How ICE's AI Systems Create a Transparency Gap

The useful starting point for understanding ICE's AI systems in immigration enforcement and detention is not a vendor demo or a lawsuit allegation. It is the Department of Homeland Security's own AI Use Case Inventory for U.S. Immigration and Customs Enforcement, updated July 15, 2026. On that page, ICE acknowledges roughly 20 AI use cases across deployed, pilot, and pre-deployment status, including biometric matching, facial recognition, language translation, address extraction, redaction, draft investigative reports, open-source intelligence support, and other operational tools.[1]

The inventory is not a complete map of every automated system that may affect a person's encounter with ICE. It is self-reported, and contractor-operated tools appear only indirectly, if at all. But it gives an official floor. It shows what ICE is willing to say it uses, how DHS labels those uses, and which systems DHS treats as triggering its more demanding risk-management obligations.

That is where the map becomes strange. Of the ICE use cases disclosed in the inventory, only four are marked as high-impact: Mobile Fortify, Facial Recognition for Vulnerable Populations, Facial Recognition for National Security, and Real-Time Language Translation Services.[1] Several other tools sit closer to enforcement than their labels suggest. ELITE extracts addresses from criminal records for immigration enforcement targeting. Draft Report Generation for Investigations would help prepare investigative documentation. AI-assisted redaction processes evidence materials. Pre-deployment open-source intelligence tools are designed to collect or organize information from public sources. Yet DHS does not classify many of these as high-impact in the inventory.[1]

The legal problem is not solved by saying a human officer still signs the final form. In immigration enforcement, the hard-to-challenge moment often arrives earlier: when a person is identified in the field, when an address is treated as reliable enough for follow-up, when a file is assembled, when detention risk is scored, or when an investigative summary starts to harden into the record a lawyer later receives. A tool can shape those moments without being the formal “principal basis” for the final agency action.

Why the “Principal Basis” Standard Leaves So Much Outside the Frame

The Brennan Center's analysis of DHS AI transparency identifies the hinge: DHS reads the Office of Management and Budget standard narrowly, so a system may avoid high-impact treatment if the AI output is not the principal basis for a rights-impacting decision.[2] That distinction matters administratively. It decides which systems must receive the fuller package of risk-management practices attached to high-impact AI. But it is a poor fit for enforcement workflows, where one output can become a lead, a lead can become a location check, a location check can become an arrest, and the downstream record may not show which upstream tool made the encounter possible.

The inventory therefore answers a narrower question than the one lawyers usually need to ask. It tells the public how DHS classifies an AI use case for internal governance. It does not establish whether the tool materially contributed to a stop, arrest, custody decision, bond posture, removal planning step, or evidentiary record. Those are different questions, and the distinction should not be treated as technical housekeeping.

Identification in the Field: Mobile Fortify and Facial Recognition

Mobile Fortify is the clearest high-impact example because DHS labels it that way and because its function is direct: biometric matching in field enforcement.[1] If the system helps an officer decide who a person is, the error does not stay inside a database. It can move into handcuffs, transportation, detention intake, and the first frantic attempt by counsel or family to understand what happened.

Public reporting has already tied AI-assisted identification concerns to wrongful-arrest allegations. PBS News reported on a federal judge's note raising accuracy and privacy concerns about immigration agents' use of AI, including the case of Jonathan Guerrero, described as a U.S. citizen arrested after an AI-driven error in 2025.[3] Context by the Thomson Reuters Foundation also reported on AI's role in immigration enforcement and discussed concerns involving Mobile Fortify and cases including Jensy Machado.[4]

Those reports do not prove that every biometric match is unreliable, and they should not be stretched into that claim. The narrower point is stronger: when biometric output is used in the field, the practical burden of correcting a false identification may fall on the person already detained or on counsel trying to reconstruct a fast-moving encounter after the fact. DHS's high-impact label at least acknowledges that Mobile Fortify belongs in the rights-and-safety category. The harder question is whether similar practical consequences are being missed when the tool is not biometric.

