Domain Atlas / Public benefits & eligibility
ID.me identity verification as an unemployment eligibility gate
During the pandemic unemployment surge, a private facial-recognition identity check operated as a de facto eligibility gate for unemployment benefits in at least 25 U.S. state workforce agencies, with a live 'trusted referee' interview queue that House investigators documented averaging nearly 10 hours in North Dakota and over 4 hours in 14 of 21 states, versus about 6 minutes in New Jersey where an in-person option existed. Oregon's own one-month study (n=10,656 routed) recorded verification-completion differences by group -- for example 41.59% for African American and 34.48% for Spanish-language claimants versus 53.44% for White claimants -- but stated the study showed differences in completion and did not show causation, so these are a friction proxy, not a measured wrongful-denial rate. The U.S. Department of Labor does not collect or report the number of workers blocked for inability to verify identity, and where verification precedes filing those workers are not counted as denied claims at all, so the scale of any wrongful lockout is undocumented.[4]
What happened
During the pandemic unemployment surge, states rushed to place a remote identity check in front of benefit payment. The private vendor ID.me signed its first state contract with Florida in June 2020, and within about a year held contracts with at least 25 state unemployment agencies for nearly $45 million, serving identity verification for roughly 30 state governments and 10 federal agencies (the IRS alone spent about $86 million on licenses). To verify, a claimant needed a smartphone or computer with a camera to upload a government ID and a live video selfie, which facial-recognition software matched; roughly 10 to 15 percent of applicants were routed onward to a live video interview with a vendor "trusted referee." A Department of Labor Inspector General audit (March 31, 2023) found 24 of 53 state workforce agencies — 45 percent — used a facial-recognition identity contractor, drawing on 10 different vendors.
The verification became a bottleneck. House investigators documented average waits for the live interview reaching nearly 10 hours in North Dakota (April 2021) and nearly 6 hours in Washington, with 14 of 21 states averaging over 4 hours — while New Jersey, the only state then offering an in-person alternative, averaged about 6 minutes. The company had removed appointment scheduling because it was judged to be "hindering efficiency," and told the IRS that waits were "about 2 hours" even as its own data showed 4-plus-hour averages across most states. The vendor's CEO publicly claimed as much as $400 billion — "as much as 50 percent" — was lost to pandemic unemployment fraud, roughly ten times the DOL Inspector General's $45.7 billion estimate of potential fraud; the House concluded the company had made "baseless claims" to increase demand for its services. The company initially represented that it used only one-to-one facial matching, then acknowledged it also conducted one-to-many matching against a database — a technique it had earlier called "more complex and problematic." The Inspector General's review of the 24 states' contracts found 18 (75 percent) did not specify whether one-to-one or one-to-many matching would be used, 15 (63 percent) did not address data-storage requirements, and 13 (54 percent) did not address destruction of the collected biometric data; even so, 22 of 24 agencies (92 percent) reported the technology had reduced improper payments.
Whether the gate turned away the wrong people is genuinely hard to see, and that is the point. Oregon's own study of one month of regular-UI claims (n=10,656 routed) found 49.37 percent had completed verification by a November 2021 snapshot and 71.06 percent eventually — implying roughly 29 percent never completed. Completion varied sharply by group: White claimants 53.44 percent versus African American 41.59 percent and Native Hawaiian or Pacific Islander 42.05 percent; English-language filers 49.65 percent versus Spanish 34.48 percent; claimants 20 and under 28.49 percent; telephone filers 38.25 percent versus internet filers 52.10 percent; the lowest weekly-benefit band 39.52 percent versus the highest 60.32 percent. Oregon stressed the study showed differences in completion and "did not show causation": its outreach to 130 non-completers found some faced a technology barrier or found the process confusing, but others had simply returned to work (22 percent) or had not finished by the snapshot. Oregon had already paused automated routing of all regular-UI claims to the vendor on October 29, 2021 specifically to study disparate impacts, then built mitigations (multilingual robocalls, trusted-referee outreach, in-person WorkSource help, loaner phones) before moving toward alternatives. The National Employment Law Project documented the structural blind spot underneath all of this: the U.S. Department of Labor does not collect or report the number of workers denied unemployment benefits for inability to verify their identity, and where verification happens before a claim can even be filed, blocked workers are not counted as denied claims at all.
