An AI assistant may be accessible to millions of people without representing their languages, experiences and cultural references to the same degree. AfroféminasGPT and Ask Hakima show what changes when sources, context and intended users become explicit design choices. But they also raise a broader question: who decides which forms of knowledge enter the systems we increasingly use to access information?
An Afghan woman who wants to start a microbusiness. A Tanzanian student looking for information about agriculture in her own language. A researcher of African descent trying to understand how structural racism operates across digital media.
All three can open a chatbot, ask a question and receive, within seconds, an answer that is fluent, articulate and apparently relevant. But the formal quality of the response tells us very little about another issue: whether the system actually has access to the knowledge, linguistic categories and cultural references required to understand the context from which that question emerges.
This is one of the issues accompanying the growing use of generative AI as a means of accessing information. Large language models are not simply repositories of universal knowledge. What they are able to represent and return depends on the data used during training and post-training, the design decisions behind them, the instructions through which they are configured, the sources they may retrieve during an interaction and the context supplied by the user.
This becomes particularly significant when moving away from the languages and environments that are most extensively represented across the digital ecosystem towards those for which fewer online resources, datasets and linguistic infrastructures exist.
According to UNESCO, at least 40% of the approximately 7,000 languages currently in use are at risk, while fewer than 2% have a substantial digital presence through localised software, websites, translation systems, text-to-speech technologies or content-moderation tools.
The issue, therefore, is not simply whether someone can physically access a chatbot. It is also what happens once access has been achieved: how effectively can a system interpret a question embedded in a particular social, linguistic and cultural context?
It is within this space that projects built around a different design logic are beginning to emerge. These are not necessarily new large language models trained from scratch, but applications that use existing models while deliberately delimiting their sources, creating specific knowledge bases, shaping their instructions and designing the service around the needs of particular communities.
AfroféminasGPT and Ask Hakima belong to this category. They do not resolve the structural issues associated with general-purpose models. They do, however, make those issues easier to observe, because they turn into explicit design decisions elements that tend to remain largely invisible in larger systems: who the intended user is, which sources are considered reliable, which language is used and what contextual knowledge is required before an answer can be produced.
TAKEAWAYS
AfroféminasGPT: when source selection becomes an epistemological choice
Antoinette Torres Soler moved from Cuba to Spain in 2007. A philosopher, activist and founder of Afroféminas, she lives in Zaragoza, in Aragón, a region that in recent years has become one of Europe’s major hubs for data-centre investment.
The regional dimension matters if the phenomenon is to be understood accurately. By 2026, around 28 major data-centre projects had been identified across Aragón, while estimates reported by El País put announced investment between 2019 and 2025 at approximately €48 billion. Some facilities are already operational, while others have been authorised or remain under development.
It is within this context that Torres Soler developed AfroféminasGPT, a technological extension of a cultural project launched in late 2013 to create a space for representation and discussion around the experiences of Black women and people of African descent.
The starting point is an apparently straightforward question: from which perspective does a system define concepts such as racism, discrimination or coloniality?
According to Torres Soler, when a mainstream chatbot is asked to explain racism, the answer often concentrates on individual prejudice, discrimination or beliefs about superiority, while the structural and systemic dimensions of the phenomenon remain less visible.
AfroféminasGPT deliberately starts from a different interpretative framework. Its references come from Black thought, Black feminism and decolonial scholarship, including authors and thinkers such as Stuart Hall, Angela Davis, Frantz Fanon and Kimberlé Crenshaw. They are not used as supplementary references, but as part of the knowledge architecture through which responses are organised.
This distinction also matters technically.
AfroféminasGPT is not an independent foundation model developed from scratch. It is a tool built within the ChatGPT infrastructure and configured around a clearly defined epistemological framework. The project itself describes it as a form of “situated” AI, grounded in critical Black thought, Black feminisms, Afrocentrism and decolonial theories.
Among Torres Soler’s design decisions is the choice not to allow the system unrestricted access to information from the open web.
“The networks are full of racism. You decide which authors to use, what tone to adopt, whether to connect to the internet or not. I said no,” she explains.
Preventing web access does not, of course, erase what the underlying model has absorbed during training, nor does it automatically eliminate potential bias. It operates at a different level: it limits the sources the system can use during retrieval and makes it possible to construct a deliberately curated documentary perimeter.
In this sense, building the archive becomes more than a matter of content curation. Deciding which authors to include, which definitions to use as points of reference and which sources to exclude means making explicit some of the epistemological decisions that shape the response.
“Artificial intelligence reproduces the biases of the world that creates it. So we asked ourselves: what would happen if an AI were designed from Black and decolonial genealogies instead of simply absorbing the noise of the internet at scale?” Torres Soler says.
The project has expanded considerably. While AfroféminasGPT recorded just over 800 requests during its first few weeks, by early February 2026 it had exceeded 10,000 consultations.
