On the North/South asymmetry in generative AI as a cultural infrastructure
When a generative AI system is tasked with depicting 'a city' without further qualification, the output is not a generic city: it is glass skyscrapers, squares with ornamental fountains, paved streets with painted lines, and urban furniture of European or North American style. The 'neutrality' of the prompt activates, by default, a specific cultural imaginary. When the same system is asked to depict Jakarta, Lagos or Lima, it typically inserts Western architectural elements into the result, as if the non-Western were merely a variation upon a background whose nature is Western.
This text examines that operation, its material mechanisms and its consequences. My thesis is that the North/South asymmetry in generative AI is not a problem of 'lack of diversity' in training data—a formulation dominant in current institutional reports, which is descriptively correct but analytically weak. It is something more structural: the active presence of a specific architecture of cultural power that operates under the language of technical neutrality. It constitutes, today, the most efficient form of cultural hegemony that human history has produced, precisely because it presents its bias not as bias but as unmarked universality.
This is the fourth text in a series on digital infrastructures that I have been developing in this public notebook. In Mass Technology and Ethical Consumption, I argued that responsibility for digital sustainability is structural, not individual. In When the Frame Paints the Picture, I showed how institutional infrastructure produces artistic experience rather than merely influencing it. In The Price of the Model, I argued that what occurs with cultural labour under generative AI is a displacement of value, not a fragmentation of work. Here, I extend the argument to its geographical and colonial dimension: if generative AI infrastructures actively shape cultural production and experience, the immediate question is who shapes them, from where, on what materials, and with what consequences for the territories excluded from the process.
1. What 'non-neutrality' conceals
The 2025 UNESCO Report of the Independent Expert Group on Artificial Intelligence and Culture opens with a statement that appears sober yet conceals a euphemism: AI algorithms 'are not culturally neutral'. The formula is correct and, nevertheless, problematic. To state that they are not neutral suggests that they could be if biases were corrected. It suggests that neutrality is the horizon and partiality the deviation. It suggests that the problem is technical, susceptible to adjustment.
The most accurate description is the opposite. Generative AI systems do not fail to be neutral: they are built upon a specific cultural position that operates as neutral because it possesses sufficient power to do so. Neutrality is not a horizon that the system could approach by improving data; it is the political effect produced by the hegemonic position when it presents itself as an absence of position.
This is no academic nuance. It radically changes the operational question. If the problem is 'the models are not neutral', the reasonable response is 'let us make the models more neutral through more diverse datasets'. If the problem is 'the models have a specific cultural position that operates as neutral', the reasonable response is very different: to question the very operation by which a concrete cultural position is established as the zero point of meaning. The two responses lead to different policies, fund different projects and produce different results.
It is prudent to keep this conceptual shift in mind before examining the data.
2. What the data indicate
Empirical evidence regarding cultural asymmetry in generative AI has multiplied over the last two years and converges on an unequivocal pattern. Five findings are particularly relevant.
Linguistic composition of training. Hugging Face, developer of Stable Diffusion, has acknowledged that the 'vast majority' of its training data is in English. Open models such as LLaMA 2 and Mistral perform significantly worse in languages like Nigerian Igbo or Kazakh than in majority languages—not due to architectural limitations of the model, but due to the relative scarcity of data available in those languages within the training corpus. The world's majority languages in terms of speakers—Mandarin, Hindi, Spanish, Arabic, Bengali—are systematically underrepresented in proportion to their actual demographic weight and far below the weight of English in datasets. The distribution does not reflect population distribution or cultural distribution: it reflects the distribution of digital content available under licences compatible with mass extraction.
Demographic biases in outputs. A systematic analysis of images generated by Midjourney found that only 23% of the people depicted are women and only 9% are people of African descent. These figures do not reflect actual population distribution or the demographic distribution of content available online. They reflect the specific composition of the training corpus, which privileges certain types of work, certain aesthetics, and certain representations over others. The model does not produce a world: it produces a specific world that presents itself as the world.
