Domain Atlas / Hiring & employment screening AI
Checkr gig-economy background screening
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 3 assumed · 19 published baseline.
Checkr, Inc. is a consumer reporting agency founded in 2014 whose automated background-check platform sits one company upstream of the gig platforms that buy from it: it retrieves criminal items from county, state and national sources, matches them to a named applicant, normalises the charge data, applies each platform customer's own eligibility matrix, and furnishes a report, after which the platform executes its own access decision. Every scale, speed and automation figure for it is the vendor's own marketing, fetched from its product pages on 28 August 2026 and independently audited by nobody: 90 percent of gig economy background checks run on Checkr; 140,000-plus businesses already run on Checkr; 260 million-plus real identities mapped, covering 96 percent of US adults; instant criminal checks that screen users in under a second; 89 percent of national criminal record checks complete within one hour; Checkr AI normalises criminal charge data; automated adjudication tools reduce manual workflows by up to 80 percent; and continuous checks are offered as a standing-monitoring product. Those claims are self-serving in both directions — they are capability marketing and they are also the clearest available statement of how unmanned the generation path is. Lyft, Hyer and GigSmart are named as gig clients on the vendor's own pages; the Uber relationship is documented litigation-side, where a defence-bar analysis records Uber moving its New York City driver checks to Checkr in mid-2017.[2]
What happened
A person applies to drive or to deliver. The platform sends an identity — name, date of birth, social security number, address history, licence — to Checkr, Inc., a consumer reporting agency founded in 2014 whose whole business is this one transaction at volume. Checkr retrieves criminal items from county, state and national sources, decides which of them belong to this particular person, normalises the charge data into its own categories, applies the platform customer's eligibility matrix, and furnishes a report. Then it stops. The deactivation, the denial, the access decision: those happen at the platform, in a different company, from the file Checkr made.
The scale is the vendor's own account of itself, and every figure in it is marketing fetched from its product pages on 28 August 2026. Ninety percent of gig-economy background checks run on Checkr. More than 140,000 businesses run on Checkr. 260 million real identities mapped, covering 96 percent of United States adults. Instant criminal checks that screen users in under a second; 89 percent of national criminal record checks complete within one hour; automated adjudication tools that reduce manual workflows by up to 80 percent; continuous checks as a standing product. Read those numbers twice, because they are self-serving in both directions. They are capability claims, and they are also the clearest available statement of how little human attention any individual report receives.
On or around 28 August 2020, one day after Checkr furnished Uber his report, Job Golightly was deactivated. He is a Black driver from the Bronx who had driven for Uber since 2014, averaging about $1,500 a week on the platform. His entire criminal history was a 2013 speeding ticket from Virginia — 22 miles per hour over the limit, which is a misdemeanour under Virginia law and a civil infraction under New York law. He received no pre-adverse-action notice, no copy of the report, no individualised analysis under New York City's Fair Chance Act, and no three-business-day window in which to respond. He learned the reason months later.
The class complaint that followed, filed in the Southern District of New York on 8 April 2021, is worth reading for what it does and does not allege. Four of its five counts are against Uber: the Fair Chance Act, disparate racial impact, and the federal and New York consumer reporting statutes. The single count against Checkr is narrower and more structural — that it furnished employment-purpose reports without obtaining the compliance certification 15 U.S.C. 1681b(b)(1) requires from the user before a report may be furnished at all. No misreporting by Checkr is pleaded. The record was accurate; the severity translated badly across a state line, and the process that was supposed to catch that did not run. The complaint estimated classes of at least 1,000 and at least 300 against a backdrop of approximately 80,000 licensed rideshare vehicles in New York City and an Uber policy of re-checking current drivers at least every two years. It also recorded, citing research from The New School's Center for New York City Affairs, who the population being screened is: 87 percent of New York City transportation independent contractors are persons of colour, 81 percent lack a college degree, and 90 percent are foreign-born.
On 21 December 2022 Judge Lewis J. Liman granted the motion to compel individual arbitration and stayed the claims. The class vehicle against the platform-and-agency pair has been parked since, and no public resolution has been located as of August 2026. The order itself could not be read from the environment that verified this record — the primary docket hosts returned errors — so its specific treatment of the claims against Checkr is not asserted here in either direction.
Alongside the litigation, a regulator was writing down what the statutes require of exactly this kind of pipeline. On 4 November 2021 the Consumer Financial Protection Bureau issued an advisory opinion holding that name-only matching is not a reasonable procedure under the Fair Credit Reporting Act's maximum-possible-accuracy duty. Its language was about the industry rather than about any company: when background screening companies and their algorithms carelessly assign a false identity to applicants for jobs and housing, they are breaking the law; the risk of mistaken identities from name-only matching is likely to be greater among Hispanic, Black and Asian communities because there is less surname diversity in those populations; and because of the sheer scale of background screening activity, even ostensibly low error rates can harm significant numbers of consumers. On 11 January 2024 two further opinions set out the store-hygiene duties: prevent the reporting of expunged, sealed or legally restricted records, ensure dispositions accompany any reported arrest or charge, eliminate duplicates, honour per-item obsolescence windows such as the seven-year bar on non-conviction arrests — and disclose, in a consumer's file, both the originating sources and any intermediary or vendor sources.