Address Intelligence and Targeting: ELITE, ImmigrationOS, and OSINT

ELITE is where the high-impact boundary looks least persuasive. In the DHS inventory, ELITE is described as using AI to extract addresses from criminal records for immigration enforcement purposes, yet it is not marked high-impact.[1] Address extraction may sound clerical until it becomes the step that decides where agents look, whose home is associated with a target, or which family members and roommates are pulled into an enforcement event.

The Immigration Policy Tracking Project reports that Palantir was awarded a $30 million contract in April 2025 to build ImmigrationOS for ICE, and describes ELITE as a component connected to deportation targeting, address-confidence scoring, and mapping of targets.[5] FedScoop's reporting on the DHS inventory and Palantir situates the system within ICE's disclosed AI environment.[6] Separately, 404 Media reporting cited by the Immigration Policy Tracking Project described a geospatial heatmap feature for raid planning.[5]

An address-confidence score is a small phrase with a large litigation shadow. If an address is scored as likely, that score may influence where officers go, which residence is entered or surveilled, and whose documents are later treated as connected to the target. But unless the score is preserved, disclosed, and explained, counsel may see only the downstream paperwork. The file may say where officers went. It may not say how that address became operationally credible.

The same issue appears in pre-deployment open-source intelligence tools. OSINT systems may collect, sort, or summarize public information, but public availability does not make the resulting enforcement use transparent. A social media post, business listing, court record, data-broker entry, or public database hit can be misread, outdated, or linked to the wrong person. If the AI system merely informs an analyst rather than dictates an enforcement decision, DHS may treat it as outside the high-impact category. For the person whose location or identity is inferred, that distinction offers little practical visibility.

Detention Risk After Hurricane Score

Hurricane Score should not be treated as the whole story simply because it had a name. FedScoop reported that the algorithmic risk-assessment tool was retired in June 2026.[6] That retirement matters, but mostly as a marker. Risk assessment did not disappear from detention practice when one tool was retired.

ICE's Risk Classification Assessment process remains important because it structures detention-related evaluation. A DHS Office of Inspector General report issued in June 2024 found documentation problems in the RCA process, including issues that make it harder to determine whether custody classification decisions were properly supported.[7] If automated or algorithmically assisted inputs feed into that environment, the documentation gap becomes more consequential. A lawyer cannot test what the record does not reveal.

The inventory's labels do not answer whether a particular detention decision relied on an automated score, a contractor-generated lead, a risk category, or an officer's independent judgment. They also do not show whether an old tool's retirement produced a less risky process, a renamed process, or a more distributed one. The responsible conclusion is limited: Hurricane Score's retirement narrows one named issue, but it does not eliminate the need to examine algorithmic inputs in detention classification and risk review.

Evidence, Redaction, and the Record a Lawyer Actually Receives

ICE's disclosed AI use cases are not limited to finding people. They also reach the evidence and documentation layer. The DHS inventory lists AI-Assisted Audio/Video Redaction as deployed and Draft Report Generation for Investigations as pre-deployment.[1] These systems may seem administratively safer than field identification or address targeting, but they affect the record through which later accountability is usually pursued.

Redaction can protect privacy and improve production speed when it is carefully controlled. It can also obscure context if the process over-redacts, misidentifies a speaker, fails to preserve an original, or makes it difficult to determine whether relevant material was withheld. Draft report generation raises a different concern: if an AI-assisted summary shapes the first written account of an investigation, later review may treat that account as ordinary officer narrative unless the AI role is disclosed.

The public inventory does not provide evidentiary standards for how AI-assisted documentation should be preserved, labeled, challenged, or authenticated in immigration proceedings or related federal litigation.[1] That absence matters even if the tools perform well most of the time. Legal process depends not only on accuracy, but on the ability to inspect the path from raw event to official record.

