Algorithmic Authority: Federal Agencies Are Making Life-Altering Decisions by Machine — and No Law Governs How
When Maria Delgado, a 58-year-old former warehouse supervisor in eastern Tennessee, applied for Social Security Disability Insurance following a severe back injury, she expected the process to be slow. She did not expect to be denied three times in fourteen months based on determinations she was never able to fully understand or meaningfully contest. It was only when a legal aid attorney filed a formal information request that Delgado learned a portion of her case assessment had been processed through an automated scoring system — one that the agency could not fully explain, whose underlying model had not been published for public review, and whose denial outputs had been affirmed by human reviewers in the overwhelming majority of cases with minimal independent analysis.
"The caseworker told me the system flagged my file," Delgado said. "But no one could tell me why, or what I could do about it."
Delgado's experience sits at the intersection of two realities that federal policymakers have been slow to confront: the rapid deployment of artificial intelligence and automated decision-making tools across the executive branch, and the near-total absence of a legal framework governing how those tools may be used, challenged, or constrained.
The Deployment Outpacing the Rulebook
The federal government's use of algorithmic systems is neither new nor limited to a single department. The Department of Homeland Security has used predictive risk-scoring tools in immigration enforcement for well over a decade. The Department of Housing and Urban Development has employed algorithmic systems to assess loan applications and flag potential fair lending violations — with the same underlying technology sometimes doing both. The Internal Revenue Service uses machine learning models to select audit targets. The Department of Veterans Affairs has piloted AI-assisted clinical decision tools in VA medical facilities across the country.
What is new — and what has outpaced the administrative law infrastructure designed to govern executive branch action — is the sophistication, opacity, and consequential scope of the latest generation of these systems. Modern machine learning models, particularly those built on large training datasets and complex neural architectures, do not produce decisions that can be easily traced to a specific rule or factual finding. They produce outputs. The reasoning that connects an input to an output may be, in a technical sense, inaccessible even to the engineers who built the system.
This creates a fundamental problem for administrative law, which is built on the premise that government decisions must be explainable, contestable, and grounded in ascertainable legal authority.
"The entire structure of the Administrative Procedure Act assumes that an agency action can be reviewed against the record that produced it," said a professor of administrative law at a major research university who has written extensively on automated governance. "If the record is an algorithm that nobody can fully interpret, the review mechanism breaks down. The legal architecture we have was not designed for this."
Statutory Silence as Implied Permission
Most federal agencies deploying AI systems do so without specific congressional authorization to use automated decision-making in the relevant context. The legal basis typically cited is a combination of the agency's general organic statute — the law that created it and defined its mission — and broad interpretive authority to determine how it carries out its functions.
This approach, which critics describe as treating statutory silence as implied permission, has been challenged in several federal courts with inconsistent results. In cases involving immigration risk assessments, courts have generally been reluctant to second-guess agency methodology choices. In cases touching on disability benefits and public housing eligibility, there is somewhat more judicial appetite for scrutiny — but the doctrinal tools available to reviewing courts remain limited.
The notice-and-comment rulemaking process established by the Administrative Procedure Act, which requires agencies to publish proposed rules, solicit public input, and respond to significant comments before a rule takes effect, has almost never been applied to the adoption of algorithmic decision systems. Agencies typically classify these systems as internal management tools or procedural mechanisms rather than substantive rules — a categorization that, if accepted, exempts them from the public participation requirements that would otherwise apply.
"If an agency changed its written policy for evaluating disability claims, it would go through notice and comment," noted a former senior official at the Office of Information and Regulatory Affairs. "If it changes the algorithm that produces the same determination, it often doesn't. The functional impact on applicants is identical. The procedural protection is not."
The Liability Vacuum
Beyond the question of how these systems are adopted lies a more immediate question for the Americans they affect: when an automated system makes a wrong decision, who is responsible?
Under the Federal Tort Claims Act, the government can be sued for certain categories of negligent conduct by federal employees. The applicability of that framework to harm caused by an algorithmic system — particularly one whose outputs cannot be attributed to a specific human decision — is, to put it charitably, unsettled. Several civil rights organizations have attempted to bring discrimination claims against agencies whose AI systems produced racially disparate outcomes, with limited success. Courts have generally struggled to apply disparate impact frameworks to systems whose discriminatory effects may be statistically demonstrable but mechanistically opaque.
The individuals most likely to be harmed by flawed automated determinations — applicants for disability benefits, public housing, and federally backed loans — are also the least likely to have access to the legal resources necessary to mount a challenge. The asymmetry is stark.
The Executive Orders That Changed Little
The Biden administration issued Executive Order 13960 in 2020 and its more comprehensive successor in 2023, establishing principles for trustworthy AI use in the federal government and directing agencies to conduct impact assessments on certain high-stakes automated systems. The Trump administration's 2019 executive order on maintaining American AI leadership emphasized innovation and competitiveness. Neither administration produced binding regulations with enforcement mechanisms that could be invoked by individuals harmed by agency AI systems.
Executive orders, by their nature, can be revoked or reinterpreted by subsequent administrations. They do not create judicially enforceable rights. They do not require congressional approval. And they do not fill the statutory gap that would give citizens a clear legal basis for challenging automated decisions that affect their lives.
Several legislative proposals have been introduced in Congress to establish minimum standards for federal AI deployment, including requirements for algorithmic impact assessments, transparency in high-stakes decisions, and individual rights to human review. None has advanced beyond committee consideration.
A Democratic Deficit With Practical Consequences
The absence of a governing framework for federal AI deployment is not merely a technical legal problem. It represents a meaningful erosion of the procedural protections that distinguish administrative governance from arbitrary bureaucratic action.
For Maria Delgado in Tennessee, the practical consequence of that erosion was twenty-six months without income while she navigated a system she could not understand, appealing decisions whose basis could not be adequately explained to her or her attorney. She was eventually approved for benefits following a hearing before an administrative law judge — a human being who reviewed her case on its merits.
The question that her case leaves unanswered, and that policymakers have yet to seriously engage, is how many applicants like her never reach that hearing. How many accept a machine's verdict because they lack the resources, the knowledge, or the stamina to contest it? And who, precisely, is accountable when the machine is wrong?
As of this writing, no federal agency has a satisfactory answer to any of those questions.