The Dignity Gap: Redefining Poverty as a Multidimensional Computational Target and Its Implications for AI-Assisted Welfare Governance

Official poverty measurement in India classifies 11.28% of the population as poor (NITI Aayog MPI 2022-23 round). We argue this figure reflects a definitional choice calibrated to administrative manageability rather than human reality, and formalise it as objective misspecification (Krakovna et al. 2020) at the institutional level: the official target is a truncated, binarised functional of the welfare objective the state is mandated to pursue. Applying a three-tier taxonomy – Survivability Failure, Dignity Poverty, and Aspirational Exclusion – grounded in published national survey data (GHI 2025; FAO-SOFI 2025; ASER 2023; NHA 2021-22; NITI Aayog 2024) and a dependency-free bound on inter-survey overlap, we estimate a corrected headcount of 58–69% of India’s 1.4 billion population, an undercount of 5.1–6.1×. We introduce the Dignity-Adjusted Poverty Index (DAPI), extending Alkire-Foster (2011) with continuous deprivation scoring over six independently verifiable dimensions weighted by an entropy-based dispersion measure (Zeleny 1982) over NFHS-5 district frequencies, robust to ±20% perturbation. We then give three results on welfare routing: under additive separability, scheme-centric allocation is strictly Pareto-dominated by a person-centric formulation on the same budget; computing the person-centric optimum is NP-hard by reduction from the Generalised Assignment Problem; and a polynomial-time cost-benefit greedy router attains a (1-1/e) approximation under a submodularity condition we verify for DAPI. On a synthetic population of 50,000 citizens calibrated to NFHS-5 marginals, the router more than doubles deprivation reduction per rupee and, with a non-exclusion constraint, eliminates false exclusion at negligible welfare cost. Critically, the status quo’s false-exclusion rate is near-invariant across an eightfold budget increase: exclusion is architectural, not budgetary. We conclude with four minimum computational requirements for AI-assisted welfare coordination at scale.