More importantly, it has not remained an isolated experiment.
In February 2026, Afroféminas launched a Socratic GPT designed for educational settings. It subsequently developed Habla con Afroféminas, which can be accessed without a ChatGPT account. In August 2026, it launched Salud Negra IA, a project focused on the health of Black women and developed as an extension of the experience accumulated through AfroféminasGPT.
The trajectory is significant for another reason. Afroféminas has explicitly stated that, in the longer term, it wants to move beyond its dependence on the ChatGPT infrastructure and develop its own technological capabilities. In its initial phase, Salud Negra still uses a GPT because this is the solution compatible with the resources currently available; the stated objective, however, is to move towards greater technological sovereignty.
This shifts the discussion from inclusion to infrastructural dependence. A community may create its own knowledge base and define its own epistemological framework while still depending on the model, platform and economic conditions controlled by an external provider.
Ask Hakima: context is not an optional variable
The same issue takes a different form in Afghanistan.
Selene Biffi has worked for years at the intersection of technology, social entrepreneurship and fragile contexts. Through She Works for Peace, she developed Ask Hakima, an AI assistant designed for women who run, or are seeking to start, microbusinesses in the country.
Here, the starting point is not primarily cultural representation in written material, but the practical usefulness of an answer.
Asking a general-purpose chatbot how to obtain funding to start a small business in Afghanistan may produce information that is formally correct but poorly aligned with the circumstances in which the person asking the question actually lives: the province, access to financial services, family constraints, available networks, the regulatory environment and the restrictions that specifically affect women.
“In a context such as Afghanistan,” Biffi explains, “building trust, addressing subjects that may be considered taboo and understanding the specific difficulties women face are fundamental. But a general-purpose model does not account for these things, because it was not designed to do so.”
Here too, it is useful to distinguish between the model and the application.
Ask Hakima is not an LLM independently trained on Afghan data. The system integrates WhatsApp, ChatGPT and the n8n automation platform. When a question arrives, the workflow queries a knowledge base built by She Works for Peace, and ChatGPT generates the response using information contained in that database. The interface supports Dari and Pashto, Afghanistan’s two official languages.
The decision to use WhatsApp is itself part of the design. It is not a secondary interface choice, but a response to the practices and infrastructure actually used by the women the service is intended to reach.
Ask Hakima is also based on the experience the organisation has accumulated through direct work with women entrepreneurs. In the interview on which this article is based, Biffi reported that more than 12,400 microbusinesses had contacted the organisation for technical support between the first and second years of the project.
By April 2026, She Works for Peace reported having built experience through supporting more than 15,000 women, while 500 entrepreneurs were already using Ask Hakima directly.
The distinction between those two figures matters: the first measures the wider population reached through the organisation’s services; the second refers specifically to use of the AI assistant.
This is where Biffi introduces another dimension of the issue: digital inclusion cannot be reduced to the abstract availability of a technology.
“Women and men do not have the same opportunities for access and use, and ignoring that difference in the demographic data we begin with produces tools that appear universal but are not,” she notes.
In a context characterised by strong social and gender asymmetries, designing around an average user risks concealing precisely those conditions that determine whether a tool can actually be used.
This does not imply that every community necessarily needs its own chatbot. It does, however, raise a design question: what contextual information must be available for a system to produce an answer that is genuinely useful?
And to what extent can a general-purpose model incorporate that complexity?
Biffi does not attribute to local projects a transformative capacity they have not yet demonstrated.
“At present, I do not see any possibility of overturning the dominant logic. These are individual, distinct projects, each of which fills gaps that general-purpose models do not address,” she says.
Attention is also distributed unevenly among contexts that are broadly described as underrepresented. According to Biffi, parts of Africa and India are increasingly able to attract investment, startup activity and media coverage, while Afghanistan’s technology ecosystem remains considerably less visible.
A single testimony is not sufficient to conclude that a stable, measurable hierarchy exists between these ecosystems. The observation does, however, point to a question worth examining: policies and investments intended to promote technological inclusion also create their own geographies of attention.
Who builds the data
To understand why the linguistic dimension cannot be separated from the cultural one, the perspective of Rebecca Ryakitimbo is particularly useful. Ryakitimbo is a Tanzanian researcher and technologist working at the intersection of technology, gender and communities.
She has been a Google Policy Fellow and an Internet Society Fellow and has also held a Mozilla fellowship. In 2024, she was included among the 100 Brilliant Women in AI Ethics.
Her work addresses, among other issues, African-language datasets and the barriers limiting women’s participation in building the digital ecosystem.
“Most widely used LLMs are built with intrinsic biases; they are trained primarily on Western data, losing the context of non-Western communities, and this becomes even more evident for marginalised groups such as women,” she argues.
