Brazil Meta AI Training Injunction: São Paulo Court Bars Scraping of News Sites

The Brazil Meta AI training injunction suspends Meta's Commercial Terms for the country's press associations, bars scraping of news sites for Meta AI and sets a daily fine — the second court order against the terms in a week.

The Brazil Meta AI training injunction landed 17 days before the rule it blocks was due to take effect. On Tuesday 3 November, a São Paulo judge suspended part of Meta’s new Commercial Terms for Brazil’s newspaper, broadcaster and magazine associations, barring the company from collecting their journalism without permission — including to train its AI systems — under a daily fine of R$1 million, capped at R$100 million (FourWeekMBA).

Information verified through 5 November 2026.

What the Brazil Meta AI training injunction actually orders

Judge Guilherme Madeira Dezem of the 44th Civil Court of São Paulo granted the preliminary injunction to three associations: ANJ for newspapers, ABERT for radio and television broadcasters, and ANER for magazine publishers (PPC Land). The order has teeth in four places. It suspends the contested clause of the updated Commercial Terms, which had an announced start date of 20 November 2026. It bars Meta from using scrapers, bots, spiders, crawlers or test credentials to access, index, cache, analyse, collect, retrieve or crawl member outlets’ domains — and from using what it gathers to “provide, develop, research, maintain, train, fine-tune or feed” Meta AI. It requires Meta to keep members’ accounts, pages and business tools on Facebook, Instagram, WhatsApp and Threads fully active, with no penalty for refusing the terms. And it sets the daily penalty that gives the order its headline number (PPC Land).

The judge ruled without hearing Meta first. His reason was timing and irreversibility: once reporting has been ingested into a model, the material “can hardly be segregated, audited or removed afterwards” (PPC Land). Meta has five days after notification to contest the measure, and the injunction stands until a final judgment (FourWeekMBA).

One detail deserves attention. The original wording of the terms, published on 16 October, expressly used the word “train.” After an immediate negative international reaction, Meta revised the text on 19 October and deleted that word — while leaving the operative verbs and the extraction machinery intact, the judge wrote (PPC Land). The court treated the edit as cosmetic. The crawl permission in the current preview still lists seven actions a business authorises, from accessing to crawling linked domains (FourWeekMBA).

Why the Meta Commercial Terms Brazil clause failed on copyright and contract

The decision rests on two legs. The first is copyright. Article 29 of Brazil’s Copyright Law requires prior, express authorisation for any use of a protected work, including reproduction, inclusion in a database and storage on a computer. Scraping a news archive, breaking it into tokens and storing it to train a large language model is, in the judge’s words, “undeniably an act of reproduction and use of protected intellectual works” — not the transient indexing that search engines do to send readers to a page (PPC Land). Brazil, the decision notes, has no general text-and-data-mining exception that would permit commercial AI training of this kind.

The second leg is contract. The Commercial Terms are a contract of adhesion — drafted by one side, offered on a take-it-or-leave-it basis. Requiring outlets to hand over their content “free of charge and universally as a toll for remaining connected to the channels through which today’s public consumes news constitutes a manifest abuse of rights,” the judge wrote (FourWeekMBA). There is a trap inside the clause, too: outlets must warrant that they hold every right needed, yet their pages carry agency photographs and contributors’ work licensed only for publication. Honour the warranty and they breach other people’s rights; refuse, and they lose the platforms their readers use (PPC Land).

The court also swept aside Meta’s California forum clause, holding that a Brazilian company earning substantial advertising revenue from Brazilian traffic answers in Brazil (PPC Land).

The Brazil media associations ANJ ABERT ANER did this as a coalition — and Norway moved first

Scale is the story here. ANJ, ABERT and ANER together represent thousands of broadcasters, hundreds of newspaper publishers and dozens of magazine publishers across every Brazilian state, and the court let the associations sue for all of them without individual authorisations (PPC Land). ANJ called the ruling the first known decision granted to a coalition of media associations against Meta’s new terms (Oossa).

It was not, however, the first court to move. An Oslo court temporarily blocked the same terms on 28 October, after a case brought by ten Norwegian media groups — six days before São Paulo acted (Oossa). Britain’s competition regulator opened its own investigation into the terms on 30 October. Three jurisdictions, three different legal theories — competition in London, and in Brazil copyright plus contract — converging on the same clause within a single week.

ANJ’s executive president Marcelo Rech welcomed the order as “a very positive sign that there are limits to the abuse of economic power” (FourWeekMBA). Meta had not responded to the Brazilian ruling in the reporting available; on 28 October its communications vice president Andy Stone said the company did not plan policy changes without a full consultation with industry groups (FourWeekMBA).

Analysis: the injunction is preliminary, and its protection stops at the associations’ membership — other Brazilian publishers are not covered, and a final judgment could differ (Oossa). But the reasoning travels. Deleting the word “train” did not save the clause, because the court read the verbs, not the label. Any platform drafting AI terms now knows that a court may treat the crawl itself as the taking, and may refuse to accept “leave our services” as a real choice for a newsroom whose audience lives on those services. That is a harder standard than a robots.txt file — and the judge dismissed robots.txt as a mere technical courtesy, not a legal answer.

Photo: ProtoplasmaKid, via Wikimedia Commons (CC BY-SA 4.0)

Sources

THE READER CONVERSATION

What’s your perspective?

Bring a question, a different reading or something the story made you notice. All comments are reviewed before they appear. Disagreement is welcome; harassment, advertising and personal attacks are not.

No approved comments yet. Start the conversation.

Comment as a guest and verify your email, or sign in or create a reader account. Guest comments reach our review queue after email verification; account holders can submit directly for review.