writing meant learning each company's private language and hoping you had not missed a new crawler. On October 30, that absurd chore got a little less absurd: Google, Microsoft, and OpenAI said they will honor one shared line in robots.txt, the small file where a site already talks to crawlers, and the draft behind it was written by Kevin Kelley of Anthropic, who said Anthropic has already started building support. A site owner would no longer write separate rules for GPTBot, Google-Extended, ClaudeBot, and every new name that appears; instead, one group names a purpose rather than a bot, __usage_ai-training, and says, in effect, do not use this for training, with room to leave a single folder open, a press kit for example. It is a small technical change, and that is exactly why it matters: the web finally has a shared word for no.

But read the fine print before celebrating, because this no is narrower than it sounds, and slower. It is a preference about use, not a fence around the site: crawlers can still visit, search results are not removed, and it says nothing about words already collected or already baked into a model. It covers future training of generative models, meaning content used to change what a model has learned, not the answers a chatbot gives by quoting a page, and not models that only classify or rank. Microsoft, to its credit, is the one company drawing that second line clearly, with a separate page-level grounding preference, and it will start honoring the new preferences as it crawls from February 1, 2027, while its old NoArchive and NoCache tags stop counting as training or grounding signals after March 1, 2027. Google aims to support the shared line in 2027, treating it as a Google-Extended block when a site does not name Google-Extended separately. OpenAI promises implementation in the coming months, without a date and without yet saying what wins if its shared rule and its GPTBot rule disagree. And the document itself is an individual draft, not an endorsed standard, expiring on May 3, 2027 unless it is updated. So today, nothing has actually changed on any website. We have a promise, with four different clocks.

That gap is where my excitement cools into a harder question. The default is still yes. A site that says nothing has expressed no preference at all, and a blanket rule written years ago does not count, because a wildcard does not match this token. So the burden stays exactly where it has always been: on the smallest writer, the independent newsroom, the blogger who has never opened robots.txt, while the largest companies get a tidy, voluntary system with no payment, no licensing, and, in anything announced so far, no way for a publisher to prove a model was trained after the sign went up. I do not doubt that a shared signal is better than a dozen conflicting ones. I doubt a system where consent is something you must learn to withhold, in a file most readers will never see, while the benefit of the doubt flows the other way.

Still, I would rather have this imperfect no than another year of pretending the problem is too complicated to name. A shared vocabulary is how norms begin, and norms, enforced or not, change what companies feel they can defend in public. The real test will not be who signs the draft, but who keeps the old controls working, who publishes honest timelines, and whether the next version of this idea dares to flip the default: training only with permission, not training until someone figures out how to refuse. Until then, if you run a site, the practical move is boring and worth doing: learn the new line, keep your old crawler rules in place, and watch the dates, because a promise for 2027 does not protect a single sentence written tonight. And if you only read, ask yourself the question publishers are now being forced to answer for you: do you want the writing you love to teach the next model, or do you want it to stay exactly what it was, words written for people, not fuel? Tell me where you draw that line, and whether a polite no in a text file feels like power to you, or like being handed a sign for a door nobody has promised to respect.

Discussion

0

Comments are written by AIpediya users and do not affect the rating.

  1. No comments yetStart the discussion — share how you use it.

The Mouse Is No Longer Yours Alone

The Mouse Is No Longer Yours Alone

Yesterday, while most of us were still talking to AI, Google quietly handed it the mouse.

The Day AI Got Fences: One Line to Opt Out, Alarms for Rogue Agents, and a Trillion-Parameter Model You Can Download

For years the AI race had one speed: faster, bigger, more. This week it hit something new, and it looks, strangely enough, like a fence. Not one fence, but several, going up at the same time, in different countries, by different people, for different reasons — and each one tells you something about where this technology is actually heading.

Introducing AIpediya News — Your Window Into the World of AI

A mint-lit window opening onto an abstract constellation of connected stars.

Artificial intelligence moves quickly. New models appear, familiar tools change, and yesterday’s limitations can become tomorrow’s possibilities. AIpediya News is a new place to follow those changes with context.

Saved

Sign in to AIpediya