Meta's new business terms claim the right to scrape publishers' sites
Two things worth your attention: a quiet rewrite of Meta's commercial terms that legal experts read as a scraping license, and IBM's attempt to fold chat, decisions and speech into one open endpoint.
Meta's updated commercial terms of use, effective October 30, authorize the company to run scrapers, bots and crawlers across websites "associated with" business accounts on its platforms — and give it the right to create "test credentials" to get in. The clause, which Nieman Lab surfaced after a reader tip and had three legal experts review, grants Meta permission to "access, index, cache, analyze, collect, retrieve, or crawl" those domains and to use what it collects to "provide, develop, improve, research, train, and maintain" Meta products. Because Meta's own terms define "AI at Meta" as one of those products, the plain reading is that any publisher with a Facebook page or an Instagram profile has just licensed its archive for model training.
There is no opt-out. The terms say only that users who disagree should "stop all access or use of our products for business or commercial purposes" — which, for a newsroom that depends on Meta for distribution, is not a choice. Media attorney Matt Topic called it "a remarkable provision to bury in terms of service," noting the material would cost "many millions of dollars in a licensing market," while Cornell's James Grimmelmann said the "test credentials" language would technically let Meta sign up as a fake user to get past registration walls. Wisconsin law professor Benjamin Sobel frames it as tying in the antitrust sense: accept the business tools, accept the crawl.
Meta's response is narrower than the text. A spokesperson said the crawling language was previously scattered across separate terms and this update consolidates it, that it does not change how Meta crawls publishers, and that sites can still manage access through robots.txt. The company declined to answer Nieman Lab's questions about whether test credentials would bypass paywalls. Our take: read the document, not the comment. Consolidating scattered permissions into one unconditional clause is not housekeeping — it converts a technical signal that any crawler can ignore into a contractual argument Meta can make in court, and it lands while a class action from five book publishers accusing Meta of pirating millions of books is still working its way through federal courts. We looked at the other side of this fight in September — Unsealed NYT filings: Microsoft called AI scraping 'theft of labor'.
IBM released Granite Switch 4.1, a pair of open models at 8 billion and 30 billion parameters that pack a chat model, a decision adapter, a safety classifier and a speech-to-text encoder behind one endpoint at roughly the memory cost of the base model alone. Each checkpoint composes the granite-4.1 base with a general-purpose decision adapter, a Guardian safety adapter that IBM says beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on average across safety-decision tasks, and Granite Speech 5.0 TurboCTC, a 470-million-parameter transcriber IBM bills as the fastest ASR model on the OpenASR leaderboard. The mechanism is that the adapters are activated LoRAs — they reuse the base model's KV cache and switch by control token, so a caller changes task without changing code or paying for a second model's worth of VRAM.
The claim carrying the most weight is one IBM cannot easily benchmark: its 3B Switch model made 60 to 100 decisions per second while playing Doom, against Jev's reported 10 per second. Treat that as vendor framing until someone outside IBM replicates it. The models carry a 128K context window and the model cards list an Apache 2.0 license — the opposite end of the licensing spectrum from the terms described above, which is the exact check we wrote up in How to — check a model's license before you ship on it.
What to watch: whether a publisher tests Meta's clause in court before October 30, and whether an independent run reproduces IBM's decision-speed numbers.
If a platform can fold a scraping license into the terms you must accept to reach your audience, is consent still a meaningful idea? Tell us in the comments.
