On Sunday, August 16, the election campaign began in Brazil and, with it, the most ambitious artificial intelligence regulatory experiment Latin America has attempted began running on a real national election. It is not a law: it is Resolution 23.610/2019 of the Superior Electoral Court (Tribunal Superior Eleitoral), updated in 2024 and 2026, and it requires all campaign material created or altered with AI to declare it in an “explicit, prominent, and accessible” way.
The detail matters because the format of the notice changes depending on the medium: identification at the start in audio, a watermark plus audio description in images, a label in video, and a note on each page in print. The resolution also prohibits deepfakes intended to disinform (synthetic videos or audio that make someone say or do something they never said or did) and prohibits an AI system from recommending candidates to users. It also closes a window from 72 hours before to 24 hours after the vote for new synthetic content featuring a candidate’s image or voice. Platforms must have a compliance plan, a reporting channel, and immediate removal; fines range from 5,000 to 30,000 reais, in serious cases they extend to the cancellation of registration or loss of office, and in certain circumstances the burden of proof shifts to whoever published the content: it is not the victim who must prove the fabrication.
What is interesting for the region is the route chosen. Brazil did not wait for its AI legal framework (the PL 2338 bill remains stuck in the Chamber) and regulated through an electoral court resolution: fast, enforceable, and limited in scope. What remains to be seen is where it breaks, because the obligation to label works when there is an operator to negotiate with, and against an open-weights model that anyone can download to their computer there is no removal order to enforce. Chile, with its bill still in the legislative process, and Mexico, which has just handed the matter to its electoral institute, now have a real case to look at instead of a European model to cite.
Also today
- Stripe closes the purchase of OpenRouter for more than $7 billion — The layer that decides which model each query goes to, and that allows routing by cost, stops being neutral: it becomes an asset of the payments company that already bills half the region. It is five times the valuation of its May round and below the $10 billion discussed in July; neither party has officially confirmed it.
- Dario Amodei responds to Wall Street: the backlash against AI is “fundamentally a crisis of trust” — It is the first time that pressure on the risk narrative comes from capital and not from the regulator. Anthropic’s CEO conceded that “we have not yet delivered on our big promises.”
- Between 20% and 25% of Peruvian workers have been exposed to automation or AI — The mechanism it describes is the hardest to measure: companies do not cut salaries, they leave vacancies unfilled. The adjustment does not show up as layoffs or as a drop in wages; it shows up as nothing.
- The AI manager of a store in San Francisco fired her first human employee — The worker was late for 17 of 23 shifts, but the agent had to be reminded of the policy it had written itself months earlier and had forgotten. Also on its record: 1,000 toilet seats purchased and a lost staff payroll.
- Alibaba’s new open model took 21 minutes to draw a pelican — It spent 22,276 units of text reasoning. It comes set to maximum effort by default: asking it for a circle produces minutes of deliberation and an animated circle nobody asked for. It is a good idea to lower its effort before evaluating it.
In the region
Brazil’s second story of the day is the flip side of the first. The first application of the new digital statute against a global platform already has a response: Discord asked for the suspension of its live streaming to be revoked and challenged the data authority’s competence to order it. The underlying argument is not about the facts but about authority: the company maintains that indefinitely suspending a feature is a sanction that only a court can impose, asks for 15 business days plus another 15 instead of the three the agency gave it, and asks whether the order extends to end-to-end encrypted calls. The deadline expires today. If the authority upholds the measure, a precedent is set in the region that none of its neighboring counterparts (the Colombian, Argentine, Uruguayan, and Chilean authorities) currently has the power to replicate; if it gives in, what is set is the opposite. Outside Brazil, today’s window brought no new publications from multilateral organizations or from the ministries and data authorities of the rest of the region.
Threads we’re following
This adds to something we had been following: a little over a week ago we noted that nine of the ten best video models in the world are Chinese and that many of them can be downloaded freely. The two stories meet right at the weak point. A labeling regime like Brazil’s relies on there being someone to demand the mark from (a platform, a provider, an account), and that chain exists as long as the video is generated on a service with an owner. The day a campaign’s synthetic material comes from a weights file that someone downloaded to their computer, the fine still applies to whoever publishes, but removal and labeling at the source no longer have anyone to be directed at.
Brazil has just set electoral AI rules running by administrative resolution while its AI law has been stuck in Congress for two years, and in the same month its data authority shut down a feature of a global platform without going through a court. If the fast route works and the legislative one does not advance, are we looking at a regional governance model made of resolutions by technical bodies, effective but without parliamentary debate, and what happens the day that same instrument is used by a government we don’t like?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.