Industries / Finance and business services / Staffing and recruiting / Background checks
Accuracy and disputes
Getting each report right about the person it describes, and fixing it when that person shows it is wrong.
7
Milestones achieved, from 1970 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1970
Maximum possible accuracy
The FCRA requires a consumer reporting agency, which includes a background screening company, to follow "reasonable procedures to assure maximum possible accuracy" in every report (section 607(b)). Since a 1996 amendment the agency must reinvestigate a disputed item free of charge within 30 days, which can be extended by up to 15 days, and delete anything it finds inaccurate, incomplete or unverifiable (section 611); before that the law asked only for a reinvestigation within a reasonable time.
Credit
US Congress
2012
First screening firm charged
On 8 August 2012 HireRight Solutions agreed to pay a $2.6 million civil penalty after the FTC charged that it reported records that had been expunged, listed the same offense more than once and failed to reinvestigate disputes. The FTC called it the first time it had charged an employment background screening firm with violating the FCRA. Two months earlier Spokeo had paid $800,000 for selling people-search profiles to human resources departments as a screening tool without following the Act.
Credit
US Federal Trade Commission; US Department of Justice
2013
Most British disputes upheld
People in England and Wales who think their certificate is wrong can dispute it with the DBS. Its published counts show 4,496 disputes in 2008, of which 2,560 were upheld, and 2,761 in 2012, of which 1,658 were upheld, so more than half succeeded in both years. DBS now reports that 99.99% of the criminal and barring information that should have appeared on certificates did so in 2024–25, against a target of 99.98%, when it issued 7,272,217 certificates.
assets.publishing.service.gov.ukdata.gov.ukassets.publishing.service.gov.uk
Credit
Disclosure and Barring Service
2014
Loose matching draws penalties
On 9 April 2014 InfoTrack agreed to a $1 million penalty, all but $60,000 suspended because it could not pay, for reporting everyone with the same name on the National Sex Offender Registry as a "possible match", although many registry entries had no birth date before late 2011. In October 2018 RealPage paid $3 million, then the largest FTC penalty against a background screening company, for tenant screening that required an exact match on last name only. In December 2020 AppFolio paid $4.25 million for reporting criminal and eviction records bought from a vendor without checking them, which mixed people up and listed cases more than once.
Lessona name alone, or a vendor's data taken unchecked, does not meet the Act's test of reasonable procedures.
Credit
US Federal Trade Commission
2019
Errors counted, in part
NCLC's report Broken Records Redux (December 2019) said no industry-wide study of accuracy in criminal background checks had ever been done, and cited court records from a lawsuit against First Advantage: of 3.5 million reports it prepared from 2010 to 2013, 17,431 were disputed and 14,346 revised, 13,346 of them after people said a record belonged to someone else. In November 2022 the CFPB reported more than 16,000 complaints about wrong information in tenant screening reports, including complaints that the same wrong information showed up again on later reports. In February 2024 Sarah Lageson of Rutgers University and Robert Stewart of the University of Maryland compared state records with commercial checks for 101 people in New Jersey. At least one false positive, a record the state file did not support, appeared for 60% of participants in regulated checks and 50% in unregulated ones, and at least one false negative, a record the check missed, for 90% and 92%.
Lessondispute counts show only the errors that people found and challenged.
Credit
National Consumer Law Center (Ariel Nelson); CFPB; Sarah Lageson and Robert Stewart
2021
Errors must reach someone
In TransUnion v. Ramirez, decided 5–4 on 25 June 2021, the US Supreme Court ruled on a class of 8,185 people whose files carried false alerts that matched them by name to the US Treasury's sanctions list. Only the 1,853 whose reports had been sent to businesses had standing to sue for damages in federal court; the other 6,332 did not. On 25 June 2026 a federal appeals court held that a screening company had acted reasonably in relying on an official state court website, although the record it reported belonged to the plaintiff's twin brother.
Lessonan error is not automatically a violation; federal damages need harm, and mistakes in a court's own records can pass through.
Credit
US Supreme Court; US Court of Appeals for the Eighth Circuit
2025
Guidance issued, then withdrawn
In November 2021 the CFPB said that screening companies that match records to people by name alone break the accuracy rule, and in January 2024 it said the same of reporting one case several times, reporting sealed or expunged records, and leaving out a case's outcome. On 12 May 2025 it withdrew both opinions along with many other guidance documents, and on 15 May 2025 it withdrew a December 2024 proposal to treat data brokers as consumer reporting agencies. On 28 October 2025 it issued a rule saying that states cannot validly regulate which categories of information, such as arrest records, appear in consumer reports.
