What the AI Index 2026 Confesses Through Its Silence on the Dominican Republic
The ninth edition of Stanford's AI Index Report does not mention the Dominican Republic even once. That arithmetic deserves to be read as a diagnosis.

The ninth edition of the AI Index Report from Stanford's Institute for Human-Centered Artificial Intelligence has just been published. Four hundred and twenty-five pages, nine chapters, dozens of datasets. And across that entire body of evidence, the words «Dominican Republic» appear exactly zero times. The Caribbean appears eleven times, almost always as a regional aggregate. Jamaica, once. Costa Rica, fourteen times. That arithmetic deserves to be read as a diagnosis.
It is worth starting by acknowledging what the report does well. Stanford maintains a rare kind of methodological honesty: it publishes raw data and open repositories, offers interactive tools such as the Global AI Vibrancy Tool, and acknowledges the limits of its own measurement. The sentence that opens the introduction, —«what we cannot yet measure matters as much as what we can»—, is methodologically honest in a field saturated with reports that are marketing pieces dressed up as research. The 2026 edition adds two new chapters on AI in science and medicine, widening the lens beyond the consumer product cycle. And for the first time it establishes AI sovereignty as a structuring category, broken down into five layers: infrastructure, data, models, applications, talent.
That conceptual backbone is the firm ground from which the limitations are best read. Generative AI adoption reached 53 percent global penetration in three years, with Singapore leading at 61 percent and the United States in twenty-fourth place at just 28 percent. Documented incidents grew from two hundred and thirty-three in 2024 to three hundred and sixty-two in 2025. U.S. private investment reached 286.9 billion dollars, against China's 12.4. The data are there. The question is what they make visible and what they leave in the shadows.
Here the first structural limit appears. More than ninety percent of the notable models recorded in 2025 came from industry. And that industry publishes less and less about how it builds what it builds. We are measuring better than ever a phenomenon that is becoming more opaque than ever. The very definition of a notable model privileges benchmark performance, academic citations and media coverage. An entire layer of cognitive activity falls outside the frame: small models trained in local languages, sector-specific models applied to concrete problems, academic models that solve real problems in their own contexts. The report itself acknowledges, in a footnote, that the ecosystem of African language models such as AfriBERTa remains unrepresented.
Langdon Winner named it decades ago: the politics of artifacts. Measurement is itself a political artifact. When we decide what counts as a notable model, we decide what kind of cognitive activity deserves visibility. And the available methodology keeps counting creators while leaving out the substrate. The people in Kenya who label images to train computer vision, those in the Philippines who moderate toxic content so American platforms can function, those in Venezuela and Mexico who annotate conversations to improve chatbot responses: all of them are doing massive, essential, poorly paid cognitive work. Without that substrate, frontier models would not exist. In the AI Index, they do not appear. The economics of artificial intelligence measures Sam Altman without measuring the people who make his product possible. Victorian Britain measured coal without measuring the miners.
Here a concept I have worked on for years applies: The invisibles. Professionals and cognitive workers whose contribution is counted in other countries' national accounts. The methodology inherits that blindness because the available databases —Zeki, LinkedIn, GitHub— capture a very specific professional layer of the value chain. The environmental footprint appears in the report with alarming figures —Grok 4 emitted 72 tons of carbon dioxide equivalent in its training alone; GPT-4o's annual inferences consume as much water as 1,200,000 people need to drink for an entire year; data center electrical capacity reached 29.6 gigawatts, equivalent to the peak demand of New York State—, yet the language remains descriptive. The footprint should operate as the first structural filter of the analysis.
The second limit is the one that challenges us most. Latin America and the Caribbean produced only two documentable models between 2018 and 2025. Two. Against sixteen hundred and eighteen from the United States, eight hundred and forty-nine from China, six hundred and sixty-six from Europe. Eight regional supercomputing clusters, against forty-four in Europe and forty-one in North America. Costa Rica and Jamaica with national AI strategies adopted. Mexico and South Africa with strategies under development. Chile leading the Latam-GPT initiative as a regional bet on a Spanish-language model adapted to the Latin American context. The Dominican Republic is absent from every one of those lists.
The paradox is clear. A country that celebrates having been the second in Latin America to auction 5G, with eighty-five percent internet penetration, promoted as the Caribbean's digital hub, fails to appear in a single productive metric of the report. Here another framework I have developed for years comes in: the Pokémon effect. We accumulate certifications, rankings, penetration percentages, submarine cables, private data centers. We catch every Pokémon around us. And we ignore the game those certifications presuppose. Connectivity is a necessary condition. We mistake it for a result and end up celebrating physical infrastructure while the cognitive kind goes unbuilt.
The national cognitive infrastructure is the missing layer. Public data that is cataloged and accessible. Models trained on local corpora in Caribbean Spanish that no Latam-GPT is capturing yet. Analytical capacities installed with permanence inside the State. Universities producing applied research with publishable rigor. Dominican companies with their own product, able to export analytical capacity. Without that layer, digital highways mainly benefit foreign platforms. The symbolic industrialization —four plans, two public consultations, dozens of events, signed memoranda, offices with grandiloquent names— hangs in the air while the metric that matters, the number of models produced by the country, remains zero.
The U.S. immigration bloc opens a mixed window. The flow of AI talent to the United States has fallen 89 percent since 2017, with 80 percent in the last year alone. India stopped being a net exporter and began to retain. We have brilliant Dominicans working remotely for Boston, Berlin and Singapore, generating cognitive value whose figures are counted in other countries' national accounts. If part of that flow is redirected toward local initiatives, the country has an opportunity. Redirection requires public policy, investment, and domestic demand that values AI work at the level the external market pays.
Five concrete levers would have to be activated in the next twelve months so that the AI Index 2027 cannot ignore us again. First, formal incorporation into the Latam-GPT consortium with a contribution of Caribbean Spanish corpus and proportional funding —we managed to take part in its launch thanks to the support of our friends at CDN—; embracing this model would move the country out of the regional zero category and onto the list of contributors. Second, a national AI strategy published and operationalized with a multi-year budget and auditable quarterly metrics, able to survive three electoral cycles. Third, an infrastructure of open public data, cataloged, with clear licenses, that allows the training of sector-specific models in health, education and transport. Fourth, a sovereign computing program under a public-private agreement with at least one of the providers with regional presence, dedicated to university research and public-good projects. Fifth, and most important, a public procurement policy that favors locally developed AI solutions, to create the domestic market that remains absent today.
Those five things already operate in other countries of the region. The knowledge exists. What is missing is execution. And execution is built through collaboration among government, the private sector, academia and civil society, with the discipline to keep the plan when cabinets change. That is where symbolic industrialization separates from real industrialization. The silence that surrounds us across four hundred and twenty-five pages works as a diagnosis, one that can change. Diagnoses only change when collaboration and execution stop being words in speeches and become infrastructure. We need AI for everyone.