The federal government built a surveillance app that pulls personally identifiable information from a dozen databases, deployed it to deportation officers hired on an expedited timeline with reduced training, and skipped the privacy assessment that a 2002 law says it should have published first.
That is the structural shape of what The Intercept reported this week about ICE’s use of Palantir’s ELITE platform, and it is worth reading slowly, because the individual pieces are less alarming than the way they fit together.
Consider the sequence.
ICE added more than 12,000 employees through a hiring surge that began in 2025. Reuters reported in February that thousands of those new hires were working without cleared background checks. During the same period, deportation officers were handed ELITE — Enhanced Lead Identification and Targeting — a Palantir-built tool that 404 Media first described as an app for finding neighborhoods to raid. The user guide shows officers can pull addresses, criminal histories, immigration records, and court rulings, and can see confidence scores generated from the underlying data.
Consider what that means at the level of a single hire.
Imagine an officer — call him a 27-year-old recent hire assigned to a field office in Phoenix — sitting down at a terminal in his first weeks on the job. His background check has not been completed. His training has been compressed. And in front of him is a tool that lets him query the addresses, criminal histories, and court records of people across the country. The gap between what he has been vetted for and what he can now see is the entire story.
An anonymous ICE deportation officer put it more bluntly to The Intercept. Some of these guys should’ve never been hired, let alone have access to the information on ELITE.
The second piece of the sequence is the missing paperwork. The E-Government Act of 2002 requires federal agencies to conduct and publish a privacy impact assessment for any new or substantially changed technology that collects, maintains, or disseminates personally identifiable information. The DHS’s own AI use-case catalog notes that ELITE handles PII. No PIA has been published for it.
Mary Ellen Callahan, the former chief privacy officer of DHS, told The Intercept that the department’s privacy office published one PIA in the past year. During her three-and-a-half years in the role, she said, the office published 228. Whatever else one thinks of the number, the change in throughput is not subtle.
The point of a privacy impact assessment is not bureaucratic hygiene. It is the mechanism by which the public learns what data an agency is collecting, how long it keeps that data, who inside the government can query it, and what remedy exists for a person whose record is wrong. Clare Garvie, a privacy lawyer at NYU Law School’s Policing Project, described PIAs to The Intercept as an essential component of transparency and privacy governance — the way the public understands what information is collected about them and what recourse exists if that information is misused.
Without the PIA, the public does not know those things. Neither, in many cases, do the officers using the tool.
The third piece is the tool itself.
ELITE is not an oracle. It is a query interface that assembles information from twelve government sources into a single view and attaches confidence scores to potential targets. Confidence scores in adjacent law-enforcement systems encode whatever biases sit in the underlying data. The 2019 PIA for ICE’s Enforcement Integrated Database, one of the systems ELITE draws from, runs to dozens of pages precisely because the risks are not obvious from outside.
Consider the compounding effect. A newer officer, less training, a tool that outputs a number that looks authoritative, and a legal framework in which the assessment that would have flagged the failure modes was never written. Each element is defensible on its own. Together they describe an accountability gap.
Consider what this looks like from outside the agency.
A lawful permanent resident — a 54-year-old cook in Queens, say, with a decades-old misdemeanor and a common name — appears in ELITE with a confidence score. An officer, three months into the job, acts on it. The person’s address, criminal history, and prior contacts with law enforcement are all visible in the same view. There is no PIA describing the retention window on that query, no published guidance on how the score was generated, and no public documentation of the internal review that would catch a mismatched identity.
Or consider a US citizen — a 31-year-old warehouse worker in suburban Chicago — whose file, pulled from a state database with known error rates, shows an outdated immigration flag. The officer querying the record has no way of knowing which of the twelve sources contributed which field, because the merged view does not surface provenance. The user guide does not appear to require it.
These are not exotic scenarios. They are the ordinary failure modes of merged government databases, which is why the PIA process exists in the first place.
Callahan, speaking to The Intercept, said no one should have been hired by DHS before background reviews were completed, and that it had not happened in previous administrations. That is a claim about administrative norms rather than statute, and it is worth noting the qualifier. The hiring surge was authorized. The compression of training was a policy choice. The absence of a published PIA is a legal question that has not been tested in court.
What has changed is the speed at which each shortcut was taken and the fact that they were taken simultaneously.
The pattern is familiar from other domains where oversight lags deployment: legislatures narrow the transparency features of a proposal while keeping the enforcement teeth. The Palantir–ICE story is a federal-scale version of the same trade: the surveillance apparatus scales up, the accountability apparatus scales down, and the gap between them is where the abuses live.
Palantir’s role here is worth stating precisely. The company builds the tool; the agency decides how to deploy it. ELITE is not autonomous. It does not make arrests. It surfaces information and ranks it. The civil-liberties question is not whether Palantir’s software is uniquely dangerous — it is whether the government agency using it has met the statutory obligations that apply to any technology handling PII at this scale. On the available reporting, it has not.
There is a temptation, in stories like this, to reach for either the sensational frame — a dystopian AI targeting immigrants — or the reassuring one — this is just a database with a nicer interface. Both miss the actual mechanism.
The mechanism is that four ordinary decisions, each individually explainable, produced a system in which under-vetted personnel operate powerful surveillance software without the public documentation that would let anyone outside the agency understand what is happening. Hire faster. Train less. Deploy the tool. Skip the assessment.
The Intercept’s reporting describes five ICE officials across different regions confirming that deportation officers without completed background checks continue to work in their jurisdictions, some hired nearly a year ago. That is not a hypothetical about future risk. It is a description of the present.
The E-Government Act was written in 2002, before Palantir existed, before ELITE existed, before machine-generated confidence scores were embedded in the daily workflow of federal law enforcement. The law’s authors did not anticipate the specific tool. They anticipated exactly this shape of problem, which is why they wrote a statute that applies to any new technology handling PII, whatever it happens to be called.
The tool got built. The officers got hired. The assessment did not get written.
That is what accountability looks like when it fails quietly, one procedural step at a time.