How much of Africa's economy is informal?

The ILO estimates that informal employment outside agriculture accounts for 66 percent of total employment in sub-Saharan Africa and 52 percent in North Africa. Vulnerable employment, covering own-account and unpaid family workers, sits at 76.6 percent against a global figure of 45.3 percent. Include agriculture and the informal share rises further. At borders, the UN Economic Commission for Africa puts informal cross-border trade at 7 to 16 percent of recorded intra-African trade and 30 to 72 percent of recorded trade between neighbouring countries.

Why do large language models perform badly on African languages?

Because the training text does not exist in usable quantity. The African Languages Lab collaboration reported in October 2025 that African languages make up close to a third of the world's living languages and that 88 percent of them are severely underrepresented or ignored entirely in computational linguistics. Speech data shows the same shape: in Mozilla Common Voice 23.0, Kinyarwanda holds roughly 2,383 hours while Swahili, spoken across a far larger region, holds roughly 719. Coverage tracks who funded a collection programme, not how many people speak a language.

Who owns data collected in Africa?

Usually whoever paid for the collection, which is the core governance problem. The African Union's Data Policy Framework, adopted in 2022, names sovereignty as a principle and says member states should self-manage and govern their own data, but enforcement is thin. Kenya's High Court ordered Worldcoin to delete iris data from over 300,000 Kenyans in May 2025, and the regulator only confirmed deletion in January 2026. Community-held corpora, like those built by the Masakhane network, are a different model, though the funding behind them is still largely non-African.

Africa's Data Gap: The Real Constraint on African AI

In July 2025 Nigeria's economy became roughly a third larger in an afternoon. Nothing was built.

The National Bureau of Statistics moved the base year for GDP from 2010 to 2019. Nominal output for 2024 went from about 277 trillion naira to about 373 trillion, a jump of 34 percent. Real estate climbed to the third largest activity in the country. Crude petroleum and natural gas, the thing Nigeria is known for abroad, dropped to fifth at under 6 percent.

None of that activity was new. Informal rentals, digital services, modular refineries, quarrying: all of it had been running for years, invisible to a 2010 basket with no line for it. The statisticians did not find money. They started counting.

That is the argument of this piece, compressed into one government press release. The binding constraint on African AI is not a shortage of people who can build. It is that the reality those systems would have to describe was never written down.

Sixty-six percent of the work leaves no trace

The ILO puts informal employment outside agriculture at 66 percent of all employment in sub-Saharan Africa and 52 percent in North Africa. Vulnerable employment, meaning own-account and unpaid family work, runs at 76.6 percent against a world figure of 45.3. For women the non-agricultural informal share is 74 percent, against 61 for men. Put agriculture back in and the informal share climbs again.

At the borders it gets worse. Research by the UN Economic Commission for Africa puts informal cross-border trade at between 7 and 16 percent of all recorded intra-African trade, and between 30 and 72 percent of recorded trade between neighbouring countries. Sit with the top of that range for a second. On some borders the goods nobody writes down approach the volume of the goods somebody does.

These are not small economies. They are large economies with no telemetry.

Capture follows money, not speakers.

What a trader's ledger looks like in Huaqiangbei

I run the Nigerians in Shenzhen Association, so I meet the same business shape most weeks. A man buys four hundred phone screens in Huaqiangbei, pays cash or sends a WeChat transfer to a personal account, takes a carbon-copy receipt with a shop stamp on it, consolidates in a warehouse out in Bao'an and air-freights to Lagos, where the screens end up in Computer Village. Real revenue. Years of it. When I am in Lagos I can buy that same screen off a table in Ikeja and the seller has no idea which shipment it came from, only who brought it.

I asked one member what he does with the notebook he keeps it all in. He throws it out in December and buys a new one.

