All twenty members of the G20 unanimously backed the Carolina Principles, the light-touch artificial intelligence regulatory framework that the United States had been pushing since Monday: United States reaches an agreement with the G20 to promote more flexible AI regulation. The announcement was made by U.S. Commerce Secretary Howard Lutnick at the close of the technology and innovation summit in Chapel Hill, North Carolina. What was a request yesterday is a signed agreement today, and three of the signatures are Latin American.
The text commits the signatories to reserving new regulation for genuinely novel considerations, preferring sectoral approaches over regulatory bodies dedicated to AI, directing public funds to fundamental research and opening commercial opportunities to emerging technologies. It recognizes each country’s sovereignty and calls for flexible application according to national circumstances. The tension appears when you look at the legislative calendar: Brazil, Mexico and Argentina signed a commitment not to create dedicated AI authorities in the same month in which three legislatures in the region are debating precisely that institutional design—PL 2338 in Brazil’s Chamber, Bill No. 16.821-19 in Chile’s Senate and bill 043/2025 in Colombia’s Senate, all on the table at once. None of the three countries that signed published a position of its own. Three days earlier, the Financial Stability Board had asked the same G20 for the opposite: to treat frontier models as a systemic risk.
There is a detail worth keeping in mind when reading any summary of this, including this one: no official text has been published on any U.S. government domain. What is known about the framework comes from what was announced at the summit and from press coverage. The document will be presented for formal adoption at the leaders’ summit in December, in Florida.
Also today
- The U.S. Department of Justice sides with OpenAI: training on protected works is fair use — A twenty-page amicus curiae brief in the New York Times case, contradicting the report of the U.S. Copyright Office itself. The copyright question becomes industrial policy, and no Latin American publisher, archive or collective management society is at that table.
- OpenAI adopts “opaque recurrence” in Astra, and safety experts warn that the trail of reasoning is lost — If the model reasons in internal loops instead of writing out its steps in readable text, the explainability required by the bills in Brazil, Chile and Colombia becomes hard to enforce even before they are passed.
- Six in ten Colombians who use AI do tasks they could not do a year ago — The Colombian cut of Microsoft’s work index: 63% do new tasks, 70% fear falling behind and only 28% believe their organization’s leadership is aligned. A workforce that is adopting on its own.
- A Texas police department used AI to write the report on a Flock camera search for a woman who had an abortion — Eighty thousand cameras mobilized to track something that is not a crime under that state’s law, and the record of the search was written by a model, with a disclaimer line at the bottom.
- A floating island appeared and disappeared on a Canadian lake, and the first official hypothesis was that the video was AI — The “liar’s dividend” in reverse: the state-owned power utility dismissed a real event because it suspected it was generated, and needed satellite images to believe it.
In the region
Latin America’s only institutional activity of its own at the summit was Mexican, and its agenda was industrial policy more than governance: Mexico negotiated access to the AI value chain and secured a committed visit from Jensen Huang, Nvidia’s chief executive. In the same room where the Carolina Principles were signed, the only country in the region with a visible presence went to talk about plants and supercomputers. Meanwhile, two developments push in the opposite direction without any regional authority following them closely: the U.S. Department of Justice’s backing of fair use, which comes just as Chile’s copyright bill is debating collective licensing, and the progressive closing of access to cybersecurity capability. Google today added its Fairwind Program, which provides its cybersecurity model only to “trusted” governments and critical operators under unpublished admission criteria: it is the third closed channel in three days, alongside OpenAI’s Daybreak Blue and Anthropic’s trusted access programs, and none of them declares a way in for an incident response team, a central bank or an electoral court in the region.
For the record: the fourth edition of Hi Ventures’ State of AI in Latin America (official page) surveyed more than 420 startups, corporations and funds. Some 71% of the region’s companies report having adopted AI in some function, and 53% of startups have agents in production, versus 30% of corporations. The readiness index by country stands as follows: Chile 72.5, Argentina 71, Brazil 70.2 (the biggest year-over-year improvement), Mexico 67.5 and Colombia 53.1. It is worth reading with a caveat the report itself states: it is produced by a venture capital fund that invests in the startups whose adoption it measures.
Launches
- Gemini 3.8 Flash — Google’s workhorse model reaches 73.7% on DeepSWE v1.1, practically tied with Claude Opus 5 (74) and above GPT-5.6 Sol, at $0.75 and $3.75 per million tokens until the end of the year (the price doubles in January 2027). It is publicly available in AI Studio, Gemini Enterprise, the app and Sheets. It is the best capability-to-price ratio available today for a public pilot or a small business in the region: about a fifth of the cost of Opus 5. Google does not publish which languages or countries it supports.
- Agentic video analysis in Gemini — Instead of slicing video at a fixed frame rate, the model decides which segments are worth watching: up to 88% fewer tokens, 66% lower cost and 7% more accuracy, especially on long video. It is enabled in the Gemini API with the “agentic” processing mode, at no additional charge for the feature. Probably the least discussed and most useful news of the day for anyone processing legislative sessions, court hearings or radio archives on a tight budget.
- Muse Spark 1.3 — Meta Superintelligence Labs’ new model uses about 20% fewer tool calls and 25% fewer tokens than its predecessor, and it brings a different behavior: it asks questions when faced with ambiguity, asks for help when it gets stuck and confirms before acting. Available in Muse Code and the Meta Model API. The weights were not opened at launch; their release remains a roadmap promise, which is precisely the variable that any regional technical sovereignty strategy depends on.
Threads we’re following
This adds to a story we have been following since Monday. When the Carolina Principles appeared, they were a U.S. proposal with verification problems and no public text. In four days they went from proposal to a unanimous agreement of the twenty, with a date set for formal adoption in December. The other thread that crosses paths with it today is access: every frontier lab that announces a cybersecurity capability pairs it with a closed channel, and the list of who gets in remains unpublished. These are two different moves pointing to the same place: the ceiling of what a country in the region can regulate, and the floor of what it can defend, are being set in conversations in which it barely participates.
If a country’s regulatory framework is negotiated at a trade summit and signed without its legislature knowing, what is left to decide for the legislator who was drafting that very law?
Correction (September 30, 2026). The original version said that Brazil, Mexico and Argentina signed in the same month in which “their legislatures” are debating that institutional design; the correct wording is “three legislatures in the region,” because the bills cited are from Brazil, Chile and Colombia, and Chile and Colombia are not G20 members, according to the Council on Foreign Relations.
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.