The matter widely described as the first artificial intelligence hiring discrimination settlement was not pleaded or settled as one. In EEOC v. iTutorGroup, Inc., the agency alleged a hard-coded instruction: tutoring application software programmed to reject female applicants at one age and male applicants at another. The term artificial intelligence appears nowhere in the amended complaint, nowhere in the proposed consent decree, and nowhere in the EEOC’s announcement of the settlement. The absence is not a curiosity. It means the case most often placed at the origin of this field does not establish how discrimination by a learned model would be proved.

Four records, verified against primary documents on August 11, 2026 and refreshed where time-sensitive on August 21, describe one problem from four positions. What they share is a single question: can the person who was turned down point to the rule that turned them down? The pleading in iTutorGroup reduces the alleged mechanism to a single explicit instruction. The complaint in Harper v. Sirius XM Radio, LLC identifies alleged proxies but no model, no score, no product version, and no decision rule. The question in Mobley v. Workday, Inc. is who answers when the tool belongs to a vendor. And New York City’s Local Law 144 of 2021 answers a different question, requiring in advance group-level selection evidence of the kind Harper’s complaint does not contain.

These records suggest that proof gets harder as the discriminating mechanism becomes less legible, and that the regulatory answer moves the work from the courtroom to a compliance filing. One city, among those examined here, requires that filing. Whether it operates is the subject of a companion piece, When the rule is written down and nobody looks.

The rule fit in one sentence

The EEOC sued iTutorGroup, Inc., Tutor Group Limited, and Shanghai Ping’An Intelligent Education Technology Co., Ltd. in the Eastern District of New York on May 5, 2022, amended its complaint on August 3, and had a consent decree so-ordered on September 8, 2023, 16 months in all. The mechanism alleged is one instruction: the defendants “programmed their tutor application software to automatically reject female applicants age 55 or older and male applicants age 60 or older.”

On the EEOC’s account, the charging party’s experience reads like a controlled experiment. The agency alleged that Wendy Pincus was rejected when she applied using her real date of birth, then offered an interview the next day after submitting an otherwise identical application with a more recent one. More than 200 applicants were affected, the complaint alleges. The complaint states the mechanism without relying on an expert, a regression, or discovery into a model.

The pleading charged only the Age Discrimination in Employment Act, section 4, 29 U.S.C. 623(a) and (b). Title VII appears nowhere in it, although the alleged cutoffs differed by sex: 55 for women and 60 for men. The agency litigated the age line and left the sex line alone.

What the settlement did not establish

The settlement was $365,000, distributed to applicants rejected because of age. The filed proposed decree at ECF 24-1 provides for payment into a qualified settlement fund and distribution at the EEOC’s sole discretion among applicants rejected in March and April 2020; the operative revised decree at ECF 25-1, which the court adopted, was not available in RECAP. The decree was entered without findings of fact or conclusions of law, and the defendants denied liability. No court found that anyone discriminated. The EEOC alleged, and the parties resolved by consent.

Most of the non-monetary relief is conditional. The injunctions, training, monitoring, and reporting obligations trigger mainly on a resumption date, because the defendants had stopped hiring tutors in the United States and attested that they had no plans to resume. The proposed decree at ECF 24-1 runs five years, or three years from resumption, whichever is later, and the agency release confirms monitoring for at least five years, or longer if hiring resumes.

Even the cutoff is unsettled in the agency’s own paper. The complaint body and both press releases put the line at age 55 or older; the complaint’s preamble, the decree recital, and the fiscal 2023 performance report put it at over the age of 55.

The word that appears nowhere in the pleading or the proposed decree

The amended complaint at ECF 8 and the proposed consent decree at ECF 24-1 were read in full, and neither uses the term AI. The EEOC’s settlement announcement contains no instance of “AI,” “artificial intelligence,” “algorithm,” or “algorithmic.” The fiscal 2023 performance report groups the matter with one other as a case “with technology at the forefront,” a weaker claim than the one built on it since. The first-AI-settlement label traces to law firm alerts and trade press rather than to the agency.

What the agency said about technology, it said at filing. Chair Burrows said in the May 2022 announcement that “Even when technology automates the discrimination, the employer is still responsible,” and called the case an example of why the EEOC had recently launched its Artificial Intelligence and Algorithmic Fairness Initiative. That is a statement about responsibility, not about what the software was. The alleged sex-differentiated age cutoffs describe conventional facial discrimination implemented in software, and proving that tells a plaintiff little about proving a learned model.

Harper has no rule to point at

Harper filed in the Eastern District of Michigan on August 4, 2025, before Judge Terrence G. Berg, No. 2:25-cv-12403. The docket was active and not terminated when refreshed on August 21, 2026, carrying 24 entries through March 12, 2026. It pleads Title VII disparate treatment and disparate impact, 42 U.S.C. 2000e-2(a) and 2000e-2(k), and a claim under 42 U.S.C. 1981, on behalf of a Rule 23 class for which no certification motion has been filed.

