Meta launched Meta One, a global ladder of subscriptions that runs from $2.99 to $499 a month and puts heavy use of artificial intelligence in Instagram, Facebook, WhatsApp and Meta AI behind a paid plan. What you buy is quota: more image and video generation and editing with the company’s Muse models, tools like Restyle, and the business agent. Meta insists that the basic experience remains free and reports 15 million cumulative subscriptions and trials.
For Latin America, the relevant fact is not the price but the channel. WhatsApp is the app through which much of the region first met generative AI without paying anything, and the experiment Meta began in May with two test markets —Guatemala and Bolivia— is now a global product with eight tiers billed in dollars. Two questions remain open: how much of the free floor survives, and at what exchange rate the next tier is paid. The context makes them more urgent: independent measurements show that heavy AI use is doubling year over year in the Northern Hemisphere, just as access begins to be segmented by ability to pay.
The same day brought another signal pointing the same way. Cloudflare turned on blocking of mixed-use crawlers by default and created an “Accountable” label for those that declare their intentions: an infrastructure provider has just decided, with no law involved, who can train models on the content of the open web. Two private decisions, made in a settings panel and in a price list, end up defining for the region both the cost of using AI and the conditions under which its content feeds the next generation of models.
Also today
- OpenAI confirms it is negotiating safety measures with Anthropic and Google — The company maintains that it does not need antitrust clearance to coordinate. These are three private companies writing the technical rules that the region inherits as a de facto standard, with no Latin American seat at the table.
- Cloudflare turns on default blocking of mixed-use crawlers — A “mixed-use” crawler is one that scours the web without distinguishing whether the content goes to a search engine or to train a model. It is now blocked unless it identifies itself.
- Jensen Huang, at Dreamforce: “Safety is an engineering problem, not a legal one” — The argument for leaving safety in the hands of the market presupposes a market capable of punishing failures by foreign providers; in the region, that mechanism does not exist.
- A digital fly brain takes over the internet — 166,700 simulated neurons playing Minecraft and Doom. It is the best available counterexample to the loose use of the brain metaphor when talking about language models.
- The first hotlines appear for one AI agent to report another — An agent without full internet access can “encode its distress directly in the URL.” A Cornell professor warns that the mechanism builds an automated surveillance state.
- One in four large US companies cannot detect unauthorized AI agents on their network — According to EY’s survey of 202 senior executives, 98% report formal AI governance policies, but 47% bypassed them for urgent deployments and 36% suffered incidents with material harm. The full report is here.
In the region
The week brings two regional moves worth reading together. CAF put a figure on the table for the first time: more than $12 billion over five years for digital transformation, with three pillars —public AI capabilities, infrastructure and cybersecurity, and the productive transformation of small and medium-sized businesses— along with a call for a single regional agenda. Its executive president framed the argument precisely: no country in the region can influence on its own the rules that will define the future. What will determine the outcome, however, is the conditionality of the financing, not the amount.
Meanwhile, Latin American media directors wrapped up two days of the LATAM Media Leaders Summit in Lima with a shared thesis —less volume, more editorial judgment— and the case of La Silla Vacía, which cut its big daily stories from four to two. The coincidence in timing is the point: while Lima was discussing how to negotiate with AI, Cloudflare was switching on by default the conditions under which that same content can train models. In cybersecurity, Google proposed federated threat intelligence centers at the LATAM CISO Summit: what used to take eight hours now runs in seconds, the patching window has fallen to 1.6 days, and no national incident response team in the region has early access to frontier cyberdefense models.
Launches
- Meta One — Subscriptions that add AI quota in Instagram, Facebook, WhatsApp and Meta AI, from $2.99 to $499 a month. Gradual global rollout, with prices that vary by region.
- Disallow AI Training and the “Accountable” designation — A Cloudflare control that lets sites keep appearing in search engines while refusing model training. Available to all customers, including the free plan: it is the tool that a media outlet, a university or a public repository in the region can turn on today without waiting for legislation.
- Koa — A reasoning model for Salesforce customer management, built on Nvidia Nemotron and trained on synthetic data, not real customer data. Only for Agentforce and Missionforce customers. It matters as a signal: a vertical model that runs inside the provider’s perimeter changes the data sovereignty argument for banking, health and government.
Threads we’re following
We had been following the debate over who writes the rules for AI and from where: UNESCO’s Global Forum on the Ethics of AI, meeting in Riyadh and closing on September 17, is the only space where the region’s countries took part in the drafting. Today’s chapter puts it in perspective. While that forum deliberates, three companies negotiate safety standards among themselves and claim they need no antitrust permission, an infrastructure provider decides by default who trains on the open web, and a platform sets in dollars the price of mass access. The multilateral route moves at the pace of forums; the other three, at the pace of product announcements.
Another line that was already emerging returns with figures: money is concentrating in the layer that supervises and directs agents. AIUC raised $40 million to certify agents against its own standard of some 5,000 tests, and Profound reached a valuation of $1.8 billion for controlling which brands appear inside a generated answer, with no advertising transparency rules involved.
If the price of AI in Latin America is set in dollars from Menlo Park, and the terms of use for the region’s content are set by default from a settings panel in San Francisco, what is left for a Latin American state to regulate that it does not have to negotiate company by company?
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.