National stereotypes. A study published in Rest of World on three thousand images generated by Midjourney with prompts adapted to different countries documented unequivocal patterns: 'an Indian' appears almost always as an older man with a beard; 'a Mexican person' frequently as a man with a sombrero; urban landscapes of New Delhi appear systematically 'dirty'; food from Indonesia is represented almost always on banana leaves. The system does not represent the complexity and heterogeneity of these cultures: it flattens them into concrete stereotypes that circulate in Western corpora regarding those places. Generative AI does not invent these stereotypes. It reproduces, amplifies and normalises them at a scale that previous human representations had never reached.
Cultural differences between models. A comparative study between GPT, a model from OpenAI, and ERNIE, a model developed by Baidu in China, found that the same questions posed in English and Chinese respectively produce markedly different answers. In English, the models tend to emphasise an individualistic orientation; in Chinese, a collective orientation. This is not a defect of the models: it is the trace of the corpora upon which they were trained, which reflect the cultural contexts in which those corpora were produced. The operational consequence matters: generative AI is not a single system with biases; it is multiple systems, each with its own cultural configuration, presenting themselves simultaneously as universal in their respective markets.
Biases in specialised representations. A study published in npj Digital Medicine in 2025 documents that image generators produce portraits of patients with specific diseases that overrepresent white people of normal weight, in proportions disproportionate to the actual epidemiological distribution of those diseases in the population. This is not a cosmetic problem: when these systems are incorporated into educational workflows, clinical materials, or medical communication tools, they propagate a specific imagery of the 'normal patient' that has material consequences on how disease is recognised or not recognised in racialised bodies or bodies with corporalities different from the represented norm.
The five findings are not independent aspects. They form a pattern. The pattern is the systematic operation by which a specific cultural configuration—predominantly Anglo-American, male, white, and urban middle-class—presents itself as the zero point from which everything else is measured. What lies outside that configuration appears as a difference from the norm, not as a cultural configuration on equal footing.
3. Algorithmic hegemony is no accident
It is worth examining why the composition of the corpora is as it is. There are three material reasons that matter because they change where possible intervention lies.
First. Differential digital availability. Not all cultures, languages and productions of the world have the same degree of digitisation. The English language possesses centuries of textual production digitised by massive projects (Project Gutenberg, Anglo-Saxon university archives, extensively digitised press). The majority languages of the Global South possess much smaller corpora in proportion to their number of speakers. Minority languages and especially oral languages without extensive written tradition are virtually invisible to the data scraping processes upon which current models are based. The digitisation of the world is not uniform: it is the product of specific historical investments, of institutions that decided what to archive and what not to, of legal frameworks that permitted or prevented extraction.
Second. Access under compatible licence. Even when digitised content exists, not all of it is available under conditions that permit its incorporation into training datasets. The American doctrine of 'fair use' has functioned de facto as broad permission for massive extraction of content published on the internet. Other jurisdictions—the European Union, various countries of the Global South—maintain more restrictive frameworks. The operational effect of this legal asymmetry is perverse: the more permissive a country's legal framework is regarding extraction, the more content from that country ends up in training corpora; the more protective, the less. The legal protection of one's own cultural production translates, paradoxically, into exclusion from the algorithmic horizon.
Third. The geographical location of development. The companies developing mass-market generative models are located in a small number of countries and cities: San Francisco, Seattle, Beijing, Hangzhou, and a few European enclaves. Their teams for development, evaluation, fine-tuning, and testing operate from these geographies. Decisions regarding what to include, what to exclude, what to label as problematic, and what to adjust through human reinforcement are taken from these contexts. Even when there is an explicit desire to incorporate cultural diversity—as has been the case in several recent initiatives—the horizon from which it is evaluated what is diverse, what is representative, and what is offensive remains culturally localised. Diversity incorporated under Western evaluation remains, ultimately, diversity legitimised by a Western gaze.
These three reasons are not neutral in relation to one another. They reinforce each other. Unequal digitisation produces unequal corpora. Unequal corpora operated from specific geographies produce models with specific biases. Models with specific biases are presented as neutral in global markets. Global markets assimilate these biases as the new zero point. Subsequent corpora, partially fed by the outputs of current models, incorporate this already consolidated zero point. The cycle closes.
To call this 'bias' underestimates what is occurring. 'Bias' suggests a correctable deviation from a neutral axis. What is occurring with generative AI is the active construction of the axis. It is not a biased system: it is a system that produces, at a global scale and at unprecedented speed, a specific cultural axis that subsequently recognises itself as universal.