The controls that arrived came from outside and after the fact. Two House investigations reported findings in 2022; the Inspector General audit landed in 2023 with three recommendations the Employment and Training Administration accepted. On May 2, 2023, the National Center for Law and Economic Justice, NYLAG, and Make the Road New York filed a Title VI federal civil-rights complaint against the New York State Department of Labor, alleging that the identity vendor combined with untranslated notices caused delays and wrongful denials for limited-English claimants — the named complainants included one man forced into a homeless shelter and a woman whose language barriers contributed to an erroneous willful-fraud finding and nearly $20,000 in fines (a national-origin and language-access claim in which the identity gate was one contributing element). After bipartisan backlash, the IRS and Treasury dropped the mandatory facial-recognition requirement in February 2022 and the vendor made facial recognition optional across government agencies; Massachusetts became the first state to follow, though there the technology had already been optional. In July 2023 the Department of Labor launched a public alternative — GSA's Login.gov plus U.S. Postal Service in-person verification — and a 2024 GAO review found federal agencies had leaned heavily on commercial identity vendors while Login.gov only recently added NIST-compliant one-to-one face matching. The pattern persists: the vendor and similar services still gate unemployment claims in a large share of states into 2025, and in 2026 the IRS proposed allowing the vendor to retain taxpayer biometric data for up to 36 months after an account is deleted. (State counts vary by source and definition — at least 25 UI agencies, roughly 30 governments, 27 agencies contracting at some point — because they measure different things; and per project rules no specific algorithm or model is named. NIST's finding that facial-recognition false-negative rates ranged from below 0.5 percent to above 10 percent, often higher for women and younger people, describes the technology class, not a named system.)
The sociotechnical reading
Almost every algorithmic-benefits failure in this Atlas produces a record you can point at: MiDAS issued tens of thousands of false fraud determinations, Robodebt raised hundreds of thousands of unlawful debts, Indiana denied benefits for "failure to cooperate." Those are visible adverse actions, and the governance question they pose is who could reverse them, and how fast. The identity gate poses a stranger question, because it produces no decision at all. It is not scoring anyone and not adjudicating eligibility — it is a fraud-control checkpoint placed in front of the claim, and the people it turns away simply never appear. A claimant who abandons an hours-long verification queue has not been denied; there is no determination to appeal, no reason code, no letter. The system's most consequential output is an absence. This is the case's first and sharpest lesson: friction can be a denial with no denial to contest, and when the checkpoint sits before the claim is even filed, the adverse action is not merely unappealable — it is uncounted. The National Employment Law Project's finding that the federal government collects no data on workers blocked by verification failure is not a paperwork gap; it means the harm is invisible by construction. You cannot govern, or even see, a denial you never recorded.
That invisibility is held in place by an asymmetry the map makes legible. Operators see a real, measurable benefit — 92 percent of surveyed agencies reported the technology reduced improper payments — while the cost, wrongful lockout of eligible claimants, stays unmeasured on the other side of the ledger, falling hardest on low-broadband, non-English-speaking, younger, and lowest-benefit filers (Oregon's completion gaps are a friction proxy, not a measured error rate, but the direction is unmistakable). A private vendor's incentive to inflate the fraud it exists to stop — a CEO's $400 billion against the Inspector General's $45.7 billion — tilts the same scale. And because which claimants get routed to the gate is driven by upstream fraud flags, the checkpoint lands twice on the people least able to clear it, including identity-theft victims flagged for the very crime committed against them. So the productive governance moves here are not accuracy fixes. The one control that actually fired was Oregon's pre-authorized pause — an internal actor exercising the authority to stop and study before litigation forced it, the domain's cheapest control. The rest are about turning the invisible visible: reconcile who was routed against who ever completed so non-completers surface as reviewable denials; assign someone to review the lockouts nobody appeals (this domain's defining governance question); fund the starved human fallback that New Jersey's six-minute in-person option shows is the real bottleneck; and, because the gate runs on biometrics the vendor searches one-to-many and retains for years, govern it at the contract — the matching type, the storage, the destruction schedule that most audited contracts never specified. The Atlas's throughput-honesty and provenance patterns trace straight back to a case whose central failure is that it kept no honest count of whom it excluded.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.