ICE Has Copied Tens of Millions of Voter Records. The Paper Trail Can’t Say Why.
Internal ICE records released on October 7 after Freedom of Information Act litigation show that Homeland Security Investigations had ingested roughly 50.6 million voter records from the Department of Justice by late May, with tens of millions more from other sources. The same documents show a hunt for voter fraud that has produced far more case numbers than prosecutions. The records, reported by Biometric Update on October 9, raise a question the agency has not publicly answered: what, exactly, is a federal immigration enforcement arm doing with the country's voter rolls?
What the documents say
The records were obtained by the group Democracy Forward. According to Biometric Update's reading of them, HSI had processed about 50.6 million voter records supplied by DOJ, prioritizing Texas, Pennsylvania, California and Georgia. A further 48.4 million records came from independent sources attributed to DHS Deputy Assistant Secretary Heather Honey. The files were routed to HSI's Innovation Lab, known as iLab, for "enrichment" and the identification of "high-value targets."
A separate workstream started from USCIS data. As of June 1, it had 26,563 prioritized records. HSI had worked through 2,290 of them, sent 1,586 leads to 24 field offices, and left 24,273 pending. In August, ICE also issued a Request for Information seeking a contractor to acquire nationwide voter-registration files, possibly with annual updates. A DHS forecast put the value at $2 million to $5 million, with an award expected by the end of September. No award has been publicly identified.
A fraud hunt with thin results
The yield is the most striking part. HSI reported more than 2,000 investigative leads across 49 field offices in 2025, and 68 cases presented for prosecution. As of May 25, 2026, it had opened 418 voter-fraud cases. Of those, 318 had not been referred for prosecution. The records do not give conviction figures, and they do not show which cases, if any, came from analyzing the voter files.
Nevada offers a sense of the error rate. DHS initially flagged nearly 16,000 potential noncitizens on that state's rolls, then narrowed the list to 185 high-confidence cases. Nevada officials reported that all 185 were citizens. The documents do not say whether iLab was involved in that particular match.
None of this proves bad faith. Data matching across messy administrative systems produces false positives, and agencies are expected to winnow them. But a pipeline that begins with nearly 100 million records and ends, in at least one documented instance, with zero actual noncitizens is not obviously a fraud-detection system. It is, at minimum, a very large collection effort looking for a purpose.
The nonprofits
The records also describe investigations into more than 120 organizations. In Pennsylvania, HSI reviewed 82 nonprofits and flagged two for further examination. In New Jersey, it identified about 41 voter-registration groups using commercial nonprofit databases, plus three more through open-source research, and reported no derogatory information on them. It reviewed 20 of 56 funds associated with the Amalgamated Foundation's Assets Under Movement program and reported no suspicious financial activity. Leads on Columbus Votes in Ohio were exhausted with no further action expected.
HSI also considered deploying undercover agents against voter-registration groups in New Jersey and Pennsylvania. The records do not show that this happened. They do show it was on the table.
Voter-registration drives are lawful, routine civic activity. A federal agency whose mission is immigration and transnational crime compiling reviews of those groups, and weighing undercover operations against some of them, is a notable expansion of whose conduct counts as an investigative matter.
The Palantir question
A July 6 HSI report lists Missouri, Michigan, New Mexico, Connecticut, Georgia and Kentucky as states whose enriched voter records were awaiting "Palantir entry into ELITE," a Palantir-built analytics tool. After a June federal injunction, HSI flagged Michigan and New Mexico for further legal review.
Palantir told WIRED that "voter roll data has never been integrated into ELITE." A DHS official said they could not "say for sure" whether the data entered the system. Those two statements sit uneasily together, and the records themselves do not establish that any transfer occurred. The distinction matters, and so does the fact that the agency holding the data cannot confirm where it went.
The picture around it is not reassuring. An August inspector general report found HSI spent more than $150 million developing an earlier analytics platform, RAVEn, before beginning to decommission it, with agents in six of eleven field offices examined having largely stopped using it. Replacement was estimated at about $39 million. The government has a track record of paying heavily for data tools it then struggles to use.
Separately, a proposed class action filed in Maine alleges that DHS agents logged ICE observers and protesters in Palantir's Investigative Case Management system. DHS calls that suit meritless and says the incidents, if they happened, were isolated. It is a different case, but it shares a theme: the question of who gets entered into federal databases, and on what grounds.
What stays unanswered
The documents leave large gaps. They do not say what legal authority governs ICE's retention of state voter data, how long it will be kept, who can query it, or what happens to a citizen who is wrongly flagged. They do not say whether the 2026 acquisition plans proceeded.
What they do say is that the voter rolls of most of the country have been copied into a federal enforcement apparatus, that the stated purpose has so far produced a short list of prosecutions and a long list of dead ends, and that the agency itself is unsure of some of the data's path. A voter file is, among other things, a list of where people live. Whether that is the intended use or merely an available one is a question the paperwork does not resolve.
