The Waiver That Makes the Rule Optional

The Department of Homeland Security’s updated directive on artificial intelligence, effective since September 2026, keeps a prohibition against using AI for “systemic or large-scale monitoring, surveillance, or tracking of individuals, or predictions of emotional state or future behavior.” [1] The same document hands the agency’s chief AI officer the power to waive any requirement it contains, including that prohibition. A safeguard and its cancellation sit a few pages apart in the same policy. That arrangement is the story: a written judgment about what AI must not do, placed in the hands of an official who can set the judgment aside.

A Prohibition With a Handle on the Side

The text of the directive is largely continuous with its 2025 predecessor in scope and content, and the surveillance prohibition survived the revision nearly intact. What the revision adds is the escape hatch. The chief AI officer may waive any requirement outlined in the directive, prohibited use cases included, and the policy attaches no independent review by the privacy or civil rights offices, no notice to Congress or the public, and no expiration date. Tom Bowman, policy counsel for the Center for Democracy & Technology’s Security and Surveillance Project, put the structural problem plainly:.” [1] The waiver is not hidden in an annex. It is a feature of the governance architecture, and it means the prohibition’s force depends on a single official’s decision not to use it.

The Contract Next Door

The timing gives the waiver its edge. In September, Immigration and Customs Enforcement awarded a $125 million contract to Thomson Reuters Special Services for services that would enable monitoring of 1 million individuals and identifying potentially criminal and fraudulent behavior before it happens. DHS had been developing the surveillance system with CACI since at least 2020, according to documents obtained through a public records request by Democracy Forward and shared with FedScoop. An ICE spokesperson said use of “Super Intelligence” by the agency is subject to applicable governance, use-case documentation and reporting requirements, and that if Thomson Reuters Special Services uses the technology in support of the contract, the agency’s chief information officer will assess the specific role and intended use for consistency with the directive. That assessment is exactly the kind of judgment the waiver provision lets an agency official make alone, and set aside.

What the Waiver Replaces

Before the revision, the directive assigned the chief privacy officer responsibility for ensuring AI “does not erode” privacy protections. The new version asks only that AI “complies” with established guardrails. The difference is not stylistic. One formulation asks an official to weigh whether a system diminishes privacy in practice; the other asks whether it stays inside lines that the agency itself draws and can move. Rachel Levinson-Waldman, director of the Brennan Center’s Liberty and National Security program, called the waiver authority one of her biggest concerns. [1] She noted the directive requires waivers to be “consistent with law, regulation, and government-wide policy,” then added: “Any CAIO worth his salt will be able to assemble a justification for a waiver that explains away any tension with other policies, especially given the sweeping authority granted by this directive.” The privacy officer’s narrowed role and the chief AI officer’s expanded one point the same direction: discretion migrates toward the office that wants to deploy.

Sharing Without a Purpose

The Waiver That Makes the Rule Optional (Image 1)

The revision also adds language on AI asset sharing and reuse, aligning with the Office of Management and Budget’s April 2025 memorandum on accelerating federal use of AI. DHS is instructed to proactively share custom-developed code, AI models, use cases and related data across the agency and governmentwide, and to release and maintain AI code as open-source software in a public repository. Bowman saw a possible transparency gain in the open-source preference, contingent on follow-through, but identified the civil liberties problem as purpose limitation. A model built, tested and reviewed for one mission will be picked up by another component for a different mission, while the privacy and civil rights analysis that justified it the first time may be entirely different. Reuse separates a model from the reasoning that once licensed its use. The judgment that cleared the tool for one job does not travel with it, and under the new policy nothing requires that it should.

The Precedent Is the Architecture

This is not the first time a government has written a restraint and then written itself permission to ignore it. Export-control regimes have long paired prohibitions with licensing exceptions, and the exception, not the prohibition, tends to govern the flow of goods. Surveillance law has followed the same pattern: a statutory wall, a carve-out for emergencies or foreign intelligence, and a practice that grows to fill the carve-out. The lesson from those regimes is consistent. Where a prohibition can be lifted by the same office that administers it, the prohibition describes an aspiration rather than a constraint. DHS has now built that structure into AI governance, and it has done so while its enforcement arm is buying the capacity to monitor a million people.

The Question the Policy Leaves Open

The removal of language on equity, underserved communities, environmental efficiency and sustainability was expected and, as FedScoop reports, reflects the administration’s priorities. The narrowing of the privacy officer’s mandate and the widening of the chief AI officer’s discretion were choices of a different kind, and they are the ones that will decide what the directive actually forbids. So the open question is not whether DHS wrote a prohibition on large-scale surveillance. It did. The safeguard at its center looks less like a wall than a door. —.

Sources

Fedscoop — Quote source (original article)

The Waiver That Makes the Rule Optional (Image 2)