Every reason capture fails is sitting in that answer. Cash settles instantly and leaves nothing behind. The negotiation happened in Igbo and shop-floor Mandarin, so even a recording would not be searchable by anything trained on English text. Nobody funds a surveyor to follow that trade, because the cost per record is high and the buyer for the record does not exist yet. And the trader has a positive reason to stay illegible: writing it down is a tax event and a disclosure to his competitors. Silence is the rational move.

The same hole sits under prices and under movement. Ask what a bag of onions cost in Dawanau market in Kano last Tuesday and everyone standing in it can tell you, while no queryable series exists anywhere. Ask how Lagos actually moves and the honest answer is that the minibus network was never planned, it accreted, and the most complete map of it lives in the heads of the drivers. Neither can be forecast. Not because forecasting is hard, but because there is no series to fit.

Half a million yield observations for a continent that mostly farms

In 2025 a team published GROW-Africa in Scientific Data, an Africa-wide crop production database. It holds 535,844 georeferenced yield observations covering 25 crops, including maize, sorghum, cassava, groundnuts and cowpeas. About 36 percent of it comes from World Bank household surveys, about 35 percent from subnational government statistics and about 27 percent from other local sources.

That is a serious piece of work. It is also, more or less, the corpus.

A useful yield model wants the plot boundary, the planting date, the variety, the input applied, the rainfall received and the weight harvested. The overwhelming majority of African plots have never generated one complete row like that. Not once. So an insurer pricing drought cover cannot underwrite the farm. It underwrites a satellite rainfall proxy instead, which pays out in seasons the farmer was fine and stays quiet in the season the crop actually failed. The farmer concludes that insurance is a scam. The insurer concludes that African smallholder agriculture is uninsurable. Both conclusions come out of the same missing table.

Kinyarwanda has more recorded speech than Swahili

Mozilla's Common Voice 23.0 landed on 30 September 2025 with 35,921 hours of speech across 286 languages, after a single release added 149 new ones. Inside that total, Kinyarwanda holds roughly 2,383 hours. Swahili holds roughly 719.

Kinyarwanda is spoken in one country. Swahili is the working language of a large part of East and Central Africa. The gap has nothing to do with how many people speak either one. It exists because donors funded a Kinyarwanda collection programme and the collection happened. Capture follows money, not speakers.

The wider picture was measured last October by the African Languages Lab collaboration. African languages are close to a third of the world's living languages, and 88 percent of them are severely underrepresented or ignored outright in computational linguistics. The research gap they found runs from twenty papers on global languages for every one on African languages in Google Scholar, up to seventy to one in IEEE. Independent benchmarking across dozens of African languages keeps returning the same verdict: even the strongest proprietary models still open wide gaps against their own English performance.

The work closing this is unglamorous and largely volunteer. Masakhane, a distributed network of African researchers, has shipped named-entity, part-of-speech and news-topic datasets across African languages plus a benchmark for testing large models on them. The South African team at Lelapa AI released InkubaLM, a small language model trained on African-language text and sized to run without a data centre standing behind it. Neither is a frontier model. Both are corpus work, which is the part nobody wants to fund and the only part that changes the outcome.

The sensor already exists and it is designed in my city

The phones most people in Lagos, Nairobi and Accra actually hold are Tecno, Infinix and itel. All three are Transsion brands. Transsion is headquartered in Shenzhen, and its music platform Boomplay ships preinstalled on the handsets.

I live twenty minutes from where those devices get specified. I read the Chinese-language coverage of the African handset market months before an English version appears, when one appears at all. What a phone measures, which input methods and keyboards it ships with by default, what telemetry goes home: those calls are made here, by people who are very good at making them, on behalf of a user twelve thousand kilometres away.

That is not an accusation. Transsion built for African conditions when nobody else could be bothered, and the market rewarded it correctly. It is a description of who currently owns the sensor.

For what a captured record is actually worth, look at mobile money. The GSMA counted about 1.4 trillion dollars of mobile money transaction value in sub-Saharan Africa across 2025, two thirds of the global total, moving through roughly 347 million accounts active in any given month. East Africa alone did 806 billion dollars, West Africa 498 billion. Every one of those movements is a row in a database, and that is the whole reason a trader with no payslip and no bank statement can be lent against at all. The transaction history is the collateral. The mechanism is proven. It works precisely where the record exists and nowhere else.

