Redistributed

Most of the world's population will never use AI in their first language. That is not an accident. It is the predictable outcome of a system designed to serve paying customers in wealthy markets. And unless something changes deliberately, it will stay that way.

The AI industry is concentrated in three regions: the United States, China, and Western Europe. These places host the major labs, attract the bulk of research funding, and control the compute infrastructure that makes frontier models possible. Together they represent roughly 2 billion people. The other 6 billion — spread across Africa, South Asia, Southeast Asia, Latin America, the Middle East — are largely afterthoughts in AI development. Not because they lack talent or need, but because they lack market leverage.

This is a redistribution problem — perhaps the clearest one in all of AI. The benefits are flowing to those who already have the most. Building AI for everyone else will require intentional investment that market forces alone will not provide. But as we will see, that investment is already happening, driven by communities who are not waiting for permission.

The Language Gap

Language is where the disparity shows up most clearly.

There are more than 7,000 languages spoken worldwide. Most large language models are trained predominantly on English text, with varying degrees of coverage for a handful of other high-resource languages: Mandarin, Spanish, French, German, Japanese. Everything else gets scraps.

The mechanism is straightforward. Language model quality depends on training data volume. English dominates the internet — roughly 60% of web content, according to W3Techs. Languages with large internet footprints produce more digitized text, which produces better models, which attracts more users and developers, which generates more text. A virtuous cycle for English. A vicious cycle for Yoruba, Swahili, Bengali, Tamil, Quechua, and thousands of others.

Even models marketed as "multilingual" perform significantly worse on non-English tasks. A 2023 study by Ahuja et al. found that GPT-4's performance degraded substantially on low-resource languages across multiple benchmarks, with the gap widening as language resource levels decreased. As Emily Bender and colleagues argued in their influential 2021 paper "On the Dangers of Stochastic Parrots," large language models risk encoding the perspectives and linguistic patterns of dominant cultures while marginalizing others.

This matters concretely. A farmer in rural Maharashtra who speaks Marathi gets worse AI assistance than a programmer in San Francisco. A Kenyan entrepreneur who works in Swahili gets a degraded experience compared to a London-based competitor. The technology that promises to democratize access to knowledge is, in practice, widening the gap. This is the access dimension of The Redistribution Stack at its starkest: same technology, radically different value depending on which language you speak.

The People Doing the Work

The good news is that researchers and communities across the Global South are not waiting for Silicon Valley to solve this. They are building their own. These projects embody what Elinor Ostrom demonstrated with natural resources: communities can govern shared resources and build shared infrastructure without waiting for either corporations or governments to do it for them.

Masakhane is perhaps the most prominent example. Founded in 2019, this grassroots community of over 2,000 African researchers is building natural language processing tools for African languages. The name means "we build together" in isiZulu. Masakhane has produced machine translation benchmarks, named entity recognition systems, and sentiment analysis tools for dozens of African languages — work that major labs have little commercial incentive to do. Their approach is community-driven: researchers who speak these languages lead the research, rather than outsiders parachuting in to collect data.

Lelapa AI, based in South Africa, is taking the commercial route — building AI products specifically for African markets. Rather than adapting Silicon Valley models, they are developing systems designed from the ground up for African languages and use cases. The company was founded by researchers from institutions including the University of the Witwatersrand and represents a growing trend of African AI companies building for local needs.

AI4Bharat, an initiative led by researchers at the Indian Institute of Technology Madras, tackles a formidable problem: India has 22 officially recognized languages and hundreds more in daily use. The team has developed IndicNLP, a suite of models and datasets covering Indian languages, along with translation, transliteration, and speech recognition tools. Their work on the IndicTrans models has produced some of the best machine translation systems available for Indian languages — better in many cases than commercial alternatives from much larger companies.

Khipu AI serves as a hub for artificial intelligence research across Latin America, connecting researchers from a region with deep mathematical and scientific talent but limited access to the compute resources that frontier AI development demands. The community focuses on building local capacity rather than depending on Northern institutions.

