Advertisement

Kenya’s SHA AI algorithm overcharges the poorest while letting the wealthy off the hook, investigation reveals

Kenya's SHA AI algorithm overcharges the poorest while letting the wealthy off the hook, investigation reveals

TL;DR

  • SHA’s AI means-testing tool was trained on outdated 2021 household survey data that over-represented middle-income households.
  • Data consultancy IDinsight warned the government before launch that the system was inequitable and should not be used without new training data.
  • Among the bottom 40% of households, the algorithm misclassifies 39% as non-poor, effectively denying them subsidised premiums.
  • Of 30 million registered Kenyans, only 5 million are actively paying, while hospitals accumulate billions in unpaid SHA reimbursements.

Every week, a community health volunteer named Grace visits homes in Embakasi, one of Nairobi’s densest informal settlements, and asks residents a series of questions that feel equal parts census and interrogation. What type of toilet do you use? What is your roof made of? Do you own a radio? She helps them enter the answers into a digital questionnaire on their phones.

The data goes to an algorithm that decides what they will pay for healthcare.

Grace is one of roughly 100,000 community health promoters embedded across Kenya to support the government’s rollout of the Social Health Authority, or SHA, President William Ruto’s flagship replacement for the National Hospital Insurance Fund, launched in October 2024.

Ruto promised that SHA would deliver universal health coverage to a country where 83 percent of workers have no formal employment, no payslip, and no verifiable income record.

A ten-month investigation published this week by Africa Uncensored, Lighthouse Reports, and The Guardian has found that SHA’s central mechanism for determining premiums, a machine-learning algorithm called the Means Testing Instrument, was structurally designed to overcharge the poorest Kenyans while undercharging the wealthiest.

The government received that warning from its own consultants before the system went live and deployed it anyway.

How the algorithm works

Because most Kenyans cannot produce proof of income, SHA does not ask what people earn. Instead, the system applies a method known as Proxy Means Testing, using 43 observable household indicators to predict income.

The indicators range from the gender of the household head to the type of roof (tiles or iron sheets). A machine-learning model then calculates an income estimate, and a percentage of that estimate becomes the annual premium.

The model was trained on data from the 2021 Kenya Continuous Household Survey. The government published the mathematical formula in the Kenya Gazette in March 2024, but did not disclose the specific variables or the coefficients assigned to each.

For nearly a year, neither SHA registrants nor independent researchers could properly assess how the system was reaching its conclusions.

Africa Uncensored submitted access-to-information requests to SHA under the Access to Information Act. When the authority refused to acknowledge those requests, reporters lodged a complaint with the Ombudsman, who ordered SHA to comply.

Kenya's SHA AI algorithm overcharges the poorest while letting the wealthy off the hook, investigation reveals
Image source: Lighthouse Reports and Africa Uncensored

In November 2025, SHA disclosed the model formula and variables, though it withheld the coefficients for each variable. Journalists used that disclosure, matched against the 2021 survey questionnaire, to reconstruct and audit the model.

“It could as well have been written in ancient Egyptian. That formula had a huge impact on real people’s lives.”

— Brian Lishenga, chair, Kenya Rural and Urban Private Hospitals Association

The most damaging element of the investigation is the documentary. Reporters obtained a pre-deployment assessment produced by IDinsight, an international data consultancy, that was shared with the government before SHA went live.

The report stated that the Means Testing Instrument was inequitable, particularly for low-income households, because the 2021 survey used as training data over-represented middle-income households and under-represented the poorest.

The report also noted the data was outdated, failing to account for the multiple economic shocks Kenyans had suffered in the intervening years.

IDinsight warned that, without new training data, the system was highly likely to miscalculate premiums for low-income households, with the poorest bearing the worst of the errors. Efforts to recalibrate the model produced only a marginal improvement but the government launched it regardless.

Also Read: Researchers demonstrate $10,000 Apple Pay theft from a locked iPhone

David Khaoya, a health economist who advised Kenya’s health ministry, told investigators that the system’s design forced a deliberate trade-off: the algorithm could either accurately assess poor households or accurately assess wealthy ones, but not both with equal precision.

The government, he said, chose to prioritise accuracy at the top end. His reasoning: if a wealthy household is misclassified as poor and charged less, that household will stay quiet. The political cost of undercharging the rich was lower than the administrative cost of fixing the model.

