Kenya’s AI healthcare pricing system overcharges the poorest

TL;DR:

• A joint investigation by Africa Uncensored, Lighthouse Reports and the Guardian has found Kenya’s AI-driven Social Health Authority (SHA) systematically overcharges poor households while underestimating the incomes of wealthier ones.
• Of more than 20 million people registered with SHA, only 5 million are regularly paying premiums; some hospitals report large deficits as government reimbursements stall.
• A pre-deployment report by data consultancy IDinsight, obtained by reporters, warned the system was “inequitable, particularly for low-income households” — but the government deployed it anyway.

President William Ruto’s flagship Social Health Authority, launched in October 2024 to replace Kenya’s national insurance scheme, was meant to expand affordable healthcare to the 83% of workers in the country’s informal economy. It instead relies on a proxy means-testing (PMT) algorithm that estimates household income from features such as roofing materials, toilet type and ownership of a radio — and then issues a premium. Investigators auditing thousands of real households found incomes routinely overestimated, including two farmers whose predicted income was double their actual earnings simply because they owned their home and had electricity.

Context

PMT is a decades-old World Bank approach now spreading across Africa, Asia and Latin America, often as a loan condition. Development economist Stephen Kidd, who has tested similar schemes in Indonesia and Rwanda, found error rates of 82% and 90% respectively. Health economist David Khaoya, who advised Kenya’s health ministry, told reporters the system’s constraints forced a choice: accurately assess poor households or accurately assess wealthy ones. The government chose the latter, on the reasoning that wealthier households would not volunteer to pay more.

For UK readers, the Kenyan story sits alongside familiar British concerns about algorithmic decision-making in public services. The Department for Work and Pensions has faced repeated scrutiny over machine learning systems used in fraud detection, and the Information Commissioner’s Office has flagged the opacity of automated decisions affecting benefits. The pattern is consistent: opaque models trained on incomplete or outdated data produce outputs that disadvantage those least able to challenge them.

Looking forward

In March, former deputy president Rigathi Gachagua predicted SHA would “collapse in another six months”, and the funding gap is widening as fewer than a quarter of registrants pay regularly. Critics including Dr Brian Lishenga of Kenya’s Rural and Urban Private Hospitals Association argue the system is a “great tool for helping the government run away from responsibility”. The wider lesson — that machine-learning means-testing systems shift accountability from policy decisions to opaque code — applies wherever governments contemplate similar automation, including UK welfare and pricing systems.