Those three counts are not equally hospitable to the theory the complaint pleads. Section 1981 reaches only purposeful discrimination, General Building Contractors Assn., Inc. v. Pennsylvania, 458 U.S. 375, 391 (1982), and a plaintiff must plead and prove that but for race he would not have suffered the loss, Comcast Corp. v. National Assn. of African American-Owned Media, 589 U.S. 327 (2020). A proxy theory offered purely as an argument about effects has no route through that count. Deliberate use of a proxy would be a different matter, and Harper pleads disparate treatment as well, so it is the effects theory specifically that has one vehicle: the Title VII impact count.

The complaint alleges that Sirius XM screens applicants using iCIMS, a third-party vendor supplying an applicant tracking system and associated AI and machine-learning tools. No model, score, or product version is identified. iCIMS is not a party to the case, the allegations about its product are Harper’s, and the only defendant denies using that product at all. The mechanism Harper alleges is proxy screening: that the system evaluates data correlated with race, including educational institutions, employment history, and zip codes.

The statistical support is one person. Harper alleges approximately 150 applications since November 2023 and 149 rejections, a 99.3 percent rejection rate, with one interview offered. The complaint contradicts itself at one paragraph by describing 150 rejections, and the answer admits only that he submitted more than 150 applications. No comparative or group-level statistics appear anywhere in the pleading.

Sirius XM’s answer, filed January 6, 2026, denies using “any ATS or AI/ML tools in its hiring process in any manner whatsoever, whether from iCIMS or any other vendor,” and states that applications are reviewed manually in chronological order. The premise of the case is contested, down to the interview, which the complaint describes as a 30-minute session ending in rejection and the answer describes as an offer he never attended.

Neither an agency nor a court has endorsed any of it. The EEOC made no cause finding, stating in its May 6, 2025 determination that it “will not proceed further with its investigation and makes no determination about whether further investigation would establish violations of the statute.” A motion for judgment filed the same day as the answer remained pending at the August 21, 2026 docket refresh; it is not in the public RECAP archive, so its basis is not characterized here.

Part of the distance between these two complaints has nothing to do with software. iTutorGroup was brought by a federal agency acting on a charge filed with it. Harper is a private action by one applicant, filed after a notice stating that the agency made no determination whether further investigation would establish a violation. That difference in posture means the whole gap cannot be laid at the door of software opacity. It does not mean opacity plays no part in it.

One distinction has to be held before going further, because these records do not run on one doctrine. The comparison here is evidentiary, not doctrinal. iTutorGroup alleged a facial age cutoff under the ADEA, which is disparate treatment and calls for no impact analysis at all. Harper pleads Title VII treatment and impact together with a section 1981 count. The statutory exception discussed below is Title VII’s alone and reaches only Harper’s impact count. The particularity requirement itself applies under both Title VII and the ADEA, on different terms. iTutorGroup appears here only as a contrast in what was visible.

The remaining distance is the argument. Disparate impact analysis, which entered Title VII through Griggs v. Duke Power Co., 401 U.S. 424 (1971), assumes a plaintiff can identify a particular employment practice and show its effect on a protected group. Where the alleged practice is an explicit threshold, both halves are visible on the face of the pleading. Where it is a ranking function trained on data held by a vendor, the plaintiff must describe at the pleading stage what he cannot see, and prove later an effect he cannot yet measure, against a defendant free to deny the practice exists at all.

Congress anticipated part of this. Title VII’s disparate impact provision, 42 U.S.C. 2000e-2(k)(1)(A), carries a demonstration rule: the complaining party must show that each particular challenged practice causes the disparity, 2000e-2(k)(1)(B)(i), “except that if the complaining party can demonstrate to the court that the elements of a respondent’s decisionmaking process are not capable of separation for analysis, the decisionmaking process may be analyzed as one employment practice.”

A ranking function whose inputs are weighted inside a product nobody outside the vendor can inspect is the obvious candidate for that argument. It is not an automatic one. Being unable to inspect a process is not the same as its elements being incapable of separation for analysis, and the statute asks the second question rather than the first. Whether the exception reaches an automated employment decision tool is the doctrinal question this cluster presents. A search of reported decisions on August 11, 2026 found none applying the exception to an algorithmic, automated, or AI screening tool. The provision also does not travel between statutes: it arrived in the 1991 amendments, which expanded Title VII and left the ADEA alone, so an age plaintiff argues particularity under the older and stricter rule with no exception attached.

