The country that asks the machine for the winning number

In Costa Rica, artificial intelligence has had more opportunities to guess the accumulated prize of the National Lottery than to enter a legislative discussion. Between January 1 and September 13, 2026, the national press published 61 articles attributing to an AI predictions of lottery results or sports scores. Only 15 linked it to the Legislative Assembly, deputies, or a draft law.

The contrast seems like a quaint oddity. In reality, it reveals how the Costa Rican press is teaching people to imagine artificial intelligence. Among the articles that do land in the country, the most common framework assigns AI the role of oracle, ahead of regulation or legislative discussion: a faceless machine that recommends numbers, anticipates scores, and lends a technological appearance to chance.

A second analysis, based on human review of a representative sample of the corpus, will be published later to compare the results of an agentic method against a manual one and thus validate methodological consistency.

The conclusion emerges from an analysis conducted with Haiku on a sample of 9,025 news articles from Costa Rican print and digital media. Within that corpus, 1,472 pieces, or 16.3%, mentioned artificial intelligence, its major brands, or associated techniques. The volume is considerable, but the content is substantially less national: only 28.6% of AI articles mention Costa Rica, a public institution, a province, or an identifiable local actor in the headline or summary. In headlines, the proportion drops to 16.3%.

Haiku is an automated news monitoring platform and media intelligence analysis tool developed by the author of this article. Some media outlets block automated tracking (web scraping) of their sites, so it is not possible to guarantee the collection of 100% of the news published in the analyzed period.

This portrait does not tell us what citizens think. Haiku measures editorial supply, not consumption or public opinion. It does allow us to observe what stories the media put before their readers. The image is uncomfortable: Costa Rican press talks a lot about AI, but much of the conversation arrives packaged from outside, with foreign protagonists, corporate vocabulary, and conclusions written before a journalist asks the first question.

An imported future#

That 28.6% has its counterside: seven out of every ten articles about artificial intelligence lack Costa Rican grounding. Many describe the launch of a model, a feature added to a phone, an investment by a multinational, or a warning issued from another capital. The country appears as a spectator of a future produced in California, Seattle, London, or Beijing. The proportion remains almost identical in both halves of the year, 28.3% between January and July and 29.2% from July onward, so it is not an effect of monitoring changes but a stable feature of coverage.

The cited entities reinforce that impression. 21.5% of all AI articles mention some major foreign company in the sector, including OpenAI, Anthropic, Google, Microsoft, Nvidia, Meta, Apple, Amazon, DeepSeek, or xAI. In the second half of the period, 551 mentions of organizations were extracted from within AI news; only 32, or 5.8%, corresponded to Costa Rican organizations. This last classification was performed by a model that used different AI APIs such as Claude, ChatGPT, Gemini, and Groq (yes, with a q at the end) and should be read as a comparative signal, not as an infallible census.

Vocabulary also shifts. Compared to the rest of national coverage, AI news uses far fewer words like government, president, country, project, public, or Costa Rica. Instead, they fill with company names, products, and models: the word "artificial" appears 6,732 times more than in the rest of national coverage; "OpenAI," 475 times more, and "Anthropic," 359. The agenda is organized by the launch schedules of those who manufacture the technology.

Covering a global industry requires looking outward. The problem emerges when local translation never arrives. A news story can explain what a company's new model does and still leave intact the questions relevant to Costa Rica: who will be able to use it, what jobs does it alter, what national data does it process, what infrastructure does it demand, which provider captures the value, and who is accountable when it fails. Without those questions, the country participates in the conversation as a market, not as an actor.

Technology receives the benefit of the doubt#

AI enters Costa Rican press accompanied by a favorable presumption. Among the 1,333 articles for which there was tone analysis, the model classified 40.1% as positive and 18.8% as negative. In the 6,612 articles with no mention of AI that also had tone analysis, almost the reverse occurred: 17.6% positive and 40.3% negative.

That automatic classification should not be converted into an emotional thermometer for each journalist. The tone was estimated by AI and serves mainly to compare groups within the corpus. Even with that caution, the inversion is too marked to ignore. The press tends to tell the story of Costa Rica's politics, security, and economy from the perspective of conflict. Technology enters by another door: launches, investments, agreements, events, and promises of efficiency.