4. Why it matters: double, not additive, gap
The consequences of this asymmetry operate on two planes that do not add up arithmetically but rather multiply.
Economic level. UNESCO data documents that the majority of the benefits of global cultural trade are concentrated in a few high-income countries. Although countries in the Global South doubled their exports of cultural goods in recent years, they still account for little more than 20% in digital cultural services. The 24% projected drop in music creator revenue by 2028 that I noted in The Price of the Model is not distributed uniformly among all creators worldwide. It falls most heavily on the most vulnerable economies, on creators with the least collective bargaining power, and on cultural sectors with the weakest institutional infrastructure. Creators in the South were already at a disadvantage in the global cultural market; generative AI amplifies that disadvantage rather than correcting it.
Symbolic level. Here the displacement is more insidious. Jill Walker Rettberg, in Issues in Science and Technology (2024), warns that without specific interventions, generative AI will tend to 'reduce cultural diversity', standardising human expression according to predominantly American patterns incorporated into the corpora. She gives the example of the Norwegian story Cardamom Town, a local narrative with a specific tradition, which could be buried if chatbots begin to water down their responses with homogenised American narratives. Multiplied on a global scale, this means the progressive loss of narrative, aesthetic, and symbolic heritage not through material destruction—archives do not disappear—but through algorithmic invisibilisation. What AI does not reproduce ceases to circulate. What ceases to circulate atrophies. What atrophies over a generation becomes unrecoverable as a living reference, even if it remains as an archival object.
The multiplication between the two planes operates as follows. A creator from the Global South has less visibility in the dominant generative market because the aesthetics of their tradition are not the aesthetics that the model reproduces with fluency. This reduces their economic access to the global circuit. Reduced economic presence reduces their capacity to exert pressure on platforms, all of which are operated from distant jurisdictions. Reduced political pressure reduces the probability that platforms will invest in improving the representation of their tradition. Reduced investment perpetuates the asymmetry in the corpus. And so on.
It is not a symmetrical vicious circle. It is a spiral. Each turn aggravates the asymmetry of the previous one. And it operates over short timeframes: what took decades or centuries for historical cultural processes occurs in months under the current regime.
5. The language of diversity is not enough
The UNESCO document to which I have been referring proposes reasonable measures: cultural data as a common good, multilingualism, transparency in datasets, and international collaboration. These proposals move in the correct direction. However, their language—the language of 'cultural diversity' as a good that must be protected—has a significant analytical limitation that should be noted.
The diversity framework presupposes that the issue is additive: that the problem is the lack of presence of non-dominant cultural traditions in the systems, and that the solution lies in adding those traditions to the existing corpus. This produces quota-style policies: x% of work in minority languages, y% of diverse representations, z% of content from the Global South. Such policies are better than their absence. But they do not address the central problem.
The central problem is not that diverse content is lacking. It is that the entire architecture of the system—from the choice of what data to extract to how outputs are evaluated, including what constitutes 'good quality' in a model—is constructed from a specific position. Adding diverse content to the corpus without questioning the position from which that content is evaluated produces 'assimilated diversity': non-dominant traditions incorporated into the system under the evaluation criteria of the system itself. It is inclusion without transformation of the rule of inclusion.
To illustrate the problem: including Quechua poetry in the training corpora of a large model improves linguistic representation quantitatively. But if the criteria by which the model learns to generate 'good poetry' are trained on Western tradition—syllabic scansion, rhyme, metre, canonical genres of European poetry—the model will learn to generate what appears to be Quechua poetry according to the aesthetic criteria of Spanish or English. The diversity of the language is incorporated; the diversity of the aesthetic regime is not. The result is technified assimilation: apparent Quechua poetry, evaluated from outside the tradition that names it.
Resolving this requires something other than increasing diversity. It requires questioning the architecture of evaluation itself. And that is conceptual and political work, not technical.
6. What they are doing, and where their limits lie
There are genuine initiatives attempting to address the problem from non-Western positions. It is appropriate to mention them with honesty: neither to idealise them nor to minimise them.
