Today UNESCO’s 4th Global Forum on the Ethics of Artificial Intelligence opened in Riyadh, with delegations from 194 countries and sessions through Thursday, and its mandate is concrete: to turn the 2021 Recommendation on the Ethics of AI, the only global framework approved by consensus and the only one Latin America helped write, into enforceable mechanisms. It arrives in a week in which two other architectures for governing the same technology moved in parallel, and neither of them includes the region. Over the weekend, The Information revealed that Anthropic, OpenAI and Google DeepMind have been holding regular meetings since July to set up a private standards body: independent testing, pre-launch safety reviews and standardized risk assessments. In other words, the slowdown plan that Dario Amodei presented on Saturday as a proposal had been under closed-door negotiation for two months, with no regulators inside.
The third architecture came from New Delhi: at the BRICS summit, Xi Jinping offered for China to lead an open-source artificial intelligence community for the Global South, with cooperation on language models and training courses, although the summit’s joint declaration ended up not taking up the proposal. The comparison is uncomfortable to look at head-on: the instrument where the region has a voice recommends and does not bind; the one that will probably determine how the models the region already uses are deployed is being drafted in private meetings among three U.S. companies; and the one that offers exactly what the region says it lacks (open weights and training) comes with a geopolitical flag and, from Latin America, includes only Brazil.
In the background, money kept moving in the opposite direction from the brake. Anthropic is guiding to a second consecutive quarter of adjusted operating profit and chose Nasdaq for an eventual IPO of up to $100 billion, after a quarter of more than 11.5 billion in revenue. SoftBank secured an expanded loan of $11.87 billion from some twenty banks to finance its investment in OpenAI, secured by its own stake in OpenAI, on the same day that stake was falling in Tokyo. And The Information identified Anthropic as RUM Group’s $13.7 billion customer, the Trump-linked compute company: six years of GPU services in Georgia with a provider that acknowledges it does not have the money to fulfill the contract.
Also today
- Trump, the Speaker of the House and the White House AI adviser reject slowing the frontier en bloc — “whoever wins AI, wins.” The chair of the science and technology advisory council called the coordination the industry is proposing a cartel.
- Asian markets punish the call to slow down and SoftBank falls more than ten percent — the first time an AI safety discussion, and not demand or margins, has measurably moved the price of the Asian memory supply chain.
- A team of seven people trains frontier-level offensive cyber agents on open models — with 164,269 well-built data trajectories and three versions of Qwen, they reached 10th place on the official CyberGym leaderboard, with a 63.24% verified success rate. It is a preprint with no peer review and no declared affiliation, so it should be read with a magnifying glass; if it holds up, none of this week’s three governance architectures reaches it, because they all assume closed models.
- China’s Minister of State Security calls AI “a new arena of strategic rivalry between great powers” — the same weekend Washington says the alarms are exaggerated, the Chinese security apparatus published its own list of risks.
- Claude Code adjusts its weekly limits — starting today, the Pro, Max, Team and per-seat Enterprise plans operate with caps 17% lower than yesterday: the temporary 50% boost in effect since May expired and is replaced by a permanent 25% increase measured against the baseline before that promotion.
In the region
The Riyadh agenda has four items where a country that hosts data centers but does not train models has something concrete to ask for: a closed-door ministerial meeting, a scientific session, the presentation of tools to measure the environmental impact and energy consumption of large models, and coordination among national regulatory agencies. This time Latin America arrives with a position of its own: the Santo Domingo Declaration, approved at the 3rd Ministerial Summit in June, sets five priorities for 2026-2027 (governance and regulation, talent and the future of work, protection of vulnerable groups, environment and sustainability, infrastructure). It is worth noting who the host is: the forum is co-organized and funded by the Saudi Data and AI Authority and the ICAIRE center, that is, a state that is also one of the world’s largest buyers of compute, and that shapes which topics are comfortable to discuss. Outside Riyadh there were no dated announcements from ministries or regulators in the region (it was Sunday and the early hours of Monday), and the relevant regional development was indirect: Xi’s offer at the BRICS summit, where Brazil participates from the inside while the rest of the continent negotiates country by country with both blocs at once. One more domestic front does land directly: the cut to Claude Code’s limits communicated as an increase, because it is measured against a different baseline, is exactly the kind of fine print over which Sernac, Profeco and Senacon already have jurisdiction and on which they have not yet ruled.
Threads we’re following
This adds to the story we have been following since last week, when the call to slow the development of the most advanced models went from hallway conversation to a proposal with a text and steps. Today’s chapter changes the order of events: the step that seemed the hardest (setting up a body that audits the labs before they launch) was not an idea but a two-month negotiation, and the U.S. government responded en bloc that it has no intention of slowing anything down. Still pending is what that private architecture does not cover, which today’s preprint puts back on the table: open-weight models, which nobody can review before launch because there is no launch to review.
If the only AI governance framework Latin America helped write is also the only one that binds no one to anything, what would have to come out of Riyadh this week for the Santo Domingo roadmap to stop being a declaration of intent and become a condition for entry into the regional market?
Correction (September 30, 2026). The original version said Dario Amodei presented his slowdown plan on Friday; the correct date is Saturday, September 12, according to CoinDesk.
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.