The news frame predicts its tone. Articles on events and conferences were classified as positive in 70.8% of cases and recorded no negative pieces. Those on fraud and impersonation reached the opposite extreme: 77.5% negative. Education and health appeared as celebratory territories, with 64% and 66.2% of positive articles, and are also the frames most rooted in Costa Rica after the oracle, with 56% and 46%. When AI reaches the national level, it usually arrives as good institutional news.

The machine takes on the tone of the scenario where it is presented: it makes promises from a corporate stage, threatens in a fraud alert, and saves time in a classroom or hospital. Few news stories pause to observe the gray area where advantages, costs, errors, and people who bear the risk coexist. A press release rarely arrives with a skeptic in tow.

The State arrives when the decision is already made#

The governance and policy framework exists, but it is thin: 113 articles, a 7.7% of the AI corpus, and the only space where tone is split almost evenly, with 26.5% of positive pieces versus 28.4% negative. Within that space, the Costa Rican institutional presence is minimal. MICITT appears in eight of the 1,472 AI articles, a 0.54%. The Legislative Assembly, deputies, or bills figure in 15. CCSS appears in 16; MEP, in 14; INA, in six; and state universities, in 22. Budget or public financing is mentioned in 11. Technological sovereignty or dependence, in seven. TSE or electoral regulation appears in two articles; the Costa Rican Judiciary, in three.

The few exceptions help explain the pattern. The press covered MICITT's questioning of a proposal on AI in electoral processes, accompanied by the TSE's own warning that it is not "in any capacity to regulate"; the Judiciary's ban on replacing essential functions through automated systems; Costa Rica's incorporation into Pax Silica; and the country's alliance with Google Cloud. These are relevant facts and share a characteristic: the institution appears when signing, objecting, announcing, or prohibiting.

Prior deliberation rarely becomes news. We don't see the alternatives considered, the risks accepted, the technical disagreements, the evidence dismissed, or the conditions imposed on a provider. The citizen learns about the finished administrative act, not the process that determined its content.

That invisibility has consequences. In technologies capable of affecting rights, employment, public services, and budgets, the procedure is also policy. When public discussion begins after the signature or prohibition, debate is reduced to agreement or rejection. Accountability comes late and with little material to work with.

The absence of the National Artificial Intelligence Strategy (ENIA) from 1,472 articles sums up the problem. Zero mentions don't prove that the strategy—which the author of this article helped create—lacks actions or that the media refuse to cover it. They show something more concrete: it has not yet produced enough public, verifiable, and journalistically readable facts to shape the conversation. A policy confined to the language of instruments usually loses to a product that promises to edit photos in ten seconds.

The creole oracle#

4.1% of all AI coverage consists of lottery and soccer predictions. It may seem like an anecdotal margin. However, those 61 articles form the group most connected with Costa Rica: 60.7% mention national actors, more than double the corpus average. There are the Social Protection Board, Saprissa, Alajuelense, and the colones of the accumulated jackpot.

In La Teja, half of the AI coverage falls within that frame, and the words "luck" and "draw" appear 76 and 47 times more than in the rest of the corpus. The data deserves more than a condescending smile. Popular media found a way to domesticate a distant technology: they assigned it a familiar profession, that of divining the uncertain. AI stops being an abstract infrastructure and becomes an everyday character.

That access has a cost. The formula "AI predicts" grants authority without identifying method, data, success rate, or responsibility. The system functions as a nameless source and the technological halo disguises a selection that could be entirely arbitrary. The reader receives the result, but not the conditions under which it was produced. The model classified 90% of these articles as neutral in tone and none as negative: an AI without risk, without authorship, and without consequence.

Chance articles are not alone. Another 36 pieces use AI as viral curiosity and 23 explore emotional connections with chatbots: virtual partners, companionship, and loneliness. Together, these records add up to 120 articles, more than all the pieces linking AI with Costa Rican public institutions. The closest media experience with AI is thus constructed around entertainment, intimacy, and digital superstition.

Dismissing that gateway would be a mistake: well used, it could also educate. It would only take adding elementary questions to the narrative: what model produced the number, why did it answer that, how many earlier predictions failed, and what does it really mean for a machine to "recommend" a bet. Without that second layer, familiarity produces obedience rather than understanding—a kind of recreational submission.

An artificial intelligence for each audience#

Each media outlet seems to cover a different technology. La Nación published 291 articles and favored consumer products; only 14% had national relevance. El Financiero published 236 and focused on business and adoption. La Teja turned AI into an oracle. IT Now presented it mainly as an agenda of events and conferences. El Mundo CR, on the other hand, achieved 78% Costa Rican relevance, the highest among media with significant volume.