The Contractor Gap: Skip Tracing Without Itemized AI Use Cases

The DHS inventory is least satisfying where the operational system sits outside the clean boundary of an ICE-owned use case. Jeelani Law Firm's April 2026 analysis describes up to 13 private companies, $1.2 billion in contracts, and AI skip tracing on up to 50,000 individuals per month.[8] The Immigration Policy Tracking Project and related reporting also describe large private-sector involvement in ICE's data and targeting infrastructure.[5]

Those figures should be read carefully. They come from outside the DHS inventory, and the inventory itself does not itemize each contractor-operated AI system. But that is precisely the oversight problem. If a contractor's model links a person to an address, phone number, workplace, vehicle, or relative, and ICE later acts on the lead through a human officer, the AI system may be absent from the official AI inventory as a named operational tool. The consequential step becomes procurement infrastructure rather than a disclosed use case.

For counsel, the practical question is not whether the government calls the system AI. It is whether a contractor-generated inference helped put the client at a location, connect the client to a record, or elevate the client for enforcement attention. A non-itemized system can still create discovery, records-request, and due-process questions if its output materially shaped the chain of events.

What the Inventory Helps Prove, and What It Does Not

The inventory is valuable because it prevents the debate from floating above the systems. It names tools. It assigns statuses. It shows that DHS can recognize at least some ICE AI uses as high-impact. It also reveals the classification gap by omission: the tools closest to targeting, address intelligence, and documentation may be treated as ordinary support systems even when their outputs move through the same enforcement chain as the high-impact tools.

That does not mean every non-high-impact system is unlawful, inaccurate, or outcome-determinative. The inventory does not support that claim. It supports a more disciplined one: DHS's classification is an administrative designation under a narrow governance standard. It should not be mistaken for an answer to whether the tool affected liberty, location, identification, detention posture, or the integrity of the evidentiary record.

The Due-Process Questions Legal Professionals Should Preserve

The most important practice implication is modest but concrete. When a client has been arrested, detained, classified, or located through a chain that may involve automated tools, counsel and researchers should not stop at the phrase “human review.” They should ask where the information entered the file, what system produced or ranked it, whether confidence scores existed, whether contractor outputs were used, and whether the original materials are preserved.

• For identification disputes, ask whether Mobile Fortify, facial recognition, or another biometric system contributed to the encounter.
• For home, workplace, or address-based enforcement, ask whether ELITE, ImmigrationOS, OSINT tools, skip tracing, or address-confidence scoring informed the location.
• For detention classification, ask whether any risk score, RCA input, automated recommendation, or contractor-generated data influenced the custody posture.
• For evidentiary records, ask whether audio, video, summaries, reports, translations, or redactions were created or altered with AI assistance.
• For records requests and policy research, treat the DHS inventory as a starting index, not as a complete list of operational systems.

These questions do not assume that a court will treat every upstream AI contribution as legally decisive. The relevant doctrine may develop unevenly, and the “principal basis” standard has not been resolved in the ICE context in a way that answers all future disputes. But case preparation and policy analysis do not have to wait for a definitive appellate rule before preserving the factual chain.

ICE's AI transparency gap is not simply that the agency uses automated tools, or that DHS discloses too little. It is that disclosure and classification are misaligned with how enforcement work actually happens. A system can be labeled non-high-impact and still help identify a person, select an address, assemble a file, shape a detention risk review, or generate the document that later becomes the government's account. Legal professionals and policy researchers should treat DHS's non-high-impact labels as administrative classifications, not as due-process conclusions.

References

1. DHS AI Use Case Inventory: ICE, Department of Homeland Security, July 15, 2026.
2. A Start on AI Transparency at DHS, With Room to Grow, Brennan Center for Justice.
3. Judge's note on immigration agents using AI raises accuracy and privacy concerns, PBS News.
4. How AI is aiding Trump's immigration crackdown, Context by Thomson Reuters Foundation.
5. Reported: Palantir awarded $30 million to build ImmigrationOS surveillance platform for ICE, Immigration Policy Tracking Project, April 2025.
6. DHS AI inventory includes Mobile Fortify, Palantir tools, FedScoop.
7. OIG-24-31-Jun24, DHS Office of Inspector General, June 2024.
8. ICE Use AI to Track Immigrants, Jeelani Law Firm, April 2026.