America's safety net has new administrators. They are not government agencies. They are nonprofits and are making decisions using algorithms.
Nationally, civil society organizations are using AI to determine who receives food assistance, emergency shelter, and medical referrals. In an era of federal and state budget cuts and struggles with the latest administrative databases that support necessary operations, nonprofits have gradually stepped into these roles. An unofficial AI welfare state is emerging, operating beyond the boundaries of official ones.
Government retrenchment over the past decade has pushed welfare functions toward civil society. Nonprofits are no longer just supplementing public services. In many communities, they are replacing them. AI accelerates this shift. Predictive tools that flag food insecurity, triage shelter demand or identify at-risk youth allow smaller organizations to scale their reach. This can be genuinely valuable. But scale introduced through automation carries its own systemic distortions. When machines allocate scarce social resources, the political judgments encoded in those models become invisible. Accountability disappears with them.
The accountability problem is structural. When a government agency determines welfare eligibility, legal safeguards apply due process rights, administrative review and legislative oversight. When a nonprofit uses a proprietary AI system to make equivalent decisions, none of those protections necessarily follow.
Algorithmic bias in social service allocation is not hypothetical. Research demonstrates that seemingly neutral models reproduce structural inequalities, systematically denying help to the communities most likely to need it. The vendors building these tools are not elected. Neither are the nonprofits deploying them. The individuals affected hold no formal right of appeal. Closing the accountability gap in algorithmic systems requires deliberate institutional design — it does not self-correct.
When nonprofits use AI, they usually transfer decision-making authority to their software vendors and proprietary models. These systems now indirectly shape who receives food assistance, job training, or mental health services. That is a welfare function. It carries welfare stakes. Algorithmic systems embedded in social institutions tend to entrench existing power asymmetries rather than correct them. The most affected citizens in communities rarely participate in designing or governing the tools deployed on their behalf.
Not all nonprofits have the same capacity. In metropolitan areas, large institutes have the budgets to acquire sophisticated AI platforms. Many of the smaller nonprofits serving rural & low-income populations in America simply cannot. The distribution of AI-assisted service delivery mirrors and monitors the geographic features of poverty itself. Economic mobility is already highly correlated with social capital and access to institutional resources. If an uneven AI infrastructure is laid on top of such a stratified social system, the result can be to exacerbate, rather than reduce, entrenched inequalities.
Federal AI policy has concentrated nearly exclusively on commercial and national security applications. There is hardly a thing that regulates the use of AI when it comes to nonprofit welfare delivery. But as long as policy interventions (e.g. anthropogenic policies) that call for human intervention in algorithms do not provide binding transparency, fairness and appeal standards, they remain structurally deficient. Nonprofits are behaving like quasi-governmental functions without the accountability of a quasi-governmental status. It is not a technicality, but a structural failure of democratic governance.
Congress and state legislatures need to do three things.
First, expand algorithmic accountability standards to all nonprofits implementing government-supported social services.
Second, require algorithmic impact assessments prior to deploying AI tools in welfare-adjacent programs.
Third, fund independent ongoing audits of all systems used by states to administer food, housing, and healthcare assistance.
America did not consciously choose to privatize its welfare state through software. It is happening by inaction, one algorithm at a time. The decision to govern this shift remains available. But the window is closing fast.
Feroz Ahmed is a Research Assistant at Lamar University, USA, where he conducts research on Management Information Systems (MIS), digital governance, artificial intelligence, and data-driven technologies. Alongside his academic work, he also serves as a 'System Security Analyst' with Amazon & Securitas Security Services USA Inc. in Seattle, Washington, where he is involved in security operations and technology risk monitoring.