The error rate

Lighthouse Reports reconstructed the SHA model and ran it against the 2021 survey data to measure its error rates across income groups. Among the bottom 40 percent of Kenyan households, roughly the official poverty line as determined by the Kenya National Bureau of Statistics, the model exhibits an exclusion error rate of 39 percent.

That means nearly four in ten households at or below the poverty line are classified by the algorithm as non-poor and assigned premiums they cannot afford.

For two farmers tested during the investigation, the algorithm predicted their income at twice their actual figure, triggered by the fact that they own their home and have an electricity connection.

Chris Kidd, a social protection researcher, described the inherent problem: poverty is a fluid category, and using a fixed set of observable assets to estimate it is intrinsically imprecise.

An iron-sheet roof or a pit latrine tells the model something, but not enough.

“It feels like a lottery,” Kidd said. “And the lottery is not a great way of building trust.”

On the ground

Florence Atieno lives in Huruma, a low-income neighbourhood on the northeastern edge of Nairobi. The algorithm assigned her a SHA premium of Ksh. 6,600 per year. She was pregnant when the system set that figure. She could not pay it.

When she tried to use the official appeals mechanism, the system rejected her application without providing a reason, because her premium did not meet the minimum threshold required for reconsideration.

Another SHA registrant, a former mama fua, a self-employed laundry worker, was left unable to work after a fall that broke her leg and required metal pins.

The algorithm assigned her a premium she could not understand. Friends helped with an initial payment. Without employment, she could not sustain it. Her ongoing treatment and medication fell outside what SHA covered for her.

“The priority is for my children to eat,” she said, “not to pay for SHA.”

Patrick Maina, a former boda boda rider whose spine was broken in a 2021 accident, contributes Ksh. 500 a month to SHA. When he falls ill, he receives no assistance. Grace, the community health volunteer, has watched families eating one meal a day receive SHA premium demands that represent between 10 and 20 percent of their total income. She has accompanied critically ill people to hospital gates and watched them turned away.

“Many people have been unable to go to hospital,” she said. “Will they pay SHA, or pay for food, or pay for the small house they live in?”

A system under strain

Of the more than 30 million Kenyans now registered with SHA, only 5 million are actively paying their premiums, according to the investigation. Hospitals are reporting large financial deficits as SHA reimbursements remain delayed or unpaid.

Brian Lishenga, chair of the Kenya Rural and Urban Private Hospitals Association and one of SHA’s most prominent critics, described the system in unambiguous terms.

“This is an experiment that has failed,” Lishenga said. “It is a really poor tool for identifying poor households. It is a great tool for helping the government run away from responsibility.”

Former Deputy President Rigathi Gachagua made a prediction in March 2026 that SHA would collapse within six months.

Kenya’s health budget for 2025/2026 sits at approximately Ksh. 132 billion, representing less than 4 percent of total government spending, at a time when close to 67 percent of government revenue is committed to debt repayment.

Health economist analysts cited by the investigation argue that universal health coverage cannot succeed under those fiscal constraints regardless of algorithmic design.

The funding shortfall has pushed ordinary Kenyans to improvise. Mchanga, Kenya’s largest medical crowdfunding platform, has recorded more than 134,000 fundraisers since its founding, with 40 percent of them raised for medical expenses, the same harambee culture SHA was supposed to render obsolete.

The investigation draws a firm distinction between a technical failure and a policy choice. An algorithm that malfunctions despite good intentions is one category of problem. A system deployed after its creators documented the exact ways it would harm the people it was built to serve is another.

The government knew the Means Testing Instrument was inequitable before it went live. It knew the training data was outdated. It knew the poorest households would be the most likely to be wrongly assessed. It launched the system, assigned premiums to millions of people and took months to acknowledge any of the concerns publicly.

For Kenyans currently locked out of healthcare by a mathematical formula, the distinction between error and design is more than academic. Florence in Huruma is still waiting for her appeal to be reviewed. Grace in Embakasi will visit more households this week. The algorithm will keep generating numbers.

 

This article draws on the investigation Hiding Behind AI / Error by Design, published by Africa Uncensored, Lighthouse Reports and The Guardian. Reporters include Gabriel Geiger, Purity Mukami, Joy Kirigia, Tomas Statius and Naipanoi Lepapa.

Muchiri Avatar

Leave a Reply

Your email address will not be published. Required fields are marked *