Meta quietly removes AI-training language from its US advertiser terms, but not everywhere

On October 17, The Verge reported that a sentence had disappeared from Meta's business terms page. It allowed the company to use the content of websites linked to its business accounts to, among other things, "train" its products. But it did not disappear everywhere: the next evening, the UK version of the same text still contained it.

The story started with Nieman Lab, the media-focused site of Harvard University's Nieman Foundation. On October 15, it revealed that Meta's new business terms, which take effect on October 30, allow its crawlers to scan the "domains, web pages or URLs" associated with Facebook, Instagram or WhatsApp business accounts. The text also allows Meta to use that content to "provide, develop, improve, research, train" its products. We detailed what this means for advertisers and website publishers.

Meta calls it a simple clarification

Three legal experts consulted by Nieman Lab said the text would let Meta argue that it needs no license to train its models on news sites. Attorney Matt Topic, of the firm Loevy & Loevy, saw it as access to content that would cost many millions to license, although he considered it uncertain that a court would side with Meta.

Meta first pushed back. On October 16, its spokesperson Andy Stone wrote on X that nothing was changing in how the company uses its crawlers. On the 17th, according to The Verge, the sentence about training had disappeared. Another spokesperson, Daniel Roberts, told the US outlet that these rules applied only to Meta's advertising and business products, and that the company had "clarified" this in the text. According to him, the update does not give Meta the right to ignore the robots.txt file, which tells crawlers which pages they may visit.

Which version will apply on October 30?

The site Elsop compared the English versions. In the US version captured on October 17, the permission to crawl now targets "advertising and business products" and the security of Meta's services. The list of uses still includes "develop, improve, research," however, without explicitly excluding training. Elsop draws no legal conclusion from this.

The UK version says something different. According to Elsop, it still contained the full sentence on October 17, and according to the specialist site PPC Land, again on October 18 at 8:44 p.m. (UTC). PPC Land found neither a limitation to advertising products nor any mention of robots.txt. We were unable to check the French version. Elsop also notes that the self-serve advertising terms, in force since August 2025, already allow Meta to crawl an advertiser's content for "research and training" when the advertiser places an order.

What advertisers and publishers can do

There is no box to check to opt out: according to the text, the only way out is to stop using Meta's products for commercial purposes. That leaves the robots.txt file, which Meta says it respects.

• Before October 30, review the version of the terms that applies to your account and keep a dated copy.
• In its documentation, Meta describes the meta-externalagent crawler as the one that collects data for, among other things, training its AI models, and meta-externalads as the one for advertising products. robots.txt rules apply crawler by crawler: blocking one does not stop the other.
• According to the same documentation, two other Meta crawlers may bypass robots.txt: the one that fetches a page at a user's request, and the one used for security checks.

Meta has eleven days left to publish identical text in all its versions. Until then, one question remains: on October 30, which sentence will apply to European advertisers?