SHORTS: Case studies of AI bias affecting monitoring and evaluation outcomes

Case studies of AI bias in M&E-like contexts reveal how skewed models distort outcomes, such as resource allocation or impact assessments, often mirroring challenges in program evaluations for development projects.pmc.ncbi.nlm.nih+1

Healthcare Diagnostics (NHS AI Systems)

In UK NHS deployments, AI triage and hematology analyzers underperformed for minority ethnic groups and rare conditions due to non-diverse training data, leading to missed abnormalities and delayed interventions—66.7% of clinicians reported misleading recommendations. Manual double-checks were needed, eroding efficiency gains; retraining with balanced datasets improved equity but required ongoing audits to prevent drift.[pmc.ncbi.nlm.nih]​

Pulse Oximeters in COVID-19 Monitoring

During the pandemic, AI-enhanced pulse oximeters overestimated oxygen levels in Black patients by up to 3%, masking hypoxia and delaying care, akin to M&E bias overlooking vulnerable beneficiaries in health program tracking. This amplified disparities, as lower baseline access led to skewed risk predictions; mitigation involved direct physiological recalibration over proxies.kodexolabs+1

Healthcare Cost Prediction Algorithms

An algorithm predicting care needs used costs as a proxy for severity, under-enrolling high-risk Black patients (17.7% vs. potential 46.5%) due to systemic access barriers, paralleling M&E selection bias in impact forecasting. Switching to chronic condition counts tripled equity but highlighted needs for continuous surveillance in longitudinal evaluations.[nature]​