This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
The Kenyan Social Health Authority deployed a machine learning-based 'Means Testing Instrument' to determine health insurance premiums for millions of citizens. Investigations revealed that the model systematically overestimates the income of the poorest households, leading to unaffordable premiums and the denial of essential medical care for vulnerable populations. Despite internal warnings from consultants about the model's inequity, the government proceeded with the rollout, resulting in widespread financial burden and restricted healthcare access.
Kenya's Social Health Authority reportedly used a machine-learning means test to set health insurance premiums. Investigators alleged that income overestimation burdened poor households with unaffordable charges. A Nairobi laundry worker was reportedly sent home before surgery and could not afford follow-up care. The technical audit used a reconstruction; the exact deployed model and implementation of proposed changes were not confirmed. SHA defended the model as equitable.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Accuracy and effectiveness of AI decisions and actions are dependent on group membership, where decisions in AI system design and biased training data lead to unequal outcomes, reduced benefits, increased effort, and alienation of users.
Also classified as
The automated system sets binding premiums and denies appeals without human review or explanation, removing human agency over critical healthcare access.
The linear regression model fundamentally lacks the capability to accurately predict income at the extremes of the distribution, leading to massive errors for the poorest Kenyans.
The model exhibits significant performance disparities across demographic groups, with poor female-headed households and households with electricity being disproportionately misclassified as non-poor and overcharged.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The financial and physical harms resulted from the automated premium calculations generated by the deployed machine learning model, which unintentionally overpredicted the incomes of the poorest households.
Added categories peak at 4: Severe (Psychological Harm, Inferred)
The Kenyan Social Health Authority deployed a flawed machine learning model that systematically overpredicted incomes for poor households, locking millions out of affordable healthcare. This administrative AI failure represents a substantial threat to societal stability and human rights, highlighting the risks of deploying unvalidated automated decision-making systems in critical public services.
The 'Means Testing Instrument' is a linear regression model that multiplies 43 static household variables by fixed weights and sums them to output a predicted income. It functions as a deterministic mathematical tool without autonomous planning, goal-directed behavior, or multi-agent coordination.
High Risk: The system is an AI used in healthcare and social security/essential public services ('Means Testing Instrument' to determine access to public healthcare), which has significant implications for safety, fundamental rights, and public interests.
The system's primary function is to predict/estimate the monetary value of household consumption (income) using linear regression, which is then used to allocate healthcare costs.
The reports provide highly detailed, peer-reviewed investigative findings, including the exact variables, training data (2021 KCHS), and mathematical methodology of the model. The journalists successfully reconstructed and audited the model, and obtained internal reports from consultants (IDinsight) confirming the system's flaws, leaving very little ambiguity about the AI's role and performance.
The government might argue the system is a necessary administrative proxy rather than an AI failure, but the mathematical bias is objectively demonstrated.
While the exact final code deployed by Safaricom is proprietary, the successful replication and peer-review of the OLS and logistic models provide sufficient detail for high-confidence classification.
Technology
Regression-to-the-Mean Bias
Linear regression inherently predicts towards the center, failing at extremes.
Flawed Two-Step Model Adjustment
Proposed logistic regression adjustments reduced accuracy and increased errors.
Opaque Algorithm Formula
Complex formula was published without public disclosure of weights or code.
Data Inputs
Out-of-Date Training Data
Model trained on 2020/2021 survey data, missing recent economic shocks.
Unrepresentative Data Sample
Training data over-represented middle-income and lacked poverty pockets.
Inaccurate Proxy Variables
Assets like roof material or radio ownership poorly proxy actual income.
Human Factors
Self-Reporting Errors
Users may intentionally or unintentionally misreport assets on online forms.
Overreliance on Automated Appeals
System rejected appeals automatically without providing reasons to users.
Intrusive Questionnaire Design
Lengthy and intrusive questions caused user confusion and discomfort.
Process and Methods
Rushed Implementation Timeline
System was deployed in months instead of years despite its complexity.
Inadequate Validation and Testing
The model was launched despite known high exclusion error rates of 39%.
Ineffective Appeal Mechanisms
Automated appeal line denied premium reductions without providing reasons.
Management
Prioritizing Revenue Over Fairness
Management chose to overcharge the poor to avoid undercharging the wealthy.
Ignoring Expert Warnings
Ignored warnings from World Bank, ILO, UN, and internal consultants.
Pressure to Increase Yield
Government pressured engineers to adjust model to extract higher contributions.
Regulatory Environment
Lack of Algorithmic Transparency
SHA refused access to model details until forced by the Ombudsman.
Absence of Independent Oversight
No regulatory body audited the means testing tool prior to national rollout.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
