"A Deleted Clause in Meta's Terms Shows Where the AI Data Fight Really Lives"

Meta's latest revision to its commercial terms of service was mostly unremarkable — right up until the line about AI training. The provision described the content Meta could use to train its models in language that covered "domains, web pages, or URLs," and legal experts quickly read it as broad enough to override a site's own robots.txt preferences. Nieman Lab reports that Meta deleted the specific clause and called the episode a misunderstanding, but the experts' conclusion is the interesting part: even after the language change, the revised terms still grant broad scraping permissions.

The reason this matters has little to do with any one company. The underlying legal question — whether scraping public web content to train a model counts as fair use, infringement, or something in between — is still genuinely unsettled in most jurisdictions. Because there's no clear rule at the level of law, the contest over who gets to use the open web's data has quietly migrated into terms-of-service fine print, the one place where both sides can stake a claim without waiting for a court to decide. That's a striking shift: the most consequential debates about AI's raw material are now being conducted in documents almost nobody reads.

The second thing worth noticing is the gap between what a terms-of-service clause says and what it can actually do. A "permission" written into a ToS binds the people who agreed to that document — it is not a property right over third-party websites that never signed anything. And robots.txt is a decades-old convention, not a statute; honoring it is a norm, not an obligation in most places. So a clause granting "broad scraping permissions" is as much a signal as a legal lever. It tells you how a platform interprets its own rights more than it creates an enforceable claim over the sites it crawls. The fight is real, but the weapons are surprisingly flimsy.

There's a balanced way to read Meta's side here. Drafting terms broadly by default and narrowing them only under public pressure is a well-worn pattern across the whole industry, not a quirk of any one company — and "this was a misunderstanding, nothing changed" is a standard (and sometimes honest) corporate response. What's genuinely useful is what the episode reveals about posture: the default is maximal data access, with carve-outs added reactively when people notice. That default, more than any single clause, is what's actually being negotiated every time a terms-of-service update ships.

The takeaway isn't that anyone should panic or pick a side. It's that the AI data debate has a new front page, and it's written in legalese. Further reading: TechCrunch's account of the provision's removal, including Meta's statement that the change did not alter how it crawls the web.