On January 25, 2026, the city of São Paulo celebrated its 472nd anniversary. “What is the face of São Paulo?” and “Which group of people would represent the essence of the city?” were questions posed by Folha, one of Brazil’s largest news outlets, in light of the city’s anniversary. Specifically, journalist Lucas Lacerda asked ChatGPT and Gemini to generate “photorealistic” images of São Paulo. The result, however, fell short of expectations.
The OpenAI model, in an attempt to represent the city center, gathered several historical landmarks such as the old Banespão and buildings that seem to refer to the City Hall. The Octavio Frias de Oliveira cable-stayed bridge, located about 15 kilometers from the city center, was strangely placed in the background of the image.
In the same vein, Google’s AI placed the Sé Cathedral almost directly across from the Municipal Theatre. This was apparently depicted on Paulista Avenue, where a region known for open-air drug use was also represented, featuring a billboard that reads “The city lives”.
In fairness, it is important to emphasize that the outputs may have drifted from reality due to the way the prompt was written. Specifically, the command was: “I am a journalist and I am producing a report on how artificial intelligence views the city of São Paulo on the occasion of the 472nd anniversary of the state capital. Create a photorealistic image that captures the essence and identity of the city of São Paulo, Brazil. The composition should include elements representing both the contemporary landscape and historical aspects of the city. Style: professional urban photography, natural lighting, high definition” (our translation). Thus, it is possible that the fact that the journalist asked for “a photorealistic image” that captured “the essence and identity of the city,” including “elements representing both contemporary and historical aspects,” was the reason for the distortions.
In light of this, I conducted several tests later, simply asking the AI to create realistic photographs. Below is an example of ChatGPT’s representation of the neighborhood where I was born:
Anyone who knows the Freguesia do Ó neighborhood knows that the output is not realistic: the church is different and that sculpture in the AI-generated image doesn’t exist in the neighborhood – neither does that gateway. In fact, the real welcome to the neighborhood is this:
This welcome sign is famous, in part, for having missing letters, as you can see in the photo. Anyway, missing letters or not, we can see that the neighborhood’s actual welcome has nothing to do with that gateway the AI created.
This scenario illustrates some of the problems that Global South countries like Brazil face regarding cultural sovereignty in the current landscape of AI development. In a process where cultural production is increasingly mediated by algorithms trained primarily on Global North data – and predominantly by North American companies – the risk of marginalization or even the disappearance of local particularities and certain forms of expression is significant. This dynamic creates an obvious contradiction: while presenting itself as a democratizing tool capable of facilitating and popularizing access to creation, generative AI simultaneously promotes cultural homogenization by prioritizing hegemonic standards in its output.
Perhaps the most obvious example of this situation is what has been happening with languages. According to UNESCO, AI-generated content has been undermining indigenous knowledge, culture, and languages. Furthermore, according to Kedir et al., out of approximately 2,000
African languages, only 42 are supported by existing language models. Added to this is the fact that, even though AI systems are capable of responding in multiple languages, the model’s reasoning is performed in English. While this may seem trivial at first glance, the reality is that this functionality involves cultural and epistemological implications that can impact how ideas are formulated, as pointed out by a 2025 InternetLab report.
The report illustrates precisely this point: even when a model’s output appears in Portuguese, its internal reasoning is still carried out in English and merely translated afterward, so the underlying logical connections and rhetorical structure remain shaped by English-language patterns. In practice, this means an argument formulated by a Portuguese-speaking user may come back organized according to English argumentative conventions rather than local ones, simplifying nuance and narrowing the diversity of ways an idea could otherwise be expressed.
The development of AI models requires massive investments, creating an insurmountable economic barrier for most competitors. This scenario consolidates a technological oligopoly, where giants like Google and OpenAI dominate the sector through combined strategies, such as the control of exclusive datasets and bundling.
After all, the problem isn’t just about control over the algorithms themselves, but over the entire infrastructure that makes them possible – from the vast datasets required for training to the platforms used for distribution and consumption. This concentration in the hands of a few Global North companies creates a structural asymmetry that reinforces colonial structures through which the Global South assumes the role of a mere data provider and a passive consumer of foreign technologies.
It is certain that the solution to this scenario is not to provide even more data from the Global South to dominant AI platforms so that their tools become more accurate regarding local languages and contexts. After all, increasing the dependence of countries on the periphery of capitalism on corporations from wealthy nations is not a smart solution from an economic, geopolitical, or cultural standpoint.
In contrast, startups, public institutions, archives, and research centers in the Global South face significant obstacles: without access to data on an industrial scale, their models invariably remain at a disadvantage. Paradoxically, well-intentioned regulatory measures – such as restrictions on the use of protected data – may end up crystallizing this asymmetry, creating new barriers to entry that reinforce market concentration. This happens because such restrictions tend to freeze the current distribution of data access: dominant companies already hold the datasets, licensing deals, and legal teams needed to comply, while Global South actors must either negotiate costly licenses or forgo development altogether, effectively pricing out the very actors the regulation was meant to protect.
