The opposite of America's AI problem is happening in Brazil
While the US debates whether to regulate AI at all, Brazil has built the most detailed AI-and-elections rulebook of any democracy and the gap between the two is becoming a headache for companies
October 2013. I’m walking the halls of Brazil’s National Congress, laptop in hand, translator at my side, showing lawmakers how to get more engagement on their Facebook pages. Our policy lead had just started — his first week, he testified before Congress, then called me. Brazilian lawmakers didn’t want to meet with a Facebook policy staffer to talk about the internet bill they were debating, the Marco Civil da Internet. But if I could show them how to use Facebook and read their engagement metrics? That got us in the door.
One minute I’m explaining click-through rates to a senator. The next, he’s grilling me about Facebook’s stance on privacy.
I’ve been to Brazil twelve times since then. I was there for the 2018 election, when hate speech spiked from the candidates themselves and we had to decide, in real time, whether politicians should be exempt from fact-checking penalties. I was there in 2022, in the months before Bolsonaro faced Lula, when Brazil’s electoral court came close to banning Telegram — not because the company refused to cooperate, but because its compliance emails were landing in an inbox nobody was checking.
That’s the Brazil I know. A country that’s been negotiating, suing, and occasionally threatening to ban its way through a relationship with tech platforms for over a decade. So when I look at what’s happening there right now, ahead of October’s election, I don’t see a country improvising. I see the result of thirteen years of building toward this, and it’s arrived at a level of specificity the US hasn’t come close to.
The private room for senior practitioners in tech, politics and democracy
Brazil is exactly the kind of case I dig into twice a month in the Anchor Change Briefing Network — the international signals that aren't getting enough English-language coverage, and what they actually mean for companies and practitioners operating in this space. Founding seats start at $400/month.
Once they’re gone, pricing moves to $500.
Where things stand. In May, Brazil’s electoral court created a permanent commission on AI in elections, with a 90-day clock to build a national catalog of enforcement tools before October, and a 30-day deadline for regional courts to stand up their own cybersecurity units. The underlying rules are already detailed: deepfakes in campaign material are banned outright, AI-generated content has to be labeled, AI tools can’t rank or recommend candidates, and there’s a blackout on AI-altered content from 72 hours before voting until polls close. A separate analysis by the Observatório de IA nas Eleições found 18 AI-generated political profiles active in Brazil between January 2025 and April 2026 — 11 of them with no AI disclosure at all, and 14 spreading misleading claims about politicians or democratic institutions.
At the same time, Lula signed decrees overhauling the Marco Civil da Internet — the same law I was walking Congress through in 2013 — requiring platforms to proactively remove deepfakes targeting women without a court order, and take down non-consensual intimate images within two hours of notification. Days later, Brazil’s Supreme Court opened, then suspended, its session on platform civil liability, with Justice Toffoli pushing back on big tech’s argument that new obligations should only apply after appeals are exhausted. Two days after that, the court gave platforms 60 days to implement the expanded framework.
One of the challenges with all these new rules: AI companies don’t yet have guidance on what “ranking” actually means under the TSE’s rules. Read literally, even listing candidates alphabetically is still a form of ordering. With fines and potential shutdowns on the table, the rational compliance choice isn’t to find the line — it’s to turn off AI summaries and responses on political topics entirely. However, a recent study shows that many AI tools are still ranking candidates - showing the challenges in operationalizing these policies.
In 2022, Telegram was briefly banned in Brazil because the court’s compliance emails were going to an outdated address nobody was checking. The ban was lifted within days, once Telegram fixed it. In 2026, AI companies may go dark on Brazilian politics because the rules are too detailed to interpret with confidence. Same outcome — less information reaching voters — from opposite directions. One was a capacity failure. The other is a caution failure. Brazil spent over a decade closing the first gap. It’s now running into the second.
What I’m watching. The major AI companies like OpenAI and Meta have said, publicly, that they’re paying attention to Brazil. They should be. Brazil moved to electronic voting in 1996 and has been a model for the region on election security ever since, and the TSE is not a regulator improvising under pressure — it’s an institution with a long track record of being taken seriously. Brazilian voters know it too: 51.6% name the TSE as the body responsible for policing AI in this election, ahead of the federal government, the platforms, or citizens themselves.
But Brazil’s approach is the opposite of what’s happening in the US, where there’s been little appetite at the federal level to govern AI at all, and where the federal government is now willing to intervene directly in how AI companies operate, for reasons unrelated to elections. On June 5, the US Trade Representative (USTR) issued a Section 301 determination against Brazil covering its digital trade practices — formalizing a trade-pressure campaign that specifically targets secret Brazilian court orders compelling US platforms to remove content and suspend accounts, sometimes globally, without notifying the account holder. So Brazil is, in effect, being asked to choose between its regulatory model and its trade relationship with the US. That’s the same fight playing out in Europe right now, with far less attention — the push for “less censorship” on one side, and a US government newly comfortable using trade and regulatory leverage on AI companies operating internationally, for whatever reason it decides matters that week.
That puts companies in a genuinely hard spot. The posture that makes sense in a vacuum — wait and see, move fast, figure it out as you go — is the posture that gets you close to a ban in Brazil. The posture that makes sense in Brazil — go dark on anything ambiguous — would be a strange way to operate in a market where regulators haven’t asked you to, and now risks reading as exactly the kind of “censorship” the US is using as leverage.
I don’t think there’s a clean answer here, and I’m skeptical of anyone who says there is. There’s no controlled experiment where Brazil tried clearer rules and we can compare outcomes. What I keep thinking about is the gap between 2013, when getting a meeting meant teaching a senator how Facebook worked, and now, when the rules are so detailed that the safest move is to not show up at all. Both are versions of the same problem: the rules and the reality of how these companies actually operate are not talking to each other.
Between now and October, I’ll be watching to see whether the TSE provides more specific guidance on what “ranking” actually means and whether it takes a single enforcement action against an AI company. I’ll then be watching to see how the Trump administration responds.