Demis Hassabis, CEO of Google DeepMind, published an essay and gave an interview to Axios proposing something no one at his level had put on the table in such detail until now: an independent global standards body for frontier AI. The idea, as TechCrunch reported, is an entity funded by the industry itself and backed by the US government, modeled on FINRA, the self-regulatory organization that oversees the US financial industry. Labs would submit their models up to 30 days before launch for testing of cyber, biological and “deception” risks.
It matters because it is the most concrete proposal for formal governance of advanced AI to come from one of the three most influential lab CEOs on the planet. In a sector where self-regulation tends to stall at vague principles, Hassabis sketches a mechanism with rules, deadlines and a mandatory prior review. The problem, seen from Latin America, is one of political architecture: a body “funded by the industry and backed by the United States” reproduces the same pattern the region has already experienced firsthand. When the US Department of Commerce used its export controls (the so-called BIS ban) to decide which frontier models, Fable 5 and Mythos 5, the rest of the world could or could not see, those rules were designed entirely between Washington and US labs. No seats reserved for anyone else. Hassabis’s proposal, however well-intentioned, risks institutionalizing that asymmetry: whoever certifies which model is “safe” decides, de facto, which AI each country gets access to.
Also today
- Musk promises to delete the data Grok Build uploaded without clear consent — xAI’s coding tool was sending entire repositories to the cloud; the company quietly patched the leak, with no deletion timeline or official statement.
- IBM plunges 25% in a single day, its worst drop ever — the AI chip shortage is pushing customers to divert budget toward servers and memory, delaying the closing of software deals.
- DeepSeek negotiates a $1.5 billion round at a $71 billion valuation — a step ahead of a possible IPO in China, weeks after closing a record $7 billion round.
- China rolls out the first specific regulation on chatbots with “emotional bonds” — it bans virtual intimate relationships with minors and requires parental consent for services that simulate emotional ties.
In the region
The contrast with the Latin American regulatory scene is direct. In Chile, the Senate votes today on final passage of the national reconstruction “mega-reform”; morning reports indicate that the article exempting AI platforms from paying to use protected works to train their models was removed for good, after being rejected twice in the Chamber. In Argentina, a decree reorganized control over the AAIP (data protection, author of the national guidelines for responsible AI) and ENACOM (telecommunications), moving it to the Chief of Cabinet’s Office in the middle of negotiations over the Telecom-Telefónica merger. And in Mexico, Congress confirmed that debate on its AI regulatory framework begins on July 19, now leaning toward a more domestic approach rather than one harmonized with the North American bloc, given the uncertainty surrounding the USMCA. In the background, a figure that explains the urgency: the region could add 4.8 million high-skilled formal jobs if it speeds up AI adoption, even though only 1 in 10 companies uses it today.
Launches
- Claude for Teachers — free access for one year to Claude Premium (including programming and assisted-work tools) for verified K-12 teachers in the United States. It is not yet available in Latin America, but it is the kind of vertical education strategy that labs tend to replicate later in other regions.
Threads we’re following
For days we have been following the debate over how far coding agents can go in real mathematics: first a proof that OpenAI attributed to its models on an open conjecture, then the discussion over whether these systems actually contribute or just speed up human work. Today’s chapter adds an optimistic note and a skeptical one at the same time. On one hand, mathematician Terence Tao built an interactive explorer of the Gilbreath conjecture using a coding agent, a concrete and sober example of the tool as an assistant. On the other, the mathematical community is still debating whether the proof OpenAI presented on the Cycle Double Cover Conjecture failed to cite a 1983 paper. Enthusiasm and careful verification, this time, are moving in parallel.
If the body that certifies which AI is “safe” is born funded by the industry and anchored in Washington, are we solving global AI governance or formalizing who is left out of the conversation?
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.