The strongest objection is that capture has usually meant extraction

Here is the case against everything above, and it is not a weak one.

Worldcoin collected iris scans from more than 300,000 Kenyans in exchange for crypto tokens. On 5 May 2025 Justice Roselyne Aburili of the High Court in Nairobi ruled the collection unlawful and ordered permanent deletion of the biometric data within seven days, finding the company had never carried out the data protection impact assessment Kenya's Data Protection Act requires. The regulator confirmed the deletion in January 2026. Two and a half years between the scanning and the erasure, and it took litigation to get there.

The pattern repeats in the labour that makes datasets usable at all. Content moderators and data annotators in Nairobi have spent years in Kenyan courts over dismissals and working conditions on contracts served to a platform headquartered in California. The value travelled one way. The trauma stayed.

So when somebody says Africa needs to capture more of its own data, the honest reply is that this exact sentence has preceded most of the bad outcomes on the continent's data record.

The answer is still not less capture. A place that goes unmeasured does not thereby go unexploited. It goes unbanked, uninsured and illegible to every system that allocates capital, while staying perfectly visible to anyone with a satellite and a scraper. The question worth arguing about is custody. Who holds the corpus, who can revoke access to it, who gets paid when a model trained on it earns.

The African Union's Data Policy Framework, adopted in 2022, names sovereignty as a founding principle: member states should self-manage and govern their own data. Four years on it has better prose than it has enforcement. Kenya's data commissioner moved after a judge did, not before.

And the tension inside the good examples deserves saying out loud. When Masakhane's African Languages Hub opened its call for community-built speech and multimodal datasets in January 2026, the backers were Google.org, the FCDO, IDRC and the Gates Foundation. The corpora are community-held. The money is not. Anyone who tells you that dependency is already resolved is selling something.

The next rebasing is about five years out

Nigeria's statisticians did not discover ninety-five trillion naira. They wrote down, once, what had already been happening, and the country's ranking of itself rearranged underneath them: oil fifth, real estate third, trade second only to crop production. Statistical convention says do it again in about five years.

Everything that happens between now and then is being recorded by somebody or by nobody. In Huaqiangbei the notebooks still go in the bin every December. In Shenzhen, this quarter, engineers are specifying the handsets that Africa's next hundred million users will hold, and deciding what those handsets will notice.

Sources

ILO, Five facts about the informal economy in Africa : https://www.ilo.org/resource/news/five-facts-about-informal-economy-africa

UNECA, Measuring informal cross-border trade is key to monitoring intra-African trade : https://www.uneca.org/stories/measuring-informal-cross-border-trade-is-key-to-monitoring-intra-african-trade

Africa Check factsheet, Nigeria rebases its economy again : https://africacheck.org/fact-checks/factsheets/factsheet-nigeria-rebases-its-economy-again-heres-what-sets-it-apart

Geyman et al., An Africa-wide agricultural production database (GROW-Africa), Scientific Data 2025 : https://www.nature.com/articles/s41597-025-05257-5

The African Languages Lab: A Collaborative Approach to Advancing Low-Resource African NLP (arXiv 2510.05644) : https://arxiv.org/abs/2510.05644

Mozilla Common Voice 23.0 release, 30 September 2025 : https://community.mozilladatacollective.com/common-voice-23-0-live-on-mozilla-data-collective/

GSMA State of the Industry Report on Mobile Money 2026, via Connecting Africa : https://www.connectingafrica.com/mobile-money/-1-4t-flowed-through-mobile-money-in-sub-saharan-africa-in-2025-gsma

Kenya High Court orders Worldcoin to delete biometric data, Techpoint Africa : https://techpoint.africa/news/worldcoin-delete-biometric-data-kenya/

The notebook goes in the bin every December