The best language models in the world do not work very well for most of the world. AfroBench, a benchmark covering 64 African languages across 15 tasks and 22 datasets, found large and persistent gaps between English and nearly every African language it tested — including for the leading proprietary systems. The most advanced AI available is least useful in the places where the largest number of future users live.
That is usually told as a story about who is behind. Danil Kerimi, a senior advisor to the UN Tech Envoy, tells it the other way around.
Developing countries are not only affected by AI, he argues. “Very often, we ask ourselves how AI will affect our country, our society, our industry or our careers. But we also need to understand that our choices will shape AI systems moving forward,” Kerimi said. The countries treated as recipients of AI governance hold inputs that frontier AI cannot generate for itself. The open question is not whether they have something to contribute. It is whether the window to contribute it is actually open — and, increasingly, whether the venue China is building for them is an opening or a new dependency.
The rules are not written yet
The sheer volume of AI governance frameworks now in circulation is usually read as a failure of coordination. Kerimi reads it as a signal that the field is unsettled — and an unsettled field is one that can still be influenced. On his account the countries that recognise their own leverage earliest are the ones that get to use it, which means the first task is diagnostic rather than declaratory: AI is a full stack, from minerals and energy through semiconductors, intellectual property, skills and business models, and a country’s realistic move is to find the layer where it can actually exert force rather than attempting a position on all of them.
John Higgins, chair of the International AI Governance Association, is blunter about what the destination looks like. There will be no unified global AI governance regime, he argues; the achievable goal is interoperability between approaches built on common standards. In his reading the principles stage is largely over — the major frameworks have converged on a broadly similar set of commitments — and the live work has moved to converting those principles into practical, executable deliverables through standards bodies.
That shift matters more than it sounds. Summit declarations are decided by geopolitical weight. Standards work is decided in technical committees, over years, by whoever shows up with a worked proposal and the stamina to defend it. It is the part of the governance system where a small delegation with a specific, defensible contribution can count for more than its GDP — which is precisely why the question of what a country brings to the table stops being rhetorical.
What they bring
In a recent interview with Chinese state media China Economic Net (CEN), Jason Slater, director of digital transformation and AI at UNIDO, said developing countries have “a huge amount of untapped talent and data”, and that they can make contributions in use cases, regulatory sandboxes and policy experience.
Data. The gap AfroBench measures is a gap in inputs, and the inputs sit with the people the models serve worst. Masakhane, the grassroots pan-African NLP community, describes itself as having over 1,000 participants across 30 African countries; by 2023, when it was recognised by Ars Electronica, it had published research covering more than 38 African languages and built translation and named-entity corpora that did not previously exist in any form. AfroBench itself came out of the overlapping McGill-NLP group, led by David Adelani with co-authors including Jessica Ojo and Odunayo Ogundepo — researchers who move between the two orbits. Ghana NLP, an open-source initiative working on Ghanaian languages, built Khaya — the first machine-learning translation app for Twi, Ewe and Ga, since extended to Dagbani, Frafra and languages beyond Ghana — after noting that Google Translate covered no Ghanaian language at all. In India, AI4Bharat at IIT Madras has been assembling supervised datasets for Indian languages on a similar logic.
Scenarios. Deployment conditions in the Global South do not merely test AI; they define what a useful model looks like. InkubaLM, built by Lelapa AI’s fundamental research team for isiZulu, Yoruba, Hausa, Swahili and isiXhosa, runs at 400 million parameters and was trained from scratch on 2.4 billion tokens — roughly three orders of magnitude below the frontier systems whose scaling curves dominate the discourse, and trained on a fraction of the data. A model designed around intermittent power, thin connectivity and no local GPU cluster is not a scaled-down version of a frontier model. It is a different design brief, and the brief is written by the deployment environment. Notably, it ships under a non-commercial licence — a reminder that even the clearest case of locally specified AI has not resolved the question of who is allowed to build a business on it.
Slater points to countries deliberately positioning themselves as first movers in specific applied domains — rice production, electrification — rather than competing across the board, and to regulatory sandboxes as the mechanism that makes that positioning legible to investors and standards bodies alike.
Talent. The first Deep Learning Indaba ran at the University of the Witwatersrand in September 2017, on the stated premise that Africans should be active shapers and owners of advances in AI rather than observers of them. The organisers had planned a master class for 40 students. They received more than 700 applications, moved into the largest lecture hall they could find, and ran the week with close to 300 participants from 22 African countries and 33 African research institutes. The constraint was never the supply of people. Its 2026 gathering is scheduled for Nigeria.