Intensity also varies. Within the analyzed corpus, 54.5% of the pieces captured from El Financiero dealt with AI. In CRHoy, a general reach medium, the proportion was 6.7%. Fifty-two media outlets published at least one AI article, but the two most active concentrated 35.8% of all coverage; the top five, 56.2%; and the top ten, 78.5%.

These figures don't allow us to assert who read each story or measure the actual size of their audiences. They do show a stratified editorial supply. The economic media delivers AI as a tool for productivity, investment, and strategy. The popular media approximates it through chance. Part of the mainstream press presents it as a catalog of functions and devices. The same acronym designates different social objects depending on the door through which the reader enters.

This fragmentation matters for public policy. A discussion about regulation, for example, will bring together people who arrive with incompatible frameworks: for some, AI is competitive advantage; for others, a labor threat; for others, a toy that answers anything. Talking about "public perception of AI" as if it were one thing is convenient and probably false. There are several artificial intelligences circulating in the national imagination, even though they all use the same name.

The press transcribes more than it interrogates#

A large part of AI coverage is not produced in Costa Rica. In the second half of the period, 32.1% of articles on the topic were signed by an international agency or a generic editorial label: 68 pieces from AFP, 26 from Summa Magazine, 19 from "IT Now Editorial" and 16 from the World Economic Forum. This is not a credits detail, because it explains a significant part of the national absence that runs through the corpus: a dispatch written for global audiences doesn't stop to ask what's happening with employment in Alajuela or with the budget of a Costa Rican institution.

The source's origin also anticipates the tone. In Summa Magazine, the model classified 78.6% of AI articles as positive and none as negative; in IT Now, 71.4%, with the same zero. El Financiero, which reports its own business stories, approaches balance with 31.7% of positive pieces versus 27.6% negative, and on Teletica the scale tips barely toward the negative. Where the speaker is the event organizer or the solution provider, conflict simply doesn't reach the text.

Even so, homogeneity is not explained by literal copies. Only 1.8% of distinct headlines appeared in two or more outlets. Newsrooms don't need to copy each other to end up in the same place; it's enough that they receive the same press releases, consult the same sources, and organize the agenda around the same events.

The result is coverage with little friction. Corporate claims travel from the presentation to the headline with few stops. In a field where almost every provider promises to transform some activity, journalism ends up measuring novelties rather than consequences.

Technical specialization would help, but scrutiny doesn't require every reporter to train a foundational model. Often a second question is enough: compared to what?, with which data?, who assumes the error?, what happened in the pilot?, how much does it cost to operate the solution once the announcement is over? The distance between promotion and journalism can fit in one well-placed question.

Work is where the promise cracks#

Coverage on employment offers a revealing exception. The model classified 104 articles within that framework. 37.6% had a negative tone and 36% showed Costa Rican anchoring. It's the territory where AI abandons abstraction and threatens to produce a recognizable loss.

But even there a key actor is missing. Only six articles in the AI corpus mention labor rights, unions, collective bargaining, or the ILO. Job displacement is narrated as a market force, similar to climate: it arrives, forces adaptation, and lacks an interlocutor. Workers appear frequently as an affected category, rarely as a source with capacity to negotiate how technology is introduced.

The same logic appears in other damage frameworks. Fraud and impersonation concentrates negativity because there is a victim, a stolen identity or money at risk. Disinformation and democracy total 37 articles; existential risk brings together 45, with 54.5% negative tone and barely 4% national anchoring. The more concrete the loss, the more chances the story has of touching Costa Rican ground.

This finding offers an editorial clue. Grand declarations about the future of work matter less than documenting what task changed, who decided to automate it, what productivity they obtained, what error appeared, and what appeal capacity the affected person retained. Ethics becomes news when it has a payroll, an invoice, a file, a victim, or when there is a conflict.

Absences also govern the conversation#

The absent topics sketch the implicit user of AI: connected, urban, and able to absorb change on their own. Gender, women, or girls appear in 34 articles, 2.3% of the AI corpus. Children and adolescents appear in 25; regions outside the Greater Metropolitan Area, in 24; human rights, in eight; intellectual property, in 11. Disability and accessibility appear in two. Digital divide or rural connectivity, also in two.

