Algorithmic Decision-Making In Welfare Systems: Equity, Bias, and Democratic Accountability In Social Service Access

The application of algorithmic decision-making systems in welfare administration has grown fast in the public sector, wherein automated model is used to determine eligibility, fraud detection and the distribution of social benefits. Such systems are touted as efficient and consistent, but the accumulating body of empirical evidence demonstrates the risk of bias (Levy, 2021; Kasy, 2024). The algorithms that learn from historical data tend to replicate social inequalities with biased proxies, including income, spending on health, or geographic factors, which could likely skew decision-making results (Kasy, 2024). These distortions ask important questions about equity, access and erosion of democratic accountability in the delivery of services. This study will examine the existence of systematic access disparities to social services in the context of algorithmic decision-making in welfare systems; and measure how demographic factors can determine the approval of welfare and assess the consequences on equity and responsibility in the systems that are under the control of algorithms.

A quantitative research design is adopted using secondary administrative-style data representing welfare applications. The dataset includes key variables: application outcome (approved/rejected), income level, employment status, age, and gender. The logistic regression modelling will approximate the likelihood of approval in various demographic groups, and helps identify statistically significant differences. Furthermore, there are measures of fairness such as disparate impact ratios and outcome parity measures, which are used to measure distributional inequality in decision outcomes (Zhou et al., 2021). To reinforce the analysis, a comparative model is created to model human-based decision benchmarking against algorithmic results. A structured survey element is included to obtain the perception of fairness and trust in automated welfare systems, as they affect legitimacy and acceptance (Starke et al., 2022).

The results show a statistically significant differences in the probability of approval based on income and employment. Low-income applicants have less probability of approval, even after adjusting factors related to eligibility, indicating a possible existence of indirect algorithmic bias. Differences in access amongst demographic groups are found to be significantly lower, according to disparate impact analysis, showing that some groups have lower access rates systematically, which is consistent with previous findings of algorithmic systems to reinforce structural inequalities (Ebrahimi et al., 2024). People feel that algorithmic decisions are less in cases where system-based processes are not transparent, which supports the idea that procedural justice and lack of trust in automated systems is a challenge to automated governance (Decker, 2025).

The implications of these findings are that there is a severe conflict between efficiency and equity in algorithmic welfare systems. Although these systems contribute to the capacity of administration, they pose the risk of entrenching and multiplying inequities unless they are regulated appropriately. The results confirm claims that a robust accountability regime should govern algorithmic systems in the public decision-making process because delegating discriminatory outcomes to automated systems does not make institutions less responsible (Alon-Barkat, 2025). This paper adds an empirical data to the discussions about algorithmic governance and highlights the importance of fairness auditing, transparency systems, and regulatory controls.