In Southeast Asia, various university labs and startups are working on AI for languages like Thai, Vietnamese, Indonesian, and Tagalog. These efforts are often fragmented — lacking the coordination and funding of their counterparts in wealthier regions — but they represent a growing recognition that AI must work in local languages to deliver local value.

Two international projects deserve particular mention. Mozilla Common Voice has crowdsourced voice recordings in over 100 languages, building the speech corpora that voice recognition systems need but that do not exist for most languages. It is an infrastructure project in the truest sense — laying pipe that others can build on, much like the rural electrification projects that brought power to underserved communities (see How Technology Gets Distributed). And the BLOOM model, developed through the BigScience initiative with significant participation from Global South researchers, demonstrated that a frontier-scale multilingual model could be built collaboratively, with input from researchers outside the traditional AI power centers. BLOOM was trained on 46 natural languages and 13 programming languages, with deliberate inclusion of low-resource languages that commercial models typically ignore. It is a proof of concept for Public AI at the international scale.

Structural Barriers

These projects are impressive. They are also swimming against a powerful current. Understanding these barriers matters because they are not accidents — they are the predictable result of how AI development is currently structured. As The Automation Debate explores in the labor context, the question is not whether AI can serve everyone, but whether the incentives are aligned to make it happen.

Compute access is the most immediate bottleneck. Training a frontier language model costs tens of millions of dollars in compute alone. Even fine-tuning existing models for a new language requires GPU clusters that most Global South institutions cannot afford. A research lab in Nairobi or Chennai might have brilliant researchers but no access to the hardware they need. As Abeba Birhane has documented, this reproduces colonial extraction patterns: data and talent flow out of Africa, while the infrastructure and returns remain in the Global North.

Talent drain compounds the problem. AI researchers from Africa, India, and Latin America are actively recruited by US and European companies and universities, which can offer salaries many times higher than local institutions. This is rational for individual researchers but devastating for local AI ecosystems. The best minds working on Yoruba NLP or Hindi speech recognition end up optimizing ad targeting in Mountain View.

Funding follows a similar pattern. The vast majority of AI research funding flows to institutions in the US and Europe. Even funding earmarked for "global" or "developing world" AI research often requires partnerships with Northern institutions, reinforcing dependency rather than building independent capacity.

Data scarcity is both cause and consequence. Low-resource languages lack the digitized text, annotated datasets, and speech corpora that machine learning requires. Building these resources is labour-intensive work that does not produce the kind of flashy benchmarks that attract citations or investment. It is infrastructure work — essential but unglamorous.

Physical infrastructure matters too. Internet connectivity, cloud service availability, and electricity reliability vary enormously across the Global South. A researcher in Lagos deals with power outages that a researcher in London never thinks about. These are not AI problems specifically, but they shape who can participate in AI development and who cannot.

The Extraction Problem

Here is the bitter irony. The Global South already contributes enormously to AI — just not in ways that benefit its own communities.

As Mary Gray and Siddharth Suri documented in Ghost Work, the data labeling, content moderation, and annotation work that makes AI systems function is disproportionately performed by workers in Kenya, India, the Philippines, and other lower-wage countries. These workers train the models. They rarely get to use them. Shakir Mohamed, Marie-Therese Png, and William Isaac laid out this dynamic precisely in their 2020 paper "Decolonial AI": AI development follows historical patterns of extraction, where raw materials (data, labour) flow from periphery to centre, and finished goods (models, products) flow back — at a price.

This is redistribution running in the wrong direction. The communities that contribute the most labour to AI development receive the least benefit from it. Ursula Franklin would recognize the pattern: prescriptive technology that reduces Global South workers to following machine instructions, while the holistic benefits — creative use, productivity gains, new capabilities — accrue elsewhere.

What Genuine Inclusion Would Require

Fixing this demands more than corporate social responsibility programmes and diversity statements. It requires structural change across several dimensions. A desirable future is a design problem — and these are the design specifications.