Pleading and proof are two different burdens, and the distinction matters. A plaintiff need not plead a prima facie case, Swierkiewicz v. Sorema N.A., 534 U.S. 506 (2002), because that is an evidentiary standard rather than a pleading requirement. What Swierkiewicz does not settle is how much factual specificity Harper’s complaint will need against a defendant who denies the practice exists at all, and no ruling in Harper has reached that question.

One objection to all of this deserves a direct answer, because it is the strongest one available. What the EEOC alleged exposed the rule was not that anybody read the code. It was a matched pair: the same application, one variable changed, opposite results. That is a black-box method. There is no reason in principle it cannot be run against a learned model.

What could make it harder against a ranking function is mechanical rather than legal. An output may be a position in an ordering rather than a rejection, a model need not return the same answer twice, and there may be no single threshold for a pair of applications to straddle. Whether any of that describes the system Harper alleges is not established. He did not plead a paired test, and the record does not say whether one was available to him.

None of the four records in this cluster contains a holding that the doctrine fails in that setting, and Harper is a complaint rather than a ruling. The narrower point is the one that is hard to argue with: the public record in Harper contained no model, no score, no version, and no group statistics at filing, and whether discovery will supply them is unresolved.

Mobley asks who answers for the tool

The Mobley docket in the Northern District of California, before Judge Rita F. Lin, was open and actively litigated with 489 entries through August 14, 2026 when refreshed on August 21. A primary-source pass in July 2026 confirmed two developments: the employer-as-agent theory survived dismissal in July 2024, and the court ordered the acquired HiredScore tool included. A direct read of docket entry 128 on August 11, 2026 established a third, that a nationwide ADEA collective was preliminarily certified for notice purposes on May 16, 2025. The August 14 entry is a text-only order on three sealing motions whose text was not available in RECAP, so no substantive conclusion is drawn from it. No court has found that Workday discriminated. Neither the denial of dismissal nor the preliminary certification is a ruling on the merits, and the second is stage-one certification under 29 U.S.C. 216(b), subject to a decertification motion after discovery.

The collective runs under the ADEA, which matters more than it looks. Age impact claims are cognizable, Smith v. City of Jackson, 544 U.S. 228 (2005), but they carry the older particularity requirement with no statutory exception, and an employer defending on reasonable factors other than age, 29 U.S.C. 623(f)(1), bears both the burden of production and the burden of persuasion, Meacham v. Knolls Atomic Power Laboratory, 554 U.S. 84 (2008). The preliminarily certified collective is an ADEA collective, and the ADEA standard is not the one that would govern Harper’s Title VII impact count.

Commentators call it the first major federal challenge to AI hiring discrimination; that ranking is theirs. The verified development concerns whether Workday can answer as an agent of the employers using its product. It does not resolve discriminatory effect, and it does not tell a plaintiff how to show what the tool did. A vendor added to the caption still leaves the plaintiff needing selection data, and the docket does not establish what such data exist or who holds them.

New York City required the evidence in advance

Local Law 144 added N.Y.C. Admin. Code sections 20-870 to 20-874, with rules at 6 RCNY sections 5-300 to 5-304. The law took effect on January 1, 2023; the rules took effect on May 6, 2023, and enforcement began on July 5, 2023, a distinction commonly collapsed into one date. It reaches automated employment decision tools: machine learning, statistical modelling, data analytics, or artificial intelligence issuing a simplified output used to substantially assist or replace discretionary decision making. The rules narrow “substantially assist” to sole reliance on the output, weighting it above any other criterion, or using it to overrule other factors, and no employer size threshold exists.

The obligation is an annual bias audit by an independent auditor, one with no employment relationship with the employer and no financial interest in it, conducted no more than one year before the tool is used. It measures selection rate and impact ratio by sex, by race and ethnicity, and by intersectional categories. A summary goes on the employment section of the employer’s website, and candidates receive notice at least 10 business days before the tool is used.

Read that against Harper’s complaint, and then against something much older. The measurement is not the new part, though the federal version of it has always been softer than it sounds. Since 1978 the Uniform Guidelines on Employee Selection Procedures have provided that every user of a selection procedure “should maintain and have available for inspection records or other information which will disclose the impact which its tests and other selection procedures have upon employment opportunities of persons by identifiable race, sex, or ethnic group,” 29 C.F.R. 1607.4(A). The four-fifths rule in the same section supplies the screen, and supplies it loosely: a selection rate below four-fifths of the highest group’s rate “will generally be regarded by the Federal enforcement agencies as evidence of adverse impact,” while smaller gaps may still count and larger ones may not. It is a screening convention, not a liability threshold.

What New York City added was not the measurement. It was publication, and an auditor with no employment relationship with the employer and no financial interest in it. Federal impact records surface on inspection or in discovery, which is to say to a regulator or to a plaintiff who already has a case. Local Law 144 makes an aggregate summary public before suit, which the federal inspection provision does not. Even so it is not a case. A published summary is an aggregate without the underlying data, and it does not separate the tool’s effect from the composition of the applicant pool. It gives a rejected applicant a threshold signal and a document to ask for in discovery, which is more than nothing and less than proof.

