AI makes it easier to ration welfare to narrowly defined recipients, risking deeper exclusion and entrenched inequality; whether automation narrows or protects social protection hinges on political choices to uphold universal, rights-based systems.

Main Finding

AI and algorithmic systems are reshaping how welfare states decide “who is deserving.” By enabling large-scale datafication, predictive profiling, and automated administration, these technologies can both improve precision in targeting and reduce non-take-up, but they also risk amplifying bias, systemic error, exclusion, and conditionality. Whether AI leads to fairer, more efficient welfare outcomes or to intensified differentiation between “deserving” and “undeserving” citizens depends crucially on technological design, data quality, administrative practice, and political choices.

Key Points

• Digitalization aligns well with the social investment / investment-intervention paradigm: AI can support early, individualized interventions, faster resource allocation, improved matching (jobseekers to vacancies), and potentially higher administrative efficiency.
• Targeting vs universality trade-off: algorithmic targeting promises reduced waste and higher impact per euro spent, but targeting historically risks lower benefits, intrusive controls, complex administration, and stigma — risks that can be magnified by automation.
• Non-take-up (NTU): Big, linked administrative datasets and automated enrollment offer paths to reduce NTU (auto-enrollment, proactive outreach). But automation can also create new barriers for digitally excluded or low-literacy populations.
• Bias and systemic amplification: Algorithms trained on administrative or historical data reproduce the biases embedded in those data; automated rules can scale up errors affecting thousands, compounding harms compared to isolated human discretion.
• Proxies and simplification: Reliance on digital proxies (e.g., click behavior to infer job search effort) risks oversimplifying complex social realities and misclassifying needs.
• Transparency and accountability gaps: Lack of explainability, weak oversight of private contractors, and reduced human oversight can produce “seeing without knowing” administrative regimes.
• Empirical harms documented: Case studies (e.g., Indiana welfare automation, UK Universal Credit and Australia Centrelink systems) show increased erroneous overpayments, debt, and hardship when automation is poorly designed or lacks human safeguards.
• Political and design contingent: National politics determine whether datafication expands rights (proactive delivery) or tightens conditionality and surveillance; data & design choices (what variables used, thresholds, human-in-loop rules) crucially shape distributional outcomes.

Data & Methods

• Nature of the chapter: conceptual synthesis and literature review drawing on empirical case studies, policy reports, and academic analyses rather than a new primary dataset.
• Evidence sources cited or discussed:
  • Policy/advocacy reports (Algorithm Watch Automating Society report; Eurofound; UN Special Rapporteur on extreme poverty).
  • Empirical case studies and book-length investigations (Eubanks on Indiana automation; studies of Universal Credit and Centrelink automation).
  • Research on public employment services using profiling and matching algorithms (Körtner & Bonoli; Brioscú et al.).
  • Scholarship on deservingness, social investment paradigm, and targeting (van Oorschot; Van Kersbergen & Hemerijck; Cantillon & Van Lancker).
• Methods used in the referenced literature:
  • Comparative case studies and investigative reporting.
  • Administrative data linkage and predictive modeling as deployed in practice.
  • Policy analysis and normative critique of welfare design.
• Gaps noted: limited systematic causal evidence on large-scale welfare AI impacts; need for rigorous evaluations (RCTs, quasi-experimental designs), distributional monitoring, and algorithmic audits.

Implications for AI Economics

• Redistribution and efficiency trade-offs: Algorithms can improve allocative efficiency and reduce administrative costs, but standard economic evaluation must incorporate distributional impacts, stigma costs, enforcement costs, and political economy (who bears errors).
• Fiscal risk and scaling externalities: Algorithmic mistakes (false positives/negatives, threshold errors) can create large-scale fiscal shocks (mass overpayment/underpayment, debt recovery costs). Economists should quantify tail risks from systemic errors, not just average gains.
• Take-up and behavioral responses: AI-mediated auto-enrollment or proactive outreach can raise take-up, changing effective coverage and consumption smoothing — incorporate these behavioral responses into welfare-benefit models and microsimulations.
• Labor market matching and human capital: Better matching via AI in public employment services could raise employment outcomes but also change incentives, job search behavior, and upskilling returns; evaluate impacts on long-run earnings, job stability, and inequality.
• Measurement and fairness metrics: Standard welfare economics should be extended to include fairness-aware metrics (e.g., disparate impact, calibration across groups) and to weight equity in cost–benefit frameworks.
• Evaluation and methodology recommendations:
  • Use randomized controlled trials, phased rollouts, or difference-in-differences on administrative outcomes to estimate causal effects of algorithmic interventions on take-up, employment, poverty, and fiscal costs.
  • Combine administrative microdata with survey data to assess misclassification, NTU, and household consequences of automation errors.
  • Conduct ex ante simulations and stress tests of algorithms to assess distributional consequences and tail-risk scenarios.
  • Implement regular algorithmic audits, bias tests, and counterfactual checks (what would human caseworkers have decided).
• Governance and policy design (economic policy implications):
  • Require human-in-the-loop for sensitive eligibility decisions or at least clear remediation pathways and rapid appeals; embed adaptive monitoring with financial contingencies for error correction.
  • Prefer designs that expand rights via automation (auto-enrollment, linkage across benefits) where feasible, rather than designs that primarily tighten conditionality and surveillance.
  • Build regulatory standards for transparency, explainability, and data quality; require disclosure of variables used, performance by subgroup, and error rates.
  • Consider alternatives or complements (e.g., universal entitlements, simplified means-testing, UBI-style transfers) where administrative automation risks unacceptable exclusion or welfare losses.
• Research agenda for AI economists:
  • Quantify net welfare effects (efficiency vs equity) of algorithmic targeting versus universal approaches.
  • Estimate NTU reductions and related fiscal implications from proactive data-driven enrollment strategies.
  • Model behavioral responses to profiling (search effort, work incentives, stigma).
  • Develop methods to internalize algorithmic risk into social cost–benefit analysis and optimal policy design.

Short summary takeaway: AI can materially improve the precision and timeliness of social investment policies, but economic analysis must explicitly account for distributional consequences, systemic risks, and governance constraints. Without careful design, transparency, and evaluation, algorithmic welfare administration risks deepening exclusion and undermining core social protection objectives.