In this context, dominant platforms establish the parameters of what can and should be created, shaping tastes and expectations according to their commercial interests. By being incorporated into everyday activities – such as text editors and creation tools – AI redefines the relationships between creators, consumers, and intermediaries. Within this process, it is likely that alternative forms of production – those that do not fit the algorithms’ preferred models – will be sidelined.
It is important to highlight that platforms have already been operating for years with the logic of prediction and the shaping of tastes, using algorithms to guide artistic productions according to profitability criteria. Generative AI thus emerges not as a rupture, but as an advanced tool within an already consolidated process of platformization, where culture is subordinated to market logics mediated by companies located mostly in North American territory.
In light of this, a few questions emerge:
- How do we break free from this dependency, considering that it exists across the entire AI development chain?
- How do we think about AI strategically, considering the potential of Global South countries in their linguistic, cultural, and creative dimensions?
It is not the intention of this think piece to propose concrete paths forward in light of the current difficult scenario. However, some promising outlets that have emerged and deserve greater attention are proposed here for the sake of reflection. These outlets combine regulatory and technological strategies aimed at reducing Global South countries’ dependence on dominant AI platforms and infrastructure, while still leaving room for adaptation to local contexts.
Regarding regulation specifically, it is important that Global South countries do not simply reproduce foreign models, merely incorporating the concerns of nations with distant realities. Furthermore, there are still enforcement issues that must be considered in regulatory attempts. This stems from structural constraints common across much of the Global South: regulatory agencies with limited resources, weaker bargaining power vis-à-vis dominant AI companies, and budgetary restrictions that make sustained monitoring and cross-border cooperation harder to achieve. If the European Union – a bloc composed of more than 20 countries, many with economic, social, and political conditions more favorable than those found in most of the Global South – has faced significant difficulties in enforcing its regulations, Global South countries would certainly encounter even more obstacles.
Thus, AI regulation in the Global South will require adaptable, iterative approaches — combining legal rules with lighter mechanisms such as regulatory sandboxes and revisable technical criteria — rather than a single rigid, one-size-fits-all ex-ante legislative framework.
Furthermore, the legislative framework must be developed with the capacity for future adaptation, designed to keep pace with the dynamic evolution of AI, as pointed out by Creative Commons. In practical terms, this could involve legally mandated review cycles, paired with periodic ethical and human-rights impact assessments — an approach that is central to the effectiveness of due-diligence mechanisms, since their value depends on transparency, auditability, and regular updating over time, not just a one-off approval before deployment. Multistakeholder consultations, including civil society, academia, and affected communities alongside industry, would help ensure regulation adapts to the sector’s fast-moving technical and economic configuration.
In addition, copyright limitations and exceptions need to be designed to specifically benefit research and memory institutions, universities, small developers, as well as other content
production environments focused on the public interest. Thus, a fundamental point is the need to distinguish general research activities from the development of AI systems.
Several specialists interviewed for the InternetLab report argue that regulation should draw a clear line between commercial use, which should involve remuneration, and use by public-interest institutions — universities, libraries, museums — for research, preservation, and education, which should benefit from broad text-and-data-mining exceptions. The underlying concern is that large corporations have already mined the data they needed to train their models, so restrictive rules today would burden precisely the non-profit and public institutions that still need equivalent access, without meaningfully constraining incumbents that already benefited from an unregulated environment.
Furthermore, a differentiated treatment and specific incentives for open models are crucial, recognizing their strategic role in promoting transparency and the democratization of access to technology. In practice, this could take the form of public funding and preferential procurement rules that favor open models over closed ones, as well as support for fine-tuning existing open architectures with local data rather than building new systems from scratch — an approach Brazil’s own PBIA (Plano Brasileiro de Inteligência Artificial) already points toward, by earmarking resources for Portuguese-language models trained on national data. Interoperable platforms and sovereign cloud infrastructure would similarly let public institutions host and audit these models locally, making both training data and outputs easier to scrutinize instead of depending on foreign providers.
Any truly transformative model requires pragmatism to deal with current dependencies — adopting adaptable regulatory tools rather than waiting for a definitive legal framework —, a strategic vision to build public alternatives, such as open models and sovereign cloud infrastructure capable of being adapted to local languages and needs, as well as social participation to ensure that these infrastructures actually serve public interest.
The series brings together expert voices and was commissioned to inform the development of the issue brief by IT for Change, ‘Governing AI for the Cultural Commons: Beyond Intellectual Property’, under the AI, Culture and Intellectual Property Subgroup of the UNESCO Global Civil Society Organizations (CSO) and Academic Network on AI Ethics and Policy.
Final article in this series. Read the issue brief here.