Masakhane is a research community distributed across Africa that works on natural language processing for African languages. It brings together researchers in South Africa, Nigeria, Kenya, Ghana, Senegal, and other countries, develops datasets in languages such as Yoruba, Igbo, Swahili, and Amharic, and publishes work in international academic venues. It is one of the most robust initiatives from the Global South in this field.
Te Hiku Media in Aotearoa-New Zealand has developed speech recognition systems for Te Reo Māori under explicit principles of indigenous data sovereignty. Its governance model—whereby data belongs to the Māori community and cannot be used without their consent—is an international benchmark for how AI can be developed while respecting cultural power structures distinct from hegemonic ones.
Ai4D and Lelapa AI are African initiatives working on AI infrastructure designed from and for African contexts, with international public and local private funding.
Latin American hubs. Various efforts in Argentina, Brazil, Mexico, Colombia, and Chile are working on the evaluation of linguistic biases in American Spanish and Brazilian Portuguese within large models, as well as on the development of corpora for Latin American indigenous languages.
These initiatives are important and deserve to be recognised. However, it is also honest to acknowledge their structural limitations. They all operate with small budgets compared to large AI corporations, with partial dependence on funding originating in the North, with computational infrastructure rented from the very companies whose hegemony they seek to counterbalance, and with publication in academic venues whose evaluation criteria are located on the same cultural axis as the problem itself. These are real material conditions that cannot be resolved through mere voluntarism.
Serious intervention in the North/South asymmetry problem in generative AI requires something that these initiatives, however robust they may be, cannot achieve alone: changing the political economy of global computational infrastructure. That is to say, changing who controls large-scale computing resources, who defines technical standards, who funds research, and who captures the rent. Without that fundamental transformation, initiatives from the South operate as a rearguard action: necessary and valuable, but structurally insufficient to reverse the asymmetry.
7. The consolidation of 2026: power is concentrated while diversity is debated
While the public conversation regarding AI and culture has advanced in the language of diversity and representation, the actual market has consolidated its architecture in a direction that is worth examining with precision. Four data points from 2026 are particularly revealing.
On 31 March 2026, OpenAI closed a funding round of 122 billion dollars, presenting it as a lever for computational infrastructure, distribution, and deployment services. On 11 May, it launched OpenAI Deployment Company, a specific division to embed implementation teams within enterprise clients, and declared that it had surpassed one million corporate users of its products. Anthropic allocated 100 million dollars in March 2026 to its partner network. Microsoft announced the general availability of its enterprise agent platform on 9 March; Google pushed its Gemini Enterprise Agent Platform in April. The signal is not merely one of investment: it is one of consolidating control over the deployment, governance, and operation layer of systems on a global scale.
Europe has simultaneously landed its regulatory and industrial capacity agenda. The European Commission announced in April 2026 that 19 AI factories are already operating on the continent, with gigafactories in preparation. The AI Act begins its enforcement on 2 August 2026; the obligations regarding general-purpose models entered into force a year earlier. It is a genuine regulatory will, without precedent in other jurisdictions at this scale.
The important point here, for the argument of this text, is what these data reveal in their geographical overlap. The concentration of power in deployment, distribution, and operation at scale is occurring in the same poles that already dominated the corpus: the United States first, China second, and Europe in an ambivalent position—possessing strong regulatory capacity, intermediate industrial capacity, and high infrastructural dependence. The Global South does not feature in these consolidation figures, neither as a headquarters, nor as a massive deployment actor, nor as a governance agent. The initiatives mentioned in the previous section—Masakhane, Te Hiku Media, Lelapa AI—operate on a different plane of magnitude.
The consequence is that while international frameworks discuss how to ensure 'cultural diversity' in generative AI systems, the material power to define what is considered good AI, what is considered responsible deployment, and what is considered effective governance is being concentrated in very few geographically localised actors. Diversity is discussed; infrastructure is built. And infrastructure is built faster and with more resources than the discussion.
This is not an apparent contradiction: it is the concrete mechanism by which hegemony is reproduced in digital systems. Not through direct political imposition, but through speed of deployment, scale of investment, and the capture of the operational bottleneck, while the conversation regarding alternatives advances at a different cadence, with different budgets, and from different geographies.
8. The operational question: who, from where, about what
I shall reformulate the question posed by the title of this text: who trains the world. If the North/South asymmetry in generative AI is not an accident correctable through technical adjustments but an active architecture of cultural power, then the operational question decomposes into three concrete questions that any serious public policy on this matter should answer.
