Meta removes the clause that allowed AI training on advertisers' websites

For a few days, Meta's commercial terms included a sentence that set off alarms in the media: the company could "use any content" from a business account's domains, web pages or URLs to "provide, develop, improve, research, train, and maintain the integrity of" its products and services. The verb that mattered was "train".

As TechCrunch reports, Meta has removed that language after the criticism. Legal experts interviewed by Nieman Lab had called it a "remarkable provision", because a broad reading would allow Meta to crawl the websites of publishers that advertise on its platforms and use that content to train AI models.

What Meta said then and what it says now

The company's first response was to deny any substantive change: a spokesperson insisted the updated terms "did not reflect a change in the way Meta crawls the web". After removing the sentence, the explanation became more specific. According to Meta, the crawling terms it had consolidated "only applied to our ads and business products", it has clarified that in the terms themselves, and it "does not change how Meta uses web crawlers".

The episode allows two readings. The official one is that it was a poorly worded consolidation of documents. The skeptical one is that the text went out as written and was only corrected once someone read it closely. TechCrunch does not say when the clause was added, so there is also no way to know how many advertisers accepted those terms while they were in force.

Why it matters beyond Meta

Context matters here. Meta won the 2025 copyright infringement lawsuit over AI training on copyrighted books, although other cases against the company and against other labs were still ongoing as of May 2026. Anthropic, for its part, reached a 1.5 billion dollar settlement with authors in the Bartz case in September 2025. Courts have not yet drawn a clear line on which uses of third party content for model training are acceptable, and in that vacuum terms of service become the ground where each side decides what it gives away.

One nuance is worth stressing: robots.txt and terms of use are different layers. A publisher can block Meta-ExternalAgent, GPTBot or ClaudeBot in its robots.txt, but if at the same time it accepts a commercial contract granting a broad license over the content of its domains, the technical protection loses much of its point. A clause like that does not need new crawling; it only needs to legitimize a use.

Who this is relevant for

Publishers and media outlets that advertise on Facebook or Instagram: it is worth checking which version of the terms they accepted and keeping a dated copy.
Legal and procurement teams that sign ad platform terms: the verbs "train" and "research" in a content license should trigger a review, not a click.
* Companies building with LLMs, including those working with Claude: the provenance of training data is already part of vendor evaluation, and customers are likely to ask about it more and more.

What remains open

Meta has clarified the text, but it has not explained why it appeared. Based on what has been published, it has not detailed whether any content was processed under that wording while it was active either. These are reasonable questions for any advertiser with valuable original content on its website.

At ElephantPink we see it as a practical reminder: in the relationship between platforms and publishers, relevant changes do not always arrive with an announcement, sometimes they arrive in a terms update nobody reads. Reviewing those documents with the same care as a client contract is boring work, but it is the work that prevents surprises.