Meta's New Terms Could Let It Train on the News

Buried in a routine update to Meta's business terms is a line that could reshape how the company feeds its AI. Starting October 30, business accounts that keep using Facebook, Instagram, or WhatsApp will consent to Meta crawling the websites linked to those accounts, and the terms say Meta may use what it collects to train its products. Nieman Lab had three lawyers read the language, and all three came away with the same worry.

The catch is who holds those business accounts. Almost every news organisation on the planet runs at least one, to post stories and buy ads. Nieman Lab pointed out that it has one itself. So a clause framed as a housekeeping change to commercial terms could, read plainly, hand Meta permission to scrape and train on the archives of the very publishers it competes with for attention.

Consent by inertia

What makes this different from an ordinary data grab is the mechanism. Meta is not asking publishers to opt in. It is treating continued use of its platforms as agreement, the way a software update slips new permissions past anyone who clicks through without reading. For a newsroom, walking away from Facebook and Instagram is not a realistic protest. That is the leverage.

Matt Topic, a lawyer at Loevy & Loevy who has litigated copyright cases, told Nieman Lab it is not clear a court would uphold the terms, but that publishers should assume they are enforceable and act accordingly. It is cautious advice with a sharp edge. If you wait for a judge to decide, the training has already happened.

A wider fight over training data

Meta has been here before. A leaked list reported earlier this year claimed the company had scraped tens of thousands of top-ranked domains for AI training while ignoring the robots.txt files sites use to wave crawlers off. Meta called that report bogus but did not fully rule out that the outlets named had been used. The new terms, by contrast, are not a leak. They are the company writing the permission down in its own contract.

The timing lands in the middle of a legal season that has not been kind to AI labs. Gannett, the owner of USA Today, recently sued OpenAI over the use of its journalism, and European regulators have begun exercising real enforcement powers under the AI Act. Against that backdrop, writing a training licence into standard business terms looks less like an oversight and more like a calculated bet that most publishers will never read the fine print, and could not afford to leave even if they did.