A new AI training opt-out could let websites tell several major technology companies, in one place, that their pages should not be used to train generative AI. OpenAI announced its backing on 30 October 2026. For publishers, artists and anyone running a website, the attraction is obvious: a single instruction could replace a collection of separate requests to different companies.
But it is not an automatic legal right, and it is not a switch that immediately stops every AI company collecting content. The proposal is voluntary, its technical specification is still a draft and participating businesses are at different stages of implementing it. Those distinctions matter more than the headline promise.
Why website owners want an AI training opt-out
A news publisher might happily allow readers to find articles through search while objecting to those same articles being incorporated into the training of a commercial model. A photographer might want portfolio pages visible to customers without inviting a crawler to treat the pictures as free training material. A small business could have an entirely different view, welcoming both discovery and AI reuse. The new approach is designed to communicate a preference, not prescribe which choice a site should make.
Until now, a webmaster wishing to communicate a restriction often had to keep track of individual crawlers and write separate instructions for each provider. That becomes cumbersome as new systems appear, crawlers change names and companies distinguish between search indexing and model training. In its 30 October announcement, OpenAI said it intends to respect a common signal alongside its existing controls. It also acknowledged that implementation would take place over the coming months, not that it had completed the work that day.
A shared signal would make the process simpler for a local newspaper or independent creator without a dedicated technical team. Whether that simplicity translates into effective control depends on who adopts it and what they actually do when a site expresses a preference.
The small website file at the centre of the plan
The proposed system uses robots.txt, a familiar file that websites use to tell automated visitors how the owner would like their pages handled. Its location and general purpose are well established, but robots.txt should not be confused with a lock on a door. It communicates instructions to software; it does not physically prevent a non-compliant visitor from requesting an accessible page.
The published Internet-Draft proposes a special user-agent policy token, written as __usage_ai-training, using the same Allow and Disallow patterns that webmasters already encounter. A site could use that group to express an objection across its pages, or permit particular areas such as a press kit. This is a proposed machine-readable convention, not an announcement that every crawler on the internet now understands the instruction.
The document does not invent a completely new language for robots.txt. That is deliberate: a familiar format reduces the learning burden and may make it easier for website platforms to incorporate the setting into their existing tools. It still requires accurate configuration. A mistyped group, conflicting rule or assumption that another control has been inherited could undermine the owner’s intention.
Training and search are not the same use
A crucial difference sits behind the word training. A search engine may fetch a page to make it discoverable. An AI product may use information retrieved from a page to answer a current question. A model developer may use a collection of documents to change a model’s underlying behaviour during training. Those activities can overlap technically, but they are not interchangeable.
The shared proposal discussed by OpenAI concerns foundational generative AI training. A webmaster should not assume it automatically prevents every search excerpt, generated answer, quotation, download or unrelated form of reuse. Nor does a preference settle legal disputes about copyright, licensing or exceptions under different national laws. That is why the existing debate over publishers and AI-driven search traffic remains relevant even if the new signal becomes widely recognised.
In particular, a publisher wanting to remain visible in conventional search should check the implications of each setting rather than applying a blanket block without understanding its effect. A technical preference and a commercial licensing agreement are also different things. One tells a crawler what the operator wants. The other sets out obligations between parties, where an agreement exists.
Which companies will actually listen?
The significance of a common approach comes from participation, not simply from the elegance of its syntax. OpenAI’s decision is evidence of support, but its published timetable explicitly leaves implementation ahead. Other providers may use different dates, technical interpretations or complementary controls. Before claiming a policy works across the industry, website owners should look for individual implementation notices and test against the actual behaviour of those systems.
The Internet Engineering Task Force process is intended to make technical standards open to discussion and refinement. The current document is an Internet-Draft, which means language can change. Website operators should therefore avoid treating examples as permanent instructions without checking later revisions. There is a gap between a widely discussed standard and a completed standard, and another gap between publication and adoption.
These practical limits do not make the proposal meaningless. Reducing the number of different signals a small publisher has to maintain could be genuinely helpful. It is simply premature to say that an opt-out has become universal or that a crawler ignoring a preference has automatically violated a new law. The separate arguments surrounding AI training and copyright permission will continue.
What a sensible website owner can do now
The immediate task is to decide which uses of the site are welcome. A publisher may value search referrals, wish to restrict model training and be open to paid licensing. A photographer may care more about use of individual image collections. Those are choices to define before changing a technical file. A rule copied from somebody else’s website may make a poor substitute for a policy that reflects the owner’s own interests.
Next, keep an accurate inventory of current crawler instructions and review the official documentation of each provider that matters to the site. If adopting the proposed common signal, record when it was added, what paths it covers and which providers publicly say they honour it. Website monitoring may help detect unexpected traffic, but access logs alone cannot establish whether a model was trained on a page.
The real advance is the possibility of a common language for expressing choices. The unresolved question is how consistently AI companies will respect that language. For creators whose work is accessible to everyone with a browser, the difference between those two outcomes is everything.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.