The clearest evidence that the premise holds is individual rather than institutional, and it runs straight into China’s programmes. At a WAIC panel this month, Lin Yonghua, vice president of the Beijing Academy of Artificial Intelligence (BAAI), cited an unnamed teacher from the first cohort of the BAAI–AU-ASRIC faculty development programme who had used the academy’s FlagOS stack for systems optimisation. That researcher is Kedir Yassin Hussen, who announced on LinkedIn — and was congratulated by name by Lin — that he had taken best paper at the GlobalSouthML workshop co-located with ICML 2026 in Seoul. His subject was deploying AI efficiently on legacy data centre hardware in the Global South: a research problem that only exists if you are working inside the constraint rather than reading about it. Neither Hussen nor Hiba Asri, an associate professor at Cadi Ayyad University in Marrakech and another first-cohort participant, responded to requests for comment.
One best paper is one data point, and it should not be asked to carry more than that. But it is the specific data point that the capacity-building case needs, because it runs the whole way through: a Chinese training programme, a researcher who was already good, a problem defined by African infrastructure, and a result that landed at a top-tier venue rather than in a programme report.
China’s table
If the Global South does hold leverage, then the test of any external AI offer is whether it runs in both directions. China has built the most conspicuous venue for that question.
The World AI Cooperation Organization’s founding agreement was signed in Shanghai on 17 July, with Kazakh President Tokayev, Cambodian Prime Minister Hun Manet, Thai Prime Minister Anutin and UN Secretary-General Guterres all speaking at the opening. Alongside it, China announced three concrete measures: 5,000 AI training places for developing countries over five years; international AI application cooperation centres oriented toward ASEAN, the Arab League, the African Union, CELAC, the SCO and BRICS; and the deployment of the Mazu intelligent weather early-warning system to 30 countries.
The precision is worth preserving. The commitment is 5,000 training places, not 5,000 scientists — a capacity-building pipeline, denominated in seats. Both Kerimi and Higgins are broadly positive about China’s convening role, and Higgins in particular credits it as a substantial contribution. His assessment is not that of a bystander: the association he chairs co-hosts an annual international AI governance forum jointly with Wei Kai of the Chinese industry ministry’s AI standards committee, held at WAIC, and Higgins co-leads the standards programme of the World Internet Conference’s AI committee. That makes him unusually well placed to judge China’s convening record, and an interested party in it. The demand is also real — WAICO was framed explicitly as a response to Global South calls for a venue. Convening power is not nothing: it determines who is in the room.
But the shape of the offer describes capability moving downward — seats, centres, a deployed system — rather than co-authorship moving upward. Whether the two converge is not yet answerable, and this piece will not pretend otherwise. What can be said is that the offer will be measured against a standard China itself set by invoking the Global South’s own demand for agency.
The gap between having leverage and using it
Higgins is most useful where he is least promotional. Asked what actually stops developing countries from participating, he does not start with governance at all. “When I talk to people in developing countries, high among their concerns is often just the basic infrastructure — digital, telecoms, data centre infrastructure. The infrastructures themselves are a barrier. Within their country they haven’t yet adopted regulatory frameworks — so society and the suppliers don’t feel they’re in a safe and secure place yet,” Higgins said.
That is the honest floor under all of this. Having leverage and exercising it are different things, and the second requires money, institutions and time that the first does not.
The sceptical case goes further, and belongs in the piece. Data annotation in the Global South is frequently low-paid, precarious work; the researchers who prove the talent thesis are also the researchers most likely to be recruited abroad; and the value created by better multilingual data has so far accrued mainly to whoever owns the model, not whoever supplied the language. A contribution narrative that skips these is a public relations narrative.
There is also a quieter structural problem, visible in the citations above. Masakhane, the Deep Learning Indaba and Lelapa AI are not three independent proof points. Vukosi Marivate — professor of computer science at the University of Pretoria and director of its African Institute for Data Science and AI — co-founded all three. Jessica Ojo appears as an author on both AfroBench and InkubaLM. The same few dozen people appear in every account of African AI capability, including this one. That is a testament to what a small, determined community can build; it is also a thin bench on which to rest a claim about a continent’s structural position.
The window Kerimi describes is real, and so is the leverage. But leverage that goes unexercised is indistinguishable, in the standards committee and in the training seat, from not having any.