First. Who decides what enters the training corpora. Today, the answer is: the technical teams of the companies developing the models, without a generalised obligation for transparency regarding the criteria applied or consent from the communities whose cultural productions are incorporated. This is unsustainable. A responsible generative AI infrastructure would require: mandatory transparency regarding the geographical, linguistic, and cultural composition of the corpora; effective mechanisms for consent from the communities whose productions are incorporated; and independent audits regarding cultural representation in the outputs.
Second. From where are the outputs evaluated. Today, the processes of 'refinement through human feedback'—which adjust models after initial training to make them more useful, safe, and acceptable—operate primarily from localised geographies, with personnel hired in specific contexts, according to criteria defined by teams that are not culturally representative of the model's global audience. This should change: evaluation processes should include, in a structural and non-decorative manner, cultural agents situated in the geographies whose representation is at stake, with real capacity to modify criteria rather than merely approve results.
Third. Upon what materials is operation conducted and how is their revenue distributed. This connects directly to the argument of The Price of the Model. If training corpora include cultural production from the Global South—and they do, albeit in a fragmentary manner—legal and economic frameworks must ensure that the resulting revenue returns, in significant part, to the communities whose productions allowed the model to be trained. Today, this does not happen practically at all. Extraction is free; exploitation is private.
These three lines of intervention are not novel. Various institutions, from UNESCO to civil society organisations in the Global South, have articulated them with varying degrees of precision. What is lacking is not diagnosis. What is lacking is the international political will to impose them, and that will shall only emerge if the symbolic and political costs of inaction become greater than the economic costs of action. Whilst public discourse on AI remains reduced to a debate on 'biases' and 'diversity' without addressing the structural question of who owns and controls the infrastructure, those symbolic costs will not increase sufficiently.
9. Conclusion
Three conclusions follow from the analysis.
The first point is descriptive. The North/South asymmetry in generative AI is not a problem of a lack of data diversity nor an accident correctable through technical adjustments. It is the active operation by which a specific cultural configuration—predominantly Anglo-American—is established as the zero point of meaning in systems that present themselves as universal. The industrial consolidation of 2026—massive investment rounds, enterprise agent platforms, concentration of deployment among very few actors—reinforces this material operation rather than counteracting it. Hegemony does not operate under the name of hegemony: it operates under the name of neutrality. This is historically new in terms of scale and speed, although conceptually it continues the logic of the colonial cultural order under new media.
The second is methodological. The institutional language of 'cultural diversity' that dominates UNESCO frameworks and current international policies is necessary but insufficient. It presupposes that the problem is additive—that diverse content is missing—when in reality it is structural—the entire architecture of the system is built from a specific position that reproduces itself even when it incorporates diverse content. Without shifting the conceptual framework of the problem, the resulting policies will produce assimilated inclusion, not transformation.
The third point is ethical. Responsibility for cultural asymmetry in generative AI does not lie with the creator in the Global South who chooses to use Western tools due to a lack of alternatives, nor with the community whose output is assimilated without consent, nor with the state in the South that lacks sufficient computational infrastructure to build its own alternatives. It lies with the actors who design, finance, and operate these infrastructures from specific geographies with the actual capacity to choose how to do so. And, secondarily, it lies with the international political order that allows these actors to operate without effective obligations of reciprocity, consent, or redistribution.
Whoever trains the world, today, trains it in the image and likeness of very few places. And the trained world tends, by the very mechanics of the system, to recognise itself in that image as if it were its own. Recognising this is not nostalgia or identity-based resistance. It is the prerequisite for any international cultural policy on generative AI to have meaning. Whilst public discourse remains trapped in the language of correctable biases and additive diversity, the decisions that truly move the needle—who owns the infrastructure, who captures the yield, who defines the evaluation criteria, who decides what is incorporated and what is excluded—will continue to be made far from the territories whose culture is being reorganised.
On the open conversation
This text concludes, for now, a series of four pieces on digital infrastructures that I have been developing in this public notebook: sustainability and ethical consumption, institutional validation and artistic experience, the displacement of value in cultural labour, and now the geographical distribution of cultural power in the era of generative AI. All four share the same underlying thesis: in contemporary digital systems, the decisive factor is not individual conduct but infrastructural design, and the ethical responsibility for its consequences is distributed in proportion to the power of those who design, finance, and operate these systems.
