MobileAid: Machine-Learning Poverty Targeting from Mobile-Phone Metadata

Bangladesh's a2i (Aspire to Innovate) program, part of the government's ICT Division, worked with UNDP, GiveDirectly and University of California, Berkeley researchers to test whether machine learning applied to mobile-phone metadata could identify poor households for a cash-transfer program more cheaply than traditional door-to-door surveys.

Anonymised call-detail records from all four national mobile operators — over 1,500 features per household, covering call frequency, SMS use, airtime recharge patterns, contact diversity and cell-tower location — were used to rank households recruited through the AI-enabled national helpline 333, across a census of 106,200 households in 200 villages in Cox's Bazar district. Selected households each received a one-off transfer of BDT 15,000 (about $136).

An independent World Bank evaluation, reported by The Daily Star, compared the accuracy of three targeting methods: traditional proxy-means testing achieved 'High' accuracy (AUC 0.78-0.85), phone-metadata machine-learning targeting scored 'Medium' (AUC 0.61-0.72), and community-based targeting scored 'Low' (0.58). Six percent of households could not be matched to any phone record at all, and only 45% of recipients perceived the phone-based method as fair.

On raw accuracy, the AI method underperformed the method it was meant to modernise. GiveDirectly's own analysis argues it remains the most cost-effective option once per-household screening budgets fall below roughly $10-50, because of its speed and low marginal cost at scale. Both the World Bank evaluation and GiveDirectly's own materials flag the same structural risk: the poorest, women and people with disabilities are the least likely to own a phone, so metadata-based targeting risks systematically excluding exactly the households a safety-net program most needs to reach.