In fourteen days, the US government went from sanctioning an artificial intelligence model to controlling one before it exists for the public. The White House —through its cybersecurity and science policy offices— asked OpenAI to restrict access to its next model, GPT-5.6, to a group of enterprise partners previously approved by the government itself, ahead of the general launch. Sam Altman told employees at an internal meeting on June 25; there is no official statement from either the government or the company, and the information came from internal sources (Axios).
It is the first known case of a preventive government restriction on a frontier model —the most advanced systems on the market. The difference from what happened at the beginning of the month is not minor. The ban on Fable 5, Anthropic’s model, was a reactive sanction: it came after the launch and was justified by a specific security flaw. The GPT-5.6 case is something else: a prior approval mechanism, with no documented incident and no public announcement. If the pattern takes hold, Washington would not be responding to specific threats, but managing in advance which models reach which markets and under what conditions. And for Latin America —which has no seat in either of the two conversations— the practical result is identical in both cases: two of the most capable models available today are out of reach without consultation or prior notice to the region.
On the same day this restriction became known, the region’s technology ministers were wrapping up a summit in Santo Domingo that was trying, precisely, to build a common position. The coincidence defines the underlying problem: the region is agreeing on principles about AI ethics while the decisions that determine which tools it can access are made at another table, without it.
Also today
- Google launches DiffusionGemma, an open model for generating text — It uses a diffusion architecture (the same idea behind image generators) instead of the usual approach, and its weights are downloadable. It makes it possible to generate text on the device itself, without depending on the cloud, in Spanish as well.
- Nearly 400 local newspapers sue Microsoft and OpenAI over the use of their content — The largest coordinated action by US local media against AI labs, for having used their archives to train models. It could serve as a template for Latin American media.
- Qualcomm buys Modular for about $3.9 billion — A platform that makes it possible to run AI models on different chips without rewriting code, a direct challenge to NVIDIA’s dominance. Relevant for countries in the region with limited access to the most expensive GPUs.
- Reuters: OpenAI is reportedly considering postponing its IPO until 2027 — Because of the regulatory complexity following the restriction on GPT-5.6. It is a single source, without confirmation from the company; if it happens, Anthropic’s stock market debut, planned for the end of the year, would be the first by a frontier lab.
- AI surveillance technology is deployed in eight World Cup host cities — The company Peregrine, which merges police and sensor data, raised $250 million; Mexico adopted its system without a withdrawal clause. The EFF warns that surveillance infrastructure tends to outlive the event that justifies it. Brazil hosts the 2030 World Cup.
In the region
UNESCO’s Third Ministerial Summit on the Ethics of AI in Latin America and the Caribbean closed today in Santo Domingo with more concrete results than its predecessors. Where Santiago (2023) and Montevideo (2024) left declarations of principles, this summit adopted the Santo Domingo Declaration and a Regional Roadmap 2026-2027 with five priority areas and a twelve-month horizon, prepared by CAF and UNESCO with the collaboration of more than twenty countries (FORO IA LAC · UNESCO). The Dominican Republic took over the pro tempore presidency and went a step further: it announced an AI Code of Ethics by presidential decree, the first national instrument adopted directly from the regional multilateral framework.
None of this is enough, on its own, to respond to a unilateral control like the one on GPT-5.6 or the ban on Fable 5; no regional instrument can reverse those decisions today. But it is the first operational AI governance roadmap the region has given itself, and it arrives in the exact week when the distance between agreeing on principles and being able to carry them out became impossible to ignore.
Launches
- DiffusionGemma (Google DeepMind) — An open model with downloadable weights for text generation through diffusion, designed to run on local hardware without the cloud. It is the first of its family at Google available for local deployment; it requires a compatible GPU.
- Grok 4.3 on Amazon Bedrock (xAI) — The first xAI version distributed on AWS infrastructure, including the São Paulo region, with a lower rate of fabrication errors and a context window of one million tokens. Accessible to companies with an AWS contract, a route to a frontier model while the Fable 5 ban remains in place.
Threads we’re following
The US ban on Anthropic’s Fable 5 and Mythos 5 —which we have been following since it was announced on June 12— reaches its fourteenth day today with no public resolution. Tom Brown, a veteran researcher in the sector, remains the company’s point of contact with US agencies, and prediction markets keep the probability of an outcome before July 1 at around 57-60% (status of the ban). This week the technical context hardened: a joint advisory from the cybersecurity agencies of the United States, the United Kingdom, Canada, Australia and New Zealand —the group known as Five Eyes, published on June 22— states for the first time in an official multilateral document that AI models capable of launching major cyberattacks could overcome the defenses of governments and companies in a matter of months, not years (ACSC). The advisory does not mention the Anthropic case, but it describes exactly the scenario invoked to justify the ban, now backed by the five most sophisticated cybersecurity agencies in the world.
Today the region agreed, for the first time, on a common roadmap on AI ethics, on the same day it learned that its access to the most powerful tools depends on decisions made without it. What concrete instrument —mandatory prior consultation, guaranteed access to audited versions, or a formal seat where these rules are negotiated— would be needed so that Latin America is not the last to find out about the next control?
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.