If anyone wishes to intervene from the fields of postcolonial studies, the political economy of culture, AI studies, the Global South projects mentioned here, or any other relevant perspective, this notebook remains open.
Sources
UNESCO. Report of the Independent Expert Group on Artificial Intelligence and Culture. UNESCO, 2025. https://www.unesco.org/sites/default/files/medias/fichiers/2025/09/CULTAI_Report%20of%20the%20Independent%20Expert%20Group%20on%20Artificial%20Intelligence%20and%20Culture%20(final%20online%20version)%201.pdf
UNESCO. “A new expert report explores how AI is transforming culture”. UNESCO, noviembre de 2025. https://www.unesco.org/en/articles/new-expert-report-explores-how-ai-transforming-culture
Creatives Unite. “AI set to slash creators’ revenues by up to 24% by 2028, UNESCO warns”. 2025. https://creativesunite.eu/article/ai-set-to-slash-creators-revenues-by-up-to-by-unesco-warns
Algorithm Watch. “Cultural Hegemony: How Generative AI Systems Reinforce Existing Power Structures”. sustAIn Magazine. https://sustain.algorithmwatch.org/en/cultural-hegemony-how-generative-ai-systems-reinforce-existing-power-structures/
MIT Sloan. “Generative AI isn’t culturally neutral, research finds”. MIT Sloan Ideas Made to Matter, 2024. https://mitsloan.mit.edu/ideas-made-to-matter/generative-ai-isnt-culturally-neutral-research-finds
Rest of World. “Generative AI like Midjourney creates images full of stereotypes”. 2023. https://restofworld.org/2023/ai-image-stereotypes/
Rettberg, Jill Walker. “How Generative AI Endangers Cultural Narratives”. Issues in Science and Technology, 2024. https://issues.org/generative-ai-cultural-narratives-rettberg/
npj Digital Medicine. “Demographic inaccuracies and biases in the depiction of patients by artificial intelligence text-to-image generators”. 2025. https://www.nature.com/articles/s41746-025-01817-6
Foro Económico Mundial. “Cómo el Sur Global está reimaginando el futuro de la IA”. 2026. https://es.weforum.org/stories/2026/02/como-el-sur-global-esta-reimaginando-el-futuro-de-la-ia/
CulturaLAB. “Inteligencia artificial y cultura”. 2025. https://culturalab.es/inteligencia-artificial-y-cultura/
Masakhane. Project site. https://www.masakhane.io/
Te Hiku Media. Information regarding the AI project in te reo Māori. https://tehiku.nz/te-hiku-tech/
OpenAI. Accelerating the next phase of AI. 31 March 2026. https://openai.com/index/accelerating-the-next-phase-ai/
OpenAI. Launching the Deployment Company. 11 May 2026. https://openai.com/index/openai-launches-the-deployment-company/
Anthropic. Claude Partner Network. 13 March 2026. https://www.anthropic.com/news/claude-partner-network
Microsoft. Introducing the first frontier suite built on intelligence and trust. 9 March 2026. https://blogs.microsoft.com/blog/2026/03/09/introducing-the-first-frontier-suite-built-on-intelligence-trust/
Google. Google Cloud Next ’26. 22 April 2026. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/next-2026/
European Commission. AI Continent Action Plan delivers major milestones. 9 April 2026. https://digital-strategy.ec.europa.eu/en/news/ai-continent-action-plan-delivers-major-milestones
Esteban Ruiz, J. A. Mass Technology and Ethical Consumption. Public notebook at juanesteban.art, 2026.
Esteban Ruiz, J. A. When the Frame Paints the Picture. Public notebook at juanesteban.art, 2026.
Esteban Ruiz, J. A. The Price of the Model. Public notebook at juanesteban.art, 2026.
Series · Infrastructures
This piece forms part of a six-part series examining the influence of digital and cultural systems upon art. The overarching thesis posits that within contemporary systems, the decisive factor is not individual conduct but rather infrastructural design. The series commences with the capture mechanisms of these systems and concludes with their exclusionary practices.