The status of this statement matters. It is Ryakitimbo’s assessment, informed by her research, rather than a uniform description of every dataset used by every existing model.
The underlying infrastructural problem, however, is well documented. The quantity of digital resources available differs enormously across languages. UNESCO reports that fewer than 2% of the world’s languages have a substantial digital presence.
And digital presence is not simply a quantitative matter.
“Women in local communities are excluded because language itself is not homogeneous: it contains nuances associated with linguistic variation, accents, context and cultural meanings. Elements such as code-switching are often not captured by LLMs designed within the context of the Global North,” Ryakitimbo explains.
A language is therefore not simply an inventory of words that can be translated. Local varieties, registers, shifts between languages, implicit references and social conventions all contribute to meaning.
When these elements are poorly represented in datasets or inadequately modelled, the loss is not limited to grammatical accuracy. It may also affect the system’s ability to recognise intentions, social relationships, cultural categories and forms of local knowledge.
A larger model is not always the answer
Ryakitimbo also introduces a technological issue that often remains secondary in discussions about inclusion: model scale.
Developing and, especially, training large language models requires computational infrastructure and capital that only a relatively small number of organisations can access. Using these models through cloud services does not, of course, require users to own that hardware directly, but it introduces other forms of dependency: connectivity, API costs, service availability, data control and reliance on external infrastructure.
For some use cases, Ryakitimbo argues, Small Language Models may provide an interesting alternative.
Not because smaller models are inherently more inclusive. A smaller system can reproduce exactly the same representational problems as a larger one. But lower computational requirements can make local or edge deployment easier, allow greater specialisation in specific domains, provide more control over data and make deployment more feasible in environments with limited infrastructure.
The question therefore shifts. It is not which model is “better” in absolute terms, but which architecture is appropriate to the purpose, the resources available and the degree of autonomy required.
At the governance level, Ryakitimbo distributes responsibility across different actors: “Regulators have a responsibility to ensure compliance with standards that protect end users; companies must take responsibility for their own tools, ensuring that safety and inclusion are fundamental standards rather than optional features.”
An answer does not emerge from nowhere
AfroféminasGPT, Ask Hakima and projects working to develop datasets for underrepresented languages do not demonstrate that general-purpose models are inherently incapable of operating in contexts beyond those most extensively represented in their data.
They show something more limited, but still significant: an answer generated by an AI system does not emerge from a neutral space.
A large language model generates sequences through probability distributions learned during training and subsequently shaped by post-training, instructions, context, retrieval systems where present, and generation parameters. It does not simply select “the most probable answer” from a pre-existing set of responses.
What it can produce nevertheless remains conditioned by the knowledge architecture within which it operates.
What data is available? Which languages are sufficiently represented? Which texts can be retrieved? What instructions guide the system? Which categories are used to interpret the question?
Linguistic fluency tends to make these intermediate processes less visible. A system can formulate an entirely plausible response even when its understanding of the specific context in which the question was asked is limited.
This may be the most interesting issue raised by the projects examined here: they make the process of selection more visible.
How many perspectives can a model accommodate?
Biffi frames the question through language.
Around 7,000 languages are currently in use worldwide, and at least 40% are considered at risk.
The question, then, is not whether we can imagine a single model that simply “contains” 7,000 languages. It is what it would mean to develop systems capable of operating meaningfully across a much wider plurality of linguistic and cultural structures.
“How would the answers we expect to see change?” Biffi asks. “Could we move from a model that is a little less Eurocentric, a little less Western, towards one that is somewhat more genuinely global?”
It is a question concerning both the quantity of data and its origin.
Increasing the volume of text available in a particular language does not automatically produce epistemic plurality. If those texts come from the same institutions, the same elites or the same digital environments, the language may be formally represented while the diversity of experiences within it remains far less visible.
Representation is therefore simultaneously a linguistic, cultural, social and economic issue.
And this inevitably leads back to the question posed by Torres Soler: “The question is not only what AI can do, but: who will it serve?”
AfroféminasGPT and Ask Hakima do not provide a definitive answer.
They remain relatively small-scale projects, built with limited resources and, at least to some extent, dependent on the same technological infrastructures they are seeking to interrogate critically. We do not yet know how scalable these architectures are, nor whether they can significantly influence the development of major general-purpose models.
They do, however, contribute something useful to the debate: they show that data, sources, language, tone, context and intended users are not merely technical parameters.
They are design decisions.
Making those decisions explicit allows us to ask a different question from the conventional search for an ever more capable model: which forms of knowledge do we want a system to be able to recognise, and what conditions are required for it to do so without treating one particular perspective as universal?
Cultural plurality in AI is not an outcome that has already been achieved. These projects instead show some of the ways in which it might be built – and, more importantly, which questions become visible when we attempt to do so.