The law is narrower than its reputation. The statute requires the candidate notice to allow a request for an alternative process, and the rules decline to require one: “Nothing in this subchapter requires an employer or employment agency to provide an alternative selection process.” The statutory penalty is not the flat $500 to $1,500 band that circulates in compliance summaries. A first violation, together with additional violations the same day, is capped at $500, and only each subsequent violation carries the $500 to $1,500 range. Each day of use in violation of the audit and publication requirement is a separate violation; notice failures multiply per failure, not per day. The current OATH hearing schedule answers a different question: it lists $375 for a first violation, $1,350 for a second, and $1,500 for a third or later violation, with higher default amounts for the first two.

Set against the problem this piece opened with, that is the shape of an answer. An independent audit produces group-level evidence of the kind Harper’s complaint did not contain, and among the jurisdictions examined here one city requires it to be produced and published before anyone sues. That law reaches only employers hiring in New York City, and nothing establishes that either live case here involved a covered position. Whether any of it operates is a different question, taken up in When the rule is written down and nobody looks.

Practical implications

  1. Retain the artifacts that make a screening decision legible. iTutorGroup resolved in 16 months, and the complaint stated the alleged rule in one sentence. Where a vendor model produces a score, the equivalent evidence is the configuration in force, the thresholds actually applied, and selection data by group. If none of it is generated and kept, no party can produce it later, including the employer defending itself.

  2. Run the bias audit for the evidence it produces, not for the posting. Local Law 144 requires selection rates and impact ratios by sex, by race and ethnicity, and by intersectional categories. Those are numbers of the kind Harper’s complaint did not plead, and an employer holding them, able to explain them, is in a different position from one holding neither. The standing objection is that they are discoverable. The answer is that an employer who has never looked is the one who learns what they say from an opposing expert.

  3. Settle the vendor question in the contract, not after a complaint. Mobley is testing whether a screening vendor can answer as an agent of the employers using it, and that theory survived dismissal in July 2024. Agreements written while the question is open should allocate audit obligations, data access, and defense costs expressly, because a vendor that will not supply selection data can leave the employer unable to audit, explain, or defend. Check the governing law before relying on an indemnity: Colorado’s statute, which takes effect on January 1, 2027, voids specified terms that indemnify a developer or deployer against Colorado anti-discrimination liability for its own acts or omissions, subject to the statute’s exceptions.

  4. Check the primary document before repeating what is commonly said about it. Two claims commonly repeated in secondary summaries are not supported by this cluster’s own records: that the EEOC called iTutorGroup its first AI hiring settlement, and that Local Law 144 carries a flat $500 to $1,500 per-day penalty. Each dissolves on contact with the pleading and the statute.

Case citations and authorities

  • EEOC v. iTutorGroup, Inc., No. 1:22-cv-02565-PKC-PK (E.D.N.Y.) (amended complaint filed Aug. 3, 2022; consent decree so-ordered Sept. 8, 2023).
  • Harper v. Sirius XM Radio, LLC, No. 2:25-cv-12403 (E.D. Mich. filed Aug. 4, 2025).
  • Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal.).
  • N.Y.C. Admin. Code sections 20-870 to 20-874, added by Local Law No. 144 (N.Y.C. 2021); rules at 6 RCNY sections 5-300 to 5-304.
  • 42 U.S.C. 2000e-2(a), 2000e-2(k)(1)(A) and (k)(1)(B)(i); 42 U.S.C. 1981; 29 U.S.C. 623(a), (b) and (f)(1); 29 U.S.C. 216(b); 29 C.F.R. 1607.4(A) and (D).
  • Griggs v. Duke Power Co., 401 U.S. 424 (1971).
  • Smith v. City of Jackson, 544 U.S. 228 (2005).
  • Meacham v. Knolls Atomic Power Laboratory, 554 U.S. 84 (2008).
  • General Building Contractors Assn., Inc. v. Pennsylvania, 458 U.S. 375 (1982).
  • Comcast Corp. v. National Assn. of African American-Owned Media, 589 U.S. 327 (2020).
  • Swierkiewicz v. Sorema N.A., 534 U.S. 506 (2002).
  • Colo. Sess. Laws 2026, ch. 131 (S.B. 26-189).

Disclaimer: AI Lex Intelligence is published for informational purposes only. It does not constitute legal advice, and no attorney-client relationship is formed by reading or receiving this publication. Readers should consult qualified legal counsel about specific legal matters.

AI Lex Intelligence with Zola Valashiya. Independent analysis of artificial intelligence and the law.