AIFoPa-2026-0024 — It Looked for the Word Officer

United States Immigration and Customs Enforcement was, through the second half of 2025, under a mandate to add ten thousand officers to its force by the end of the year, supported by signing bonuses of fifty thousand dollars drawn on the appropriation made under the One Big Beautiful Bill. Applications arrived at a rate no established process had been built for. To sort them, the agency put an artificial intelligence tool on the résumés. The tool’s function was narrow and consequential: to determine which applicants were already law enforcement officers, and therefore belonged in the agency’s LEO programme — four weeks of training, online — and which were not, and therefore belonged at the Federal Law Enforcement Training Center in Georgia for eight weeks in person, covering immigration law, the handling of a firearm, and a physical fitness assessment. The eight-week course is itself the shortened form; it ran to twenty weeks until it was reduced, in the agency’s published account, to cut redundancy and to incorporate advances in technology.

The tool looked for the word officer. It found the word officer in a number of places, and the Bureau, which has spent this year cataloguing failures of nuance, wishes to record that the failure here was not one of nuance but of the entire remaining category. A compliance officer was an officer. So, according to the two law enforcement officials who described the error to NBC News, was an applicant who had written on his résumé that he aspired to become an ICE officer — a statement of intention which the instrument read as a statement of history. The majority of new applicants, those officials said, were categorised as serving law enforcement officers. Many had never served in any police force, local or federal. One of them put the finding in a sentence the Bureau considers exemplary in its economy: “They were using AI to scan résumés and found out a bunch of the people who were LEOs weren’t LEOs.”

The error was identified in mid-autumn, rather more than a month into the recruitment surge, at which point the agency introduced a corrective measure the Bureau records with a certain professional sympathy: it began reading the résumés by hand. Recruits who had been misrouted were recalled to Georgia for the full course. The officials, who were not authorised to speak publicly, told NBC News that they could not establish how many people had been improperly trained, nor how many had been sent out before the mistake was found.

The Department of Homeland Security is on record and its account differs. A departmental spokesperson described the matter as a “technological snag” which had been quickly rectified, put the number affected at about two hundred hires, and stated that all of them subsequently reported to the Federal Law Enforcement Training Center for full training. No candidate, the spokesperson said, was placed on enforcement duties without “appropriate training and credentials,” and “the vast majority of new officers brought on during the hiring surge are experienced law enforcement officers who have already successfully completed a law enforcement academy.” The Bureau sets the departmental figure and the officials’ account down beside one another, in that order, and offers no reconciliation between them, having none. It notes only that one of the two parties has a number and that the number is not the same as no number.

There is a further finding, and it concerns arithmetic rather than training. One of the officials observed that the agency met its target on paper — ten thousand hires, recorded, by the end of 2025 — but that recalling the misidentified recruits to be trained properly meant that ten thousand officers were not in fact added to the street in that year. The Bureau has encountered this shape before in other contexts and has never found a satisfactory place to file it. A count was produced. The count was accurate as to the documents. The documents had been classified by a system which was measuring something adjacent to, but not the same as, the thing the count was understood to be counting.

The Bureau declines, finally, the temptation to record this as a malfunction, because on the evidence available the instrument did not malfunction. It was asked to find a word and it found the word, thoroughly, at scale, and without a single instance of the hesitation that the same task performed by a person would have produced somewhere around the fourth compliance officer. What was delegated to it was not the search. What was delegated to it was the judgement of whether a stranger’s prior life had prepared him for four weeks of instruction or eight, and that judgement was rendered on the presence of a substring. The Bureau files the incident under the classification it has had to create for it, and observes that the classification will not be lonely for long.