The regulatory framework governing federal artificial intelligence, specifically the Office of Management and Budget guidance known as OMB M-25-21, contains a structural vulnerability that is currently being exploited to bypass essential oversight. The guidance defines high-impact AI as any system whose output serves as a principal basis for decisions or actions that have a legal, material, binding, or significant effect on rights or safety. By design, this definition is intended to capture the most consequential deployments of machine learning. However, the Department of Homeland Security has identified a semantic loophole within this language, effectively arguing that if a human is involved in the final decision, the AI output cannot be considered the principal basis for that action.
This interpretation functions as a categorical exemption. By asserting that human review acts as a buffer, DHS claims that its AI tools are merely advisory, regardless of their actual operational influence. This logic is being applied systematically across the department. In its own AI Use Case Inventory, DHS initially acknowledged that several of its systems were presumed to be high-impact under the criteria set by M-25-21. Yet, after applying the principal basis loophole, the department reclassified these tools to exclude them from the more rigorous transparency and safety requirements mandated for high-impact systems.
The practical reality of these systems contradicts the department’s defensive framing. Consider the Autonomous Surveillance Tower, an Anduril-manufactured system utilizing Lattice OS for real-time sensor fusion, object detection, and tracking. With over 300 units deployed along the southern border, the tower operates autonomously: it slews cameras to items of interest, classifies objects, and pushes alerts to agents. While DHS maintains that the AI merely alerts to the presence of an item it was trained to detect, the system independently selects what to ignore and what to flag. Because border agents cannot physically scan the entire border themselves, they are entirely dependent on the tower’s output. These alerts directly dictate where armed forces are dispatched, making the AI the functional, if not the formal, principal basis for tactical deployment.
A similar dynamic exists with ELITE (Enhanced Leads Identification and Targeting for Enforcement), a Palantir application used by ICE deportation officers. The tool maps potential deportation targets by aggregating data from sources including HHS/Medicaid, USCIS, DHS records, and Thomson Reuters CLEAR, assigning each target an address confidence score. Although DHS claims that outputs are limited to normalized addresses and that officers review and validate the information, this oversight is illusory. Agents cannot review or validate the underlying confidence-scoring algorithm. In practice, the system selects the targets, and the officers follow the algorithmic lead.
The defense of human review ignores the well-documented phenomenon of automation bias. As detailed in a 2025 systematic review by Romeo and Conti, which analyzed 35 studies involving nearly 20,000 participants, humans possess a cognitive tendency to favor automated recommendations over their own judgment, even when contradictory or more accurate information is available. This bias is driven by cognitive overload and misplaced trust, and it is often exacerbated by the perceived accuracy of the AI. The review found that Explainable AI and transparency mechanisms designed to mitigate automation bias often backfire: overly technical or even simplistic explanations can reinforce misplaced trust, especially among less experienced professionals.
As Emily Froude, a research analyst at Democracy Forward, noted in Tech Policy Press, agents are likely demonstrating automation bias—“an over-reliance on algorithmic output without being positioned to exercise meaningful scrutiny.” When the system is designed to filter reality, the human operator is not a check on the AI; they are a participant in an automated workflow.
The implications of this DHS interpretation extend far beyond a single department. By successfully utilizing this loophole to avoid high-impact classification, DHS has created a template that other federal agencies will inevitably copy. If the presence of a human in the loop is sufficient to negate the high-impact designation, then almost any AI system—no matter how consequential—can be shielded from the transparency requirements intended to protect the public. This creates a race to the bottom in regulatory compliance, where the definition of high-impact AI is rendered functionally meaningless.
Crucially, this evasion is entirely unnecessary. The high-impact classification under M-25-21 does not mandate the shutdown of these systems; it merely triggers basic transparency and accountability measures. These include conducting thorough privacy impact assessments—a requirement DHS has failed to meet for ELITE, as the cited document was last updated in 2019, well before the tool was acquired. By resisting this classification, the department is not protecting operational security; it is avoiding the basic documentation of how its most powerful tools function.
As of September 22, 2026—the compliance deadline for these regulations—the White House page for OMB M-25-21 returns a 404 error, a fitting symbol for the current state of federal AI oversight. The principal basis loophole allows agencies to maintain the appearance of human control while delegating critical decisions to opaque algorithms. Until the definition of high-impact AI is clarified to account for the reality of automation bias and the practical dependency of human operators, the regulatory framework will continue to fail in its primary objective: ensuring that the most significant AI systems are subject to meaningful public scrutiny.