It is a small textual change, possibly legally meaningless — and still an admission. The Verge reports that Meta removed language from a policy document that critics said gave it cover to scrape news sites. Publishers who have spent two years arguing over exactly those words read the deletion less as a concession than as the tidying away of evidence.
Both readings can be true at once. That is the problem.
What changed, and what did not
A sentence in a terms-of-service document is not a crawler. As long as the requests keep arriving, deleting a line changes nothing about what publishers see in their server logs. Technical access control and legal permission are two different things, and they are constantly confused: a page that is publicly reachable is not thereby free to harvest. A robots.txt entry is a request, not a lock. Terms of service are a contract that only someone who can afford to litigate will ever enforce.
That is the entire economics of the training-data market in one paragraph.
How publishers broke out of silent consent
For years the rule was: whoever does not object gets read. That has flipped in the past two years. Cloudflare moved to pay-per-crawl in the summer of 2025 and blocked by default — a step that created a market where previously there had only been forbearance. Reddit and Stack Overflow sold licence deals and demonstrated that content has a price. The New York Times is suing OpenAI not over technique but over copies. Each of those moves shifted the burden of proof from the publisher who must ask permission to the model builder who must explain what it is standing on.
Against that backdrop, Meta's wording was not a detail. It was one of the places where a large provider disclosed the legal theory under which it collects public web content. Remove it and the behaviour remains — just without the rationale you could pin it to.
The sceptical reading is the correct one
You do not need to be a cynic to notice that voluntary self-restraint rarely arrives unprovoked. Companies tend to delete wording precisely when that wording could be used against them: what was never written does not have to be defended. Without the internal reasoning, no one outside can tell whether the deletion announces a change in conduct or merely cleans up a liability.
That is not an accusation, it is a description of incentives. And it makes unavoidable a demand publishers have been making for years: not promises, but behaviour that can be checked.
What can actually be checked
Publishers can measure far more today than three years ago. Server logs show which user agent arrives and how often. Cloudflare and others offer bot management and paid gates for crawlers. Anyone who wants to know whether something changed does not need to read the terms — they need to count the requests. A company that genuinely stops shows it in traffic statistics, not in fine print.
That is the practical consequence of this week: the argument moves from what is written in contracts to what servers measure. And therefore to the question of who controls the measuring.
Why this goes beyond Meta
Every provider training a model on web data has the same balance sheet: benefit from public content, cost borne by the people who made it. That balance sheet was squared for years by forbearance and by the assumption that the law was murky enough to ignore. Both foundations are crumbling. Forbearance ends at the paywall, and the law is being settled by cases that take years and land differently in different jurisdictions.
For readers this is bad news. If news content reaches models only through licensing deals, then curation migrates to whoever can pay. Small and specialist newsrooms — the ones with the least bargaining power — drop out of the training set first.
What to watch
Whether Meta's actual crawl traffic declines measurably, or only its prose did. Whether other providers quietly adjust their terms in the coming weeks — a pattern that would suggest either coordination or a wave of legal advice. And whether regulators finally close the gap between robots.txt and copyright, instead of leaving it to courts.
Until then: if you want to know what a company does with the open web, do not read what it writes about it.