Invest in local research institutions. Not partnerships where a Global South university collects data for a Northern lab, but genuine capacity building — funding, equipment, and autonomy. The model should be closer to how the Green Revolution invested in agricultural research institutions in developing countries (for all its flaws, it built lasting local capacity) than to how tech companies run satellite offices to access cheap talent.

Fund data collection and curation for underserved languages. This is slow, expensive, unglamorous work. It does not produce Nature papers or demo videos. But without it, AI will continue to work well in English and poorly in everything else. Projects like Mozilla Common Voice show the model: sustained investment in linguistic infrastructure.

Expand hardware access. Cloud compute grants, GPU-sharing initiatives, and subsidized access programmes can partially close the compute gap. Some programmes already exist — Google's TRC programme provides free access to TPUs for researchers, and several cloud providers offer academic credits. These need to scale dramatically.

Create competitive opportunities at home. The talent drain will not stop until researchers in Lagos, Chennai, and Sao Paulo can build careers in AI without emigrating. This means local funding, local companies, and local recognition for work on local problems. Lelapa AI and AI4Bharat demonstrate that this is possible, but they remain exceptions rather than the norm.

Build community-driven, not corporate-driven. Masakhane works because it is led by the communities it serves. Outside organisations bringing AI "to" Africa or South Asia often misunderstand local needs, impose inappropriate frameworks, or extract more than they contribute. The most effective model is supporting communities to build their own tools.

Develop international Public AI infrastructure. Just as the international community has invested in public health infrastructure, scientific research facilities, and communications networks, there is a case for public AI infrastructure — shared compute, shared datasets, shared models — that is not controlled by any single company or country. Mariana Mazzucato's argument about the entrepreneurial state applies here: AI depends on decades of public research investment, and the returns should flow back to public benefit, globally. AI Cooperatives offer one model for how this can work at the organizational level.

What You Can Do

This is not only a story about institutions and governments. Individual choices add up.

Support the projects. Masakhane, Mozilla Common Voice, AI4Bharat, and similar initiatives welcome contributions — financial support, volunteer data collection, code contributions, and signal boosting. If you work in AI, consider contributing your skills to open multilingual projects.

Push for multilingual evaluation. If you build, procure, or evaluate AI systems, insist on testing across languages. A model that works brilliantly in English and fails in Hindi is not a good model — it is a model that works for some people and not others. This should be treated as a quality issue, not a nice-to-have.

Advocate for inclusive research funding. Research funders, philanthropies, and governments should direct more resources to AI research institutions in the Global South, with minimal strings attached. The goal is building independent capacity, not creating dependency.

Question who benefits. Every time a new AI capability is announced, ask: who can actually use this? In what languages? At what cost? In what countries? These questions are not peripheral. They determine whether AI becomes a tool for narrowing global inequality or widening it.

The researchers and communities profiled here are proof that AI for the rest of the world is possible. It is not a technical problem. It is a resource problem, a priorities problem, and ultimately a redistribution problem. The question is whether the rest of the AI industry — and the policymakers and funders who shape it — will treat global linguistic and geographic inclusion as essential, or continue to treat it as someone else's job.

The answer will say a great deal about what kind of technological future we are designing. If you want to help shape that answer, see How to Contribute.

Sources

Birhane, Abeba. "Algorithmic Colonization of Africa." SCRIPTed 17, no. 2 (2020). SCRIPTed
Mohamed, Shakir, Marie-Therese Png, and William Isaac. "Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence." Philosophy & Technology 33 (2020). arXiv
Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" FAccT (2021). ACM
Gray, Mary L., and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019. Publisher
Ahuja, Kabir, et al. "MEGA: Multilingual Evaluation of Generative AI." EMNLP (2023). arXiv
Mazzucato, Mariana. Mission Economy: A Moonshot Guide to Changing Capitalism. Penguin, 2021. Publisher

General

Who Builds AI for the Rest of the World