An AI system used to predict how much Kenyans can afford to pay for healthcare has systematically driven up costs for the poor, an investigation has found.
The healthcare scheme, rolled out nationally as a key electoral promise of President William Ruto, launched in October 2024 to replace Kenya’s long-standing national insurance. Billed as “accelerating digital transformation”, it aimed to extend cover to the country’s large informal economy — day labourers, hawkers, farmers and other non-salaried workers who make up about 83% of the workforce.
Ruto campaigned in 2023 promising that “no Kenyan will be left behind.” Instead, the system — described by the government as AI-powered but actually based on a predictive machine-learning algorithm using proxy means testing (PMT) — has sparked protests and anger. Premiums for millions are now set by an opaque formula sources have described as “flawed,” and the investigative audit by Africa Uncensored, Lighthouse Reports and the Guardian found the algorithm routinely overestimates the incomes of the poorest while underestimating those of wealthier households.
Volunteers such as Grace Amani* visit homes and enter dozens of data points into a digital questionnaire on people’s phones: type of toilet, roof material, ownership of radios and other assets. The algorithm then returns a figure for the household’s annual public health insurance contribution. Amani says many of the poorest families she registers are charged amounts they cannot afford — sometimes 10%–20% of their meagre incomes — and some critically ill people are denied treatment because they cannot meet the demanded payments. “People are dying, people are suffering,” she said.
The new Social Health Authority (SHA) has been flooded with complaints about misclassification, incomprehensible premiums and unaffordable charges. Kenyans without private insurance who do not pay SHA premiums risk being turned away from facilities or hit with large hospital bills. Social media is full of accounts of people who were previously paying modest sums now being billed far higher amounts; one single mother reported her monthly contribution was set at 3,500 Kenyan shillings.
David Khaoya, a health economist who advised Kenya’s health ministry, said the government faced a trade-off because of the system’s limitations: it could either aim to identify poor households correctly or wealthy ones correctly. He said the choice was made to prioritise accurately evaluating the wealthy, even if that produced systematic overcharging of poorer households. “If you identify a richer person as poor and therefore ask him to pay less, this person will never own up and say, ‘I’m actually supposed to be paying more,’” he said.
PMT estimates household income from observable characteristics and assets rather than direct earnings and has been used in World Bank-funded programmes across Africa, Asia and the Pacific. In Kenya, it has meant deploying volunteers to record roofing materials, livestock and other indicators and feeding those details into an opaque model that attempts to predict income and set premiums. But the audit tested the SHA model against thousands of real households and found repeated overestimation of the poorest households’ means; for example, two farmers were predicted to earn twice their actual income because they had electricity and owned their house.
Researchers have long warned PMT is imprecise because poverty is fluid and asset proxies do not reliably reflect current earnings. Stephen Kidd, a development economist, noted that similar poverty-targeted schemes have very high error rates: one tested in Indonesia excluded 82% of the intended population; another in Rwanda had an error of 90%. In Kenya’s case, the audit suggests the SHA system overcharges more than half of poor households while underestimating incomes of higher-income households.
A separate report by international data consultancy IDinsight, shared with the government before implementation and obtained by reporters, warned the SHA system was flawed and “inequitable, particularly for low-income households.” It said the model “over-represents middle-income households and has very few data points from poverty pockets” and was “out-of-date with the current socioeconomic condition” in Kenya after multiple economic shocks. Despite these warnings, the SHA was deployed.
The rollout has struggled: of more than 20 million people registered for SHA, only about 5 million regularly pay premiums. Hospitals report large deficits as promised reimbursements from SHA remain unpaid. In March, former deputy president Rigathi Gachagua predicted the scheme would collapse within months. Critics such as Dr Brian Lishenga, head of Kenya’s Rural and Urban Private Hospitals Association, call the system a failed experiment and say it is a poor tool for identifying poor households while enabling the government to shirk responsibility.
PMT-based systems are spreading globally, often promoted by international donors, as a way to extend state services to previously uncounted informal workers whose inconsistent earnings make traditional income-based schemes difficult. But auditors and researchers argue the models are inherently imprecise, opaque and damaging to public trust. “It feels like a lottery,” Kidd said. “The lottery is not a great way of building trust.”
The SHA’s formula has generated widespread frustration and real human costs: families forced to choose between paying for SHA or food, others unable to access lifesaving care because they were classified as able to pay when they could not. The algorithm’s lack of transparency and the decision to prioritise accuracy for wealthier households over protecting the poorest have turned a flagship promise into a source of harm and uncertainty.
*Name changed to protect identity
Read a fuller report of the methodology used by the reporters at Africa Uncensored here
