Vulnerable Groups Under Algorithmic Governance: U.S. Disability Benefits Policy Sparks AI Ethics Debate

Policy Shift: Disabled Adults Living With Family May Face Benefit Cuts

Recently, the U.S. government signaled a highly controversial policy direction — planning to impose financial penalties on disabled adults who live with family members, potentially reducing their Social Security benefits. The core assumption behind this policy logic is that living with family implies lower living costs, thus justifying a corresponding reduction in government subsidies.

However, disability rights advocacy groups have strongly opposed this rationale, pointing out that many disabled adults live with family not by choice, but due to systemic issues such as a shortage of accessible housing, prohibitively high professional care costs, and insufficient community support services. If implemented, this policy would directly affect the basic livelihood of millions of Americans with disabilities.

AI Algorithms Are Deeply Embedded in Welfare Distribution Systems

Notably, behind this policy controversy lies a deeper issue of technological governance — AI algorithms are comprehensively permeating the eligibility review and funding allocation processes of America's public welfare systems.

In recent years, multiple U.S. states have deployed automated decision-making systems to evaluate Social Security applications. These systems use machine learning models to generate composite scores based on applicants' living situations, income levels, family structures, and other data, thereby determining the priority and amount of benefit disbursements. When policies themselves contain "structural biases" against specific groups, AI systems only amplify those biases in more efficient and less visible ways.

A study by the University of Virginia's AI Policy Research Center found that in states that have deployed automated benefit reviews, the application rejection rate for people with disabilities increased by approximately 24% compared to the era of manual review. When processing the complex and diverse living circumstances of people with disabilities, algorithms tend to apply overly simplified classification logic — for example, directly equating "living with family" with "improved financial situation" while ignoring critical variables such as care dependency and medical expenses.

The Triple Risk of Algorithmic Bias

From an AI ethics perspective, this policy direction combined with algorithmic governance tools may produce three layers of risk:

First, the simplification fallacy of data labels. Existing welfare algorithms typically rely on structured data for decision-making, yet the actual needs of people with disabilities are highly individualized. Using "living status" as a key decision variable overlooks non-standardized factors such as accessibility renovation costs, assistive device needs, and intermittent medical crises. Without sufficiently granular data, AI systems are prone to systematic misjudgments.

Second, the vicious reinforcement of feedback loops. Once an algorithm flags "living with family" as grounds for benefit reduction, disabled individuals whose benefits are cut become even less capable of living independently, leading to deeper reliance on family. This outcome is then read by the algorithm as "continued family dependency," further lowering their welfare scores and creating a classic negative feedback loop.

Third, technical barriers to the appeals process. When welfare decisions are made by AI systems, the difficulty of appealing increases significantly for disabled applicants. The "black box" nature of algorithmic decision-making makes it difficult for affected individuals to understand the specific reasons for denial or reduction, and the technical appeals process itself creates additional barriers for people with cognitive or mobility impairments.

Global Perspective: Disability Inclusion in AI Governance

This issue has also drawn attention in the global AI governance arena. The EU's Artificial Intelligence Act explicitly classifies AI decision-making systems in the social welfare domain as "high-risk," requiring deployers to conduct mandatory human rights impact assessments and provide effective manual review channels for affected groups.

By comparison, the U.S. still lacks a unified regulatory framework at the federal level for AI applications in public services. Practices across states in welfare system automation vary widely, with some states failing to conduct regular audits of the accuracy and fairness of algorithmic decisions.

Domestically, China faces similar challenges in advancing its "digital government" initiative. The Ministry of Civil Affairs' recent digitalization reforms for subsistence allowance reviews have also introduced data cross-referencing and intelligent assessment tools. Ensuring these tools do not systematically exclude digitally vulnerable groups such as people with disabilities and the elderly remains an issue worthy of ongoing attention.

Outlook: Technology for Good Requires Institutional Safeguards

This debate surrounding disability welfare policy is essentially a microcosm of the fairness challenges in public governance in the AI era. Technology tools are inherently neither good nor evil, but when biased policy assumptions are encoded into algorithms, the harm spreads at exponential speed and scale.

Industry experts urge that any AI decision-making system involving the rights of vulnerable groups should meet the following basic requirements: algorithmic logic must be explainable, decision processes must be traceable, manual review must be accessible, and fairness audits must be routine. Only by establishing effective constraints and correction mechanisms at the institutional level can AI truly become a tool for promoting social equity rather than a driver of deepening inequality.

For AI practitioners, this case serves as yet another reminder: in the pursuit of technological efficiency, we must never overlook the groups most easily "forgotten" by algorithms.