Meta Drops 'Train' From Its Terms — The First Frontier-Lab Concession Reshapes the Data Bill

The Oct. 16 draft expanded Meta's authorization for businesses using Facebook, Instagram, WhatsApp, Ads Manager or Meta Business Suite to consent to automated crawling of any associated domain — and gave Meta the right to use that scraped content \"to provide, develop, improve, research, train, and maintain the integrity of, the Meta Company Products.\" The single word \"train\" was the only hook for AI training in the document, but it was enough. By October 17 the clause had been rewritten: the \"train\" verb was deleted and the permitted use was narrowed from \"Meta Company Products\" to \"advertising and business products.\"

On June 25, 2025 Judge Vince Chhabria ruled in Kadrey v. Meta Platforms that training Llama on lawfully acquired books was fair use — a procedural win Meta has leaned on ever since. But the ruling carried an asterisk: market harm under the fourth fair-use factor could tip future cases the other direction. That asterisk is now load-bearing. Anthropic won the parallel fair-use question in Bartz v. Anthropic on June 23, 2025, then paid $1.5B — the largest U.S. copyright settlement on record, approved July 20, 2026 — for the acquisition of those books through LibGen and PiLiMi. The message that has traveled from both rulings is that the training itself can be defensible, but the chain of custody that delivers the training corpus cannot be sloppy, and the licensing market that publishers say is being undercut cannot be ignored.

Read narrowly, the Oct. 17 edit is cosmetic. Meta still gets the right to crawl every website a business account touches and to use what it finds for product development, advertising recommendations and integrity/security. Read broadly, the edit does three jobs at once. It reduces the legal surface area for a fresh class-action theory that \"Meta promised in writing it could train on our pages.\" It signals to enterprise advertisers — a million-plus accounts on WhatsApp and Messenger alone — that Meta is not asking them to waive AI-training rights against their own domains. And it gives Meta room to claim it was never trying to use the commercial terms as a workaround for licensing markets. The hedge costs Meta nothing commercially, and saves it a litigation category.

Every rung of the AI training-data supply chain reprices when the largest consumer of that data publicly concedes the language. Upstream, Reddit and News Corp gain leverage: Reddit's Q2 2026 data-licensing revenue of $43M (+24% YoY) is a leading indicator for the contract values labs will accept, and News Corp joined a May 2026 publisher class action against Meta that now has stronger optics. Smaller data licensors — Shutterstock and Getty Images — sit further out on the risk curve but get the same tailwind when frontier labs move from \"scraping\" to \"paying.\" Downstream, Microsoft carries the heaviest disclosed exposure through its OpenAI partnership as the NYT summary judgment approaches, Amazon carries the Anthropic exposure via its multi-billion-dollar investment, and Alphabet is the rare platform competitor whose Gemini stack leans on licensed Reddit data and explicitly honors robots.txt in customer contracts.