The infrastructure framework brings together 57 articles and is counted in celebratory key, with 50.9% of positive pieces versus 7.5% negative. The environmental dimension reaches 58 articles, but almost all look outward: data centers in the United States, global water consumption or European energy transition. The only series that connects an AI data center to Costa Rican territory is the one published about Limón, and it entered the agenda because of a problem with the free trade zone regime, not because of water or electricity use, or at least that's how the model detected it.

These absences don't prove that the effects don't exist, but rather that they don't yet count as central subjects of coverage. The fiction of frictionless adoption leaves out those who share a phone, depend on unstable connectivity, need an accessible interface, or live in a territory where new infrastructure will compete for resources.

It also undermines the quality of regulatory debate. If news presents AI as software floating in the cloud, the demands for energy, water, chips, data centers, and hiring disappear from the calculation. If employment is discussed without labor rights and education without the digital divide, policies will inherit the same blind spot as coverage.

Two stories that asked the second question#

The corpus contains two modest examples of different coverage. In August, La Nación investigated the presence of songs by Costa Rican artists in databases used to train AI models. The story identified 146 songs by Tapón and collected the musician's reaction. A global debate over intellectual property acquired a name, work, and possible local impact, within a framework—culture and authorship, 66 notes—that barely reaches 11% of Costa Rican anchoring.

In September, La Nación and La República followed the announcement of an AI data center in Limón. Two days later, coverage revealed that the company behind the project had lost its free trade zone status since May. The project stopped being a promise of investment and became a verifiable object, with administrative background and pending questions.

Neither of the two stories resolved on its own the debate over model training or digital infrastructure, but their value lies in the method: both located a global phenomenon, identified who could be affected, and sought evidence beyond the initial press release. There appears the journalism that AI needs.

The possible agenda is broad: public contracts, training databases, licenses, electricity and water consumption, pilot results, documented errors, appeal mechanisms, and benefit distribution. There's no need to chase science fiction. The documents, affected people, and budget decisions already offer enough conflict.

Public policy that can become news#

The response to zero mentions of ENIA shouldn't be a campaign to repeat its name. The challenge consists of producing observable decisions while they can still be discussed. A public registry of AI systems, impact assessments, rules for technology purchases, criteria for entry and exit from a sandbox, pilot results, and incident reports would give the press facts to examine, not talking points to reproduce.

That would also change the State's place in the story. Instead of appearing only during signing or prohibition, those of us leading public policy design like this from institutions could expose alternatives, publish evidence, and explain the risks we're willing to accept. Early transparency can generate conflict. Precisely because of that it has democratic value: it allows discussion to occur before technology becomes embedded in a public service or a contract difficult to reverse.

The media, for their part, need to treat AI as a cross-cutting source and not as a subsection of technology. Infrastructure is relevant to environmental and economic journalism; decision-making algorithms, to judicial and political journalism; automation, to labor journalism; training data, to cultural journalism. Confining the topic to the gadgets page makes it easier for companies to define the vocabulary and for consequences to arrive late.

Haiku functions as a mirror, with the limitations inherent to any monitoring and automatic classification system. What appears in it is an abundant conversation, still dependent, in which foreign companies put the protagonists, local institutions appear only with already-completed actions, and entertainment takes care of familiarity.

The press in Costa Rica talks a lot about artificial intelligence. Too often, the script arrives already written from elsewhere. Public conversation will mature when national decisions are visible before they close and when journalism follows technology to its consequences. Until then, the country will continue reading a future drafted somewhere else and asking it, every now and then, for the winning number.

How to read this data?

The analysis integrates 9,025 deduplicated notes from digital versions of different Costa Rican media outlets published between January 1 and September 13, 2026. It excludes live broadcasts from podcasts, radio, television, and 87 records from institutional broadcasters, among them MICITT, CAMTIC, CICR, and PROCOMER, which don't constitute journalism. The identification of terms and institutional mentions correspond to verifiable counts. The tone, frames, relevance, and some entities were classified by an AI model developed by the author and serve to compare internal patterns, not to judge an individual note or editorial quality.

The corpus combines two monitoring efforts with different logics: a general feed for the first part of the year and keyword tracking from late July. For that reason, coverage volume data was not presented. Nor does it measure audience, readership, or effect on public opinion. In a manual validation of a sample of 60 notes, the frame classifier registered a 17% error rate, concentrated in nearby categories. Broad differences between frames retain interpretive value.

Source: Haiku, analysis of artificial intelligence coverage in Costa Rica, 2026.