On 12 May 2025 all four of those Fair Credit Reporting Act advisory opinions were withdrawn together, inside a mass withdrawal of sixty-seven guidance documents. Name-Only Matching, Permissible Purposes, Background Screening, File Disclosure. The statutes they interpreted were untouched: the accuracy duty, the reinvestigation duty, the public-record currency-or-notice duty and the certification bar all remain in force and privately enforceable, and courts may still reach the same results. What was removed is the regulator's stated reading of them. That is the whole of what happened, and it is neither a repeal nor nothing.
The pipeline kept running. On 26 January 2026 a putative nationwide class was filed in the Southern District of Florida — Davis v. Checkr, Inc., No. 0:26-cv-60088 — alleging under 15 U.S.C. 1681e(b) that a report included criminal case records plainly unassociated with the consumer, and that the company knowingly and willfully maintains deficient procedures because reporting more information is more profitable. The class is defined as consumers who within two years received a Checkr consumer report that incorrectly included criminal case records belonging to someone else. It is pending, pre-certification, and its allegations have not been tested. Note that it is a different mechanism from the 2021 case: mis-association, not a missing certification. The two are not the same claim and this file does not merge them.
Two comparisons close the record, and neither is about Checkr. In October 2023 the Federal Trade Commission and the Bureau together took $15 million from TransUnion's rental-screening arm over duplicate eviction entries, misreported outcomes, sealed records not removed and undisclosed third-party sources. In September 2023 the Commission took $5.8 million from TruthFinder and Instant Checkmate over background reports marketed as most accurate and assembled from sources that disclaimed accuracy, with a Flag as Inaccurate button that triggered no investigation. Those are the joint-regulator template for automated-screening accuracy enforcement in neighbouring markets. No equivalent public action against Checkr was located. And the one completed money resolution anywhere in this record landed on a platform rather than on the agency: $3.35 million in the Kings County Supreme Court over Uber Eats couriers denied in New York City between 24 October 2015 and 28 July 2021, with a final approval hearing on 8 February 2024, in a case where Checkr was not a defendant.
The sociotechnical reading
Start with what is unusual about the position of the failing component. In most deployments in this atlas the system that errs and the organisation that acts on the error are the same organisation. Here they are two, and the second one is many. A consumer reporting agency manufactures a record; platforms consume it and take their own actions from it. So the harm this deployment can do is misattribution rather than denial, and its geometry is not a pipeline but a fan.
That geometry has a specific consequence the sources make explicit. A contaminated person-record does not produce one wrong decision. It is furnished to the platform that requested it, it recurs on every later check by any platform sharing the vendor — and Uber's own policy re-checks incumbent drivers at least every two years — and a continuous-monitoring product re-furnishes on a new record event, so a person who applied once and has been working since can be re-screened without applying for anything. The correction channel, by contrast, exists at exactly one place: the single upstream store. That asymmetry is the deployment, and it is why a market-share claim is a governance fact here rather than a marketing one. If ninety percent of the checks in a labour market run through one matching rule, a defect in that rule does not average out across competitors. It repeats.
Now the part that makes this board different from the screening boards beside it, and it is a strength rather than a weakness. The check here is compulsory. Federal law does not merely permit a correction; it requires one. On a consumer's dispute the reporting agency must reinvestigate the item and correct the file, within thirty days and forty-five with an extension, and because the file sits upstream of everyone, one dispute that succeeds fixes the record for every platform at once. That is a wider correction pathway than a screening deployment usually has. It is also lodged behind two structural conditions that the record documents directly. It opens on the subject's initiative, and the subject in the pleaded case was told nothing: the report was furnished and access ended the following day, with no notice, no process and no communication. And its clock and the deployment's clock were written by different people. A loop that is legally compelled, correctly specified and structurally outrun is a different object from a loop that is merely discretionary, and reading it as simply absent misses what is actually wrong.
The certification gate is the same shape at a smaller scale. The statute bars an agency from furnishing an employment-purpose report unless the user certifies that it will comply with the notice and pre-adverse-action duties — the vendor-side gate that was supposed to verify the platform-side process existed. It is a real precondition and it is bounded by construction: a written representation about somebody else's process, obtained once for a relationship rather than read against each report. The 2021 complaint's only count against the agency is that this gate was skipped. That allegation has never been adjudicated, and nothing in this reading rests on it being true.
The third dynamic is about the governance surface rather than the system. Between November 2021 and January 2024 a regulator built the interpretive layer for this deployment class with unusual specificity: how to match, what to suppress, what must accompany a charge, what to deduplicate, what to disclose. On 12 May 2025 all four documents were withdrawn together while the deployment continued unchanged. This is worth stating precisely because the imprecise version is wrong in both directions. The duties did not end; they are statutory and privately enforceable, and a court may still read them exactly as the withdrawn opinions did. But the specificity did end, and specificity was the thing this deployment class most obviously lacked. What the record shows is a governance surface contracting on its own schedule, independent of anything the system did.
Two boundaries hold across the whole reading. The screened worker is the subject of the report, which is a served-person position: no eligibility outcome, deactivation or employment consequence for any person is computed from the network drawn here, and the worker enters it as the signal that opens a dispute rather than as a class of operators. And the register is allegation throughout for everything specific to this vendor. What can be asserted as fact is what the complaints plead, attributed as pleadings; the procedural history; the advisory opinions and their withdrawal; the enforcement in neighbouring markets; and the platform-side settlement's terms. Everything else — an error rate, a regulator finding, a merits ruling, a documented multi-platform simultaneous deactivation — does not exist in this record, and the absence of a published error rate is itself one of the findings: at the scale the regulator described, the size of the rate is the entire question, and nobody has measured it.
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.