Governments have long struggled with a fundamental administrative dilemma: how do you get financial help to the people who need it most when those same people often leave the smallest footprint in official records? As traditional methods of identifying the poor become increasingly digitized, many are looking to machine learning as a high-tech solution to an age-old bureaucratic headache. However, a new analysis from the Development Policy Centre suggests that while algorithms can certainly sharpen our ability to spot poverty, they may also be creating a new set of risks that could undermine the very social protections they are meant to bolster.
The Limits of Digital Targeting
For decades, many developing nations have relied on the proxy means test, a system that estimates a household’s welfare based on observable traits like the type of roof on a home or the education level of its occupants. It is a practical, if imperfect, tool. Now, machine learning is being deployed to find more complex patterns, pulling from satellite imagery and mobile phone data to identify those slipping through the cracks. In Togo, for instance, the government’s Novissi program used these methods to reach vulnerable populations during the pandemic. By analyzing mobile phone activity, the program managed to identify 47% of poor individuals in rural areas, a notable improvement over the 33% reached through traditional geographic targeting.
But the data itself holds a mirror to existing inequalities. If you do not own a phone or lack a valid SIM card, you are effectively invisible to the algorithm. In Togo, this meant that roughly 15% of the population remained at risk of exclusion, regardless of how sophisticated the model became. Other regions have seen even more complicated results. A 2025 working paper by Emily Aiken and her co-authors examining programs in Bangladesh found that while mobile phone-based targeting was cheaper and more scalable, traditional survey-based methods remained more accurate at predicting actual consumption levels.
When Eligibility Rules Change Behavior
The challenge is not just about finding people; it is about how they respond to being watched. When governments tie social assistance to specific eligibility criteria—like employment status or income thresholds—they inadvertently influence the choices households make. It is a phenomenon that goes beyond simple work disincentives.
In a study published in the journal Labour Economics, researchers examined Indonesia’s large-scale unconditional cash transfers, which were rolled out in 2005, 2008, and 2013. The findings were revealing: following the introduction of these programs, overall employment dipped by about 2%, but formal employment—the kind that leaves a clear paper trail for government agencies—fell by roughly 8%. It appears that households were not necessarily leaving the workforce, but they were actively avoiding formal, documented jobs for fear that a stable, visible income would disqualify them from future assistance. This pattern is not unique to Indonesia; similar evidence has emerged in Uruguay, where formal employment rates declined when eligibility was explicitly linked to registered earnings.
The findings suggest that labour-market adjustments occurred primarily through reduced entry into formal employment, rather than widespread withdrawal from work or exits from existing formal jobs.
The Unseen Risks of Algorithmic Complexity
This creates a difficult paradox for policymakers. Machine learning models are often more complex and less transparent than traditional proxy means tests. While this might make it harder for individuals to “game” the system, it also introduces deep uncertainty. If a household does not understand exactly what characteristics—or what combination of digital footprints—determine their eligibility, they may become hyper-cautious. This could lead people to turn down economic opportunities or avoid formalizing their work simply because they cannot predict how those changes will affect their standing in the eyes of an opaque, algorithmic gatekeeper.
We are left with an uncomfortable reality. Improving the “accuracy” of a predictive model does not automatically result in better social protection. If we continue to improve our ability to target the poor while the same segments of the population remain invisible to our data—or if our systems inadvertently penalize those who try to climb the economic ladder—the fairness of our safety nets remains in question. As we lean further into these digital tools, the most important work may not be refining the algorithms themselves, but understanding the human responses they trigger in the communities they are intended to serve.
