Two days before Colombia’s presidential runoff, the campaign’s sixth deepfake appeared: a fake video claiming that candidate Iván Cepeda has advanced colon cancer with metastasis and takes morphine daily. Cepeda responded by publishing a medical certificate that rules out any active recurrence and announced legal action (El Tiempo). It is the sixth fabrication generated with artificial intelligence in 21 days of campaigning, and this time the target is not the candidate’s ideas, but his body.
We had been following this case as the most documented electoral AI experiment in the region, and today it comes to a close with an uncomfortable paradox. From 6 p.m. today until June 22 a public order law is in effect: the Technology Ministry (MinTIC) and the ColCERT response team are deploying real-time deepfake monitoring, hour by hour; the United States issued a formal alert about the process, and more than 50 civil society organizations asked both candidates to commit publicly to recognizing the result. But there is a timing problem that defines everything: Law 2502, the rule that provides the mechanism to sanction deepfakes, does not take effect until July 2026. The state can detect, it can warn and it can watch minute by minute —but it cannot punish.
That is the lesson Colombia leaves as a regional precedent, and it applies to everyone coming behind it. What has been shown is not just a record of six deepfakes in one campaign: it is that detection capacity without sanctioning capacity does not protect democracy but rather documents precisely how it erodes. Brazil, which votes in October 2026, Chile and the rest of the region reach their own electoral cycles with this experiment as the only precedent at hand. What happens on June 21 will be watched closely.
Also today
- A Chinese lab launches GLM-5.2 and uses the Fable 5 ban as a banner — Zhipu AI (Z.ai) released GLM-5.2 with open weights and an MIT license, and for the first time a frontier lab explicitly names a US regulatory block —the ban on Anthropic’s Fable 5 model— as the geopolitical motivation for a launch. Accessible worldwide, Latin America included.
- Anthropic ban, day seven: Trump says it is “going well” — the first direct presidential statement on the block on Anthropic’s most powerful models, which remains in place. The proposal for an exemption for the United Kingdom fell through, and the deadline to request refunds expires on June 20.
- The Évian G7 closed without a common AI framework — no binding commitment on the governance of frontier models. According to the Washington Post, the ban “strained” relations with US allies during the summit. Latin America had no seat at any table.
- The last week of June promises the biggest model competition in years — GPT-5.6 still has no date (the prediction market Polymarket gives it 83% odds of coming out between June 22 and 28) and Gemini 3.5 Pro remains in limited preview, while Fable 5 continues to be blocked.
In the region
Next week, on June 25 and 26, Santo Domingo hosts the Third Ministerial Summit on the Ethics of Artificial Intelligence in Latin America and the Caribbean, organized by UNESCO, the development bank CAF and the Dominican digital government office. It brings together the region’s science and technology ministers just four days after the Colombian vote, and in a charged context: a ban on Anthropic’s most capable models decided without regional consultation, six electoral deepfakes in Colombia and the region’s absence from the G7. Today Latin American developers can choose among three available frontier models —xAI’s Grok 4, Zhipu’s GLM-5.2 and OpenAI’s GPT-5.5—, but none is based in the region, has governance obligations toward it or went through a regional consultation. Whether the summit produces concrete commitments or just another declaration of principles will say a lot about the region’s real capacity to coordinate.
Launches
- GLM-5.2 (Zhipu AI / Z.ai) — a 744 billion parameter model with a “mixture of experts” architecture (40 billion active per query), a context window of one million tokens and an MIT license, downloadable on HuggingFace and from Z.ai. The company claims it outperforms GPT-5.5 in long-context programming at about one-sixth of the cost, but the benchmarks are its own: it is worth waiting for independent evaluations, ideally in Spanish, before taking them as fact.
Threads we’re following
The ban the United States imposed a little over a week ago on Fable 5 and Mythos 5 —Anthropic’s two most powerful models, “switched off” across the planet through an export control— keeps sending out ripples. Two chapters were added today: a Chinese lab that turns the block into a marketing argument for its own open model, and a new Anthropic privacy policy, effective July 7, that incorporates government identity verification and biometric data. That last piece suggests a possible path to restoring access based on verifying the user’s citizenship; if implemented that way, the ban would stop being geographic and become personal, with different —and still uncertain— consequences for those in Latin America.
Can a region that documents so precisely what does not work —deepfakes without sanctions, models blocked without consultation, summits without a seat— turn that record into capacity for action? Next week, between Colombia’s ballot boxes and the table in Santo Domingo, the answer begins to take shape.
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.