A single sentence buried in Meta’s updated commercial terms would have turned every advertiser into a data supplier. It said Meta could crawl and scrape websites “associated” with the businesses that use its platforms, and use that content to “develop, improve, research” and “train.” Nieman Lab spotted it. The Verge picked it up. Within days the line was gone from the page.

The door was wider than it looked

Think about who counts as a business on Meta’s platforms. A local newspaper running Facebook ads for subscriptions. A games site with a Page. A tech blog boosting a post. All of them operate websites “associated” with that business relationship, which means the archive sitting behind the homepage was, on paper, in scope. Buying ad inventory would have doubled as a training-data license, granted by clicking through terms nobody reads.

That is not an edge case or a drafting accident you can wave off. It is the cleanest acquisition path anyone has designed for publisher text: no negotiation, no payment, no robots.txt to respect, because consent arrived through a commercial agreement the publisher signed for an unrelated reason. Crawler blocking at the network layer does not help when the crawler has a contract.

Deletion is not a commitment

So the language is off the page. Good. It also cost Meta nothing. There is no new clause saying Meta will not train on publisher sites, no carve-out for news domains, no mechanism a publisher could point to in a dispute. The absence of a permission is weaker than the presence of a prohibition, and the difference matters precisely because the pressure that produced the clause has not changed. Frontier models need text. High-quality, edited, continuously updated text is scarce and expensive, and publishers produce it by the ton.

The pattern worth watching is where this fight is happening now. Not in crawler headers. Not in court, at least not first. In boilerplate. Terms-of-service updates ship on a schedule, apply retroactively to existing relationships, and get read by approximately nobody at the businesses they bind. Scraping language in a commercial agreement is faster and cheaper than any licensing deal, and it only fails when a reporter happens to read the diff.

Which is what happened here, and it is a thin kind of safety. The clause survived internal review. It made it to a published page with an effective date. It was removed because of coverage, not because of a process that catches this before publication. Next time it might be phrased less plainly, or split across two documents, or placed in a developer agreement nobody thinks to check.

For publishers, the practical takeaway is unglamorous: your training-data exposure is now a contracts problem, and whoever handles your ad buying is making AI policy for your newsroom whether they know it or not. Somebody should be reading every platform terms update line by line, the same way you would read a syndication deal.

Meta’s new commercial terms still take effect on October 30. What the page says on that date is what binds.