When AI dreams of Havana: how algorithmic imagery is rewriting Cuba's past
On 9 June 2026, Cuban-language Telegram channels began circulating an essay-length thread under a punchy banner: "AI returns Cuba to the 19th century." The piece, posted by CubaDebate at 12:15 UTC, draws on a 2026 thesis by the historian Michael J. Bustamante and argues that generative image tools, when prompted with the word "Cuba," keep producing an oddly period-bound country: columns, wrought-iron balconies, colonial-era dresses, horse-drawn carriages, the warm sepia of an old engraving. Not the actual Cuba of 2026, with its Soviet-era tower blocks, its dollarised corner shops and its diasporic youth remixing reggaetón over WhatsApp voice notes. The thread is short, but the argument lands a heavier punch than its 200-odd characters suggest: the way algorithms picture a country is itself a kind of foreign policy, and Cuba, in this telling, is being quietly re-fantasised into a pre-revolutionary postcard.
The claim is not that AI "thinks" Cuba is old. It is that the training data overwhelmingly is old — and old in a particular way. Stock photography libraries, travel magazines and 19th-century engravings are over-represented in the corpora that feed the most-used image generators, while contemporary Cuban photography, blocked for decades by the US embargo's information side-effects and by Cuba's own connectivity constraints, sits in thinner piles. The result, the thread argues, is a hallucination that flatters the algorithm's colonial archive and erases a sovereign present.
A historian's thesis, distilled
Bustamante, a Cuba-born historian who teaches at Florida International University and has written extensively on Cuban memory and migration, has been making a related case in academic venues for several years. The Telegram thread summarises the move in plain terms: the algorithmic "default Cuba" is a Cuba that never existed as a whole — not for the enslaved people who built the sugar economy the engravings romanticise, not for the Chinese-Cuban workers of the 19th-century coolie trade, not for the revolutionary generation that followed. The image generator, in other words, doesn't just lag the present; it actively compresses a country into a sliver of its own tourist-memory.
This is not a uniquely Cuban problem. Comparable critiques have been levelled at how image models default to African savannahs when prompted with country names, at the generic "Asian city" they produce for any prompt mentioning Tokyo, Bangkok or Hanoi, and at the European-cathedral bias that flares up whenever a user asks for a "church" without further specification. What makes the Cuba case sharp is the political weight already loaded onto the country's image. For more than six decades, the visual iconography of Cuba — the cigar, the classic American car, the moustachioed guerrilla — has been contested ground, with Miami, Havana and the diaspora each laying claim to the canonical frame. AI now adds a fourth claimant, and it does so without anyone in particular intending it.
The counter-read: nostalgia, training gaps, or design choice?
There is a more charitable reading. Image generators are statistical machines; they learn what the internet shows them, and the internet has, in plain numerical terms, more old pictures of Cuba than new ones. The Travel & Tourism industry, the Hemingway industry, the Buena Vista Social Club industry — all have spent decades pushing a particular Cuba into the global visual archive. From this angle, the algorithm is simply being honest about its diet.
That defence only goes so far. Other countries with similarly old-skewing archives — Greece, Egypt, Italy — nevertheless produce a wider visual range in generative outputs, because travel publishers, news outlets and everyday users have pumped a steadier stream of contemporary imagery into the training pool. Cuba's contemporary visual record is thinner in part because of US policy: the embargo has long had a secondary information embargo attached, with US reporters historically restricted in what they could shoot and ship back. Cuban state media, for its part, projects its own curated present. The two streams rarely meet, and the gap is exactly where the model's defaults rush in.
A second counter-narrative insists that the problem is overrated: users can prompt their way out of it, typing "modern Havana," "2026 Cuba," "Soviet microdistrict in Centro Habana" and getting closer to the mark. The thread's reply is unsparing — that the burden of correction should not fall on the user, and that the very act of needing to type the year is itself the bug. Default outputs are normative outputs. They tell the world what a place "looks like" before anyone has asked a question.
The structural frame: who owns the image of a country?
The deeper question is not technical but political. Algorithmic image generation has, in the space of three years, become a low-cost mass producer of visual stock. Newsrooms, travel publishers, educational slide-decks and small businesses all draw on it; the prompt has replaced the photo commission. That makes the model — and the curators of its training data — a de facto ministry of visual culture for the parts of the world whose own image production has been thinned by embargo, under-investment or war.
The pattern is recognisable from earlier media transitions. The 19th-century illustrated press, the 20th-century wire services, the early-2000s rise of stock-photo libraries — each in turn set the canonical look of places that could not afford to set it themselves. The current shift is faster and quieter: there is no masthead, no byline, no editor a Cuban photographer can write to. The defaults are set, and the defaults favour the archive that already had the most pictures.
For the Cuban government in Havana, the implications are uncomfortable but not unfamiliar. The official iconography — Che, the schoolchildren, the May Day parade — has long been pushed back against by Miami's memory entrepreneurs. Algorithmic nostalgia, in this reading, is a third force that erases both the revolutionary present and the diasporic present in favour of an older, more saleable Cuba. For Cuban artists working in Miami, Madrid and Mexico City, the concern is more practical: their work, contemporary and self-defined, has to swim against a stream of free, plausible-looking, period-bound substitutes.
Stakes: the user, the country, the prompt
If the thread's diagnosis is right, the stakes are larger than a curiosity about image quirks. Countries are increasingly sold to investors, tourists and diaspora kin through a visual layer that no national tourism board fully controls. When the cheapest, fastest way to picture a place is to type its name into a generator, the output of that generator shapes first impressions in boardrooms, classrooms and family group chats. A Cuba that looks like 1890 is a Cuba that can be patronised, period-dressed and set to a soundtrack of son cubano — and, perhaps more consequentially, a Cuba that is harder to take seriously as a sovereign actor with a 21st-century economy.
The remedies being floated are familiar from other AI-governance debates. Training-data transparency — disclosure of the sources and weighting behind image models — is the most-cited, and the hardest to enforce. Compensated data partnerships with Cuban photographers and institutions, including state ones, would address the gap directly, though the political optics would be delicate. A lighter-touch option is geographic and temporal prompting baked into the tools themselves, the way some models now refuse or label certain categories of output.
What none of these options can do is make the question go away. The Telegram thread is unlikely to be the last word, and Bustamante's thesis is unlikely to be the last academic treatment. The harder question — who gets to set the default image of a country, and on whose archive — sits at the intersection of platform governance, cultural sovereignty and the long tail of the information embargo. Cuba is the example on the table; the pattern is wider.
What the sources leave uncertain
The thread is brief and the thesis summarised, not quoted at length. Independent verification of Bustamante's specific 2026 argument — its venue, length, and exact claims — was not possible from the material in hand. The structural claim about training-data imbalance in Cuba's case is well-attested in adjacent literatures on Global-South image representation, but a direct, primary-source link to a peer-reviewed paper remains to be located. Readers should treat the specific historical and visual examples sketched above as illustrative of a documented pattern rather than as findings unique to this article. Where the picture is genuinely thinner, the honest move is to say so.
Desk note: Monexus framed this piece around the cultural-sovereignty question raised by the CubaDebate thread, with the algorithmic-defaults argument foregrounded over a purely technical critique. The wire-style approach would have led with product-launch news; the longer read tries to put the image generator in line with earlier visual-archive transitions and to take both the nostalgia defence and the user-correction defence seriously before weighing them.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://t.me/cubadebate