Lessonagency guidance can be undone within months, while the accuracy duty in the statute stays.
Credit
Consumer Financial Protection Bureau
Watch for
Sealed records on appeal
A federal district court in Virginia held that it is accurate and lawful under the FCRA for a screening company to report records that have been sealed or expunged. The case, Doe v. InformData, is on appeal to the Fourth Circuit, and NCLC filed a brief supporting the appeal on 9 July 2026.
Achieved whenThe Fourth Circuit issues a published opinion.
Claimed by
US District Court for the Eastern District of Virginia, 2026; disputed by NCLC, July 2026
AI claims without error rates
Checkr says it became the first background check provider certified to ISO/IEC 42001, a standard for managing AI systems, in October 2025, after an audit by A-LIGN, and that its national criminal database covered 76.5% of felonies by May 2026. It gives no figures for false matches or upheld disputes, and coverage is not accuracy.
Achieved whenAn outside auditor, regulator or court publishes measured false-match or upheld-dispute rates for its reports.
Claimed by
Checkr, October 2025 and August 2026
Open
01
Lawsuits sorted by error
A public table of federal lawsuits that accuse background screening companies of inaccurate reports, coded by the error alleged: the wrong person, a sealed or expunged record, a record older than the law allows, one case reported twice, a missing or wrong outcome, or an offense put in the wrong class. Publish the docket number, court, year, defendant and error type only, with no plaintiff names and no facts about anyone's record.
The briefCredit · unclaimed02
Made-up criminal record disputes
About 300 fictional cases, each with a line from a made-up background report, the made-up person's dispute and the made-up court record, and the correct outcome under section 611 (delete, correct, confirm as accurate, or treat as frivolous) set by two reviewers working apart. Include twins and father-and-son pairs with the same name, matches on name alone, one case listed twice, missing outcomes and records too old to report, plus controls where the person disputing is wrong. Mark cases about expunged records as legally contested while Doe v. InformData is pending.
The briefCredit · unclaimed
Open 01 · the brief
Lawsuits sorted by error
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of federal lawsuits that accuse background screening companies of inaccurate reports, coded by the error alleged: the wrong person, a sealed or expunged record, a record older than the law allows, one case reported twice, a missing or wrong outcome, or an offense put in the wrong class. Publish the docket number, court, year, defendant and error type only, with no plaintiff names and no facts about anyone's record.
Why it is still open
NCLC says no industry-wide accuracy study has been done, and its own reports describe example cases rather than counts. This research found no docket study that sorts these cases by type of error.
Built with what exists
CourtListener's REST API and its RECAP archive of federal court records; the federal nature-of-suit code 480 (consumer credit), under which FCRA cases are filed; a model that scores fixed-answer questions, such as Gemini Decision 1, to propose a code for each complaint (see the Note "A thousand questions, one pass"), with a second model pass and a person checking every code (see the Note "Seven more flagged").
First step
Write the codebook and code 50 RECAP dockets by hand.
Done looks like
At least 200 complaints filed from 2020 to 2025 coded, a published codebook, and two coders agreeing at a Cohen's kappa of 0.7 or more.
Open 02 · the brief
Made-up criminal record disputes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
About 300 fictional cases, each with a line from a made-up background report, the made-up person's dispute and the made-up court record, and the correct outcome under section 611 (delete, correct, confirm as accurate, or treat as frivolous) set by two reviewers working apart. Include twins and father-and-son pairs with the same name, matches on name alone, one case listed twice, missing outcomes and records too old to report, plus controls where the person disputing is wrong. Mark cases about expunged records as legally contested while Doe v. InformData is pending.
Why it is still open
This research found no public test set of criminal-record disputes. Complaints the CFPB analyzed in 2022 included wrong information that showed up again on later tenant reports, and the court figures in NCLC's 2019 report show that most corrections were for records that belonged to someone else.
Built with what exists
Python with Faker, which makes up names, dates and addresses; Splink, open-source record-linkage software, for a matching baseline; current AI models as the dispute handlers under test; Label Studio for the reviewers.
First step
Draft 30 cases and have a second reviewer label them without seeing the first's answers.
Done looks like
At least 300 cases published with the reviewers' agreement (Cohen's kappa of 0.8 or more), a scoring script, and baseline scores for at least two systems.