Domain Atlas / Hiring & employment screening AI
iTutorGroup Tutor Application Screen
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 · 7 published baseline.
The U.S. Equal Employment Opportunity Commission alleged that iTutorGroup, Inc., Shanghai Ping'An Intelligent Education Technology Co., Ltd. and Tutor Group Limited — three integrated companies providing English-language tutoring to students in China through United-States-based tutors working fully remotely from their homes — had programmed their tutor application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older, rejecting more than 200 qualified United-States applicants because of their age. After conciliation failed the Commission sued under the Age Discrimination in Employment Act on 5 May 2022, No. 1:22-cv-02565 (E.D.N.Y.); then-Chair Charlotte Burrows framed the filing for the agency's algorithmic-enforcement agenda with the words 'Even when technology automates the discrimination, the employer is still responsible.' The parties filed a joint notice of settlement with a proposed consent decree on 9 August 2023 and the court approved the decree in September 2023, announced by the Commission on 11 September 2023. iTutorGroup pays $365,000 to be distributed among the more than 200 rejected applicants through a claims process, without admitting liability; per-claimant amounts were not made public. iTutorGroup denied the allegations and disputed that the tutors were employees at all, characterizing them as independent contractors, a question the settlement resolved without adjudication.[3]
What happened
iTutorGroup, Inc. and two affiliated companies — Shanghai Ping'An Intelligent Education Technology Co., Ltd. and Tutor Group Limited — provided English-language tutoring to students in China. The tutors were adults in the United States working fully remotely from their own homes, a pool that skews older and includes many retired teachers, which is why a rule written against age cut so deep. The three companies were treated throughout as one integrated employer, and the EEOC's coverage theory was about control: "Where, as alleged here, companies closely control the way fully remote workers perform their jobs, those workers are employees." iTutorGroup disputed that, characterizing the tutors as independent contractors. The question was resolved by settlement, not by a court.
What the Commission alleged is unusually simple to state. The companies had "programmed their tutor application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older". That is a rule someone wrote down. There is no model in this record, no score, no training data, no proxy variable and no drift: a birthdate field on an application form went into a threshold and a rejection came out, every time, on exactly the applicants the rule named. More than 200 qualified United-States applicants were rejected on it. The decree's reapplication invitations cover applicants rejected in March and April 2020, which fixes the documented window at two months.
Nobody inside the company found it. There was no review step between the rule and the rejection notice, so the staff who interview and select tutors only ever saw applicants the software had already passed; the people it turned away were told the outcome and not the reason, which left them nothing to complain about. What found it was an experiment run by one of them. The charging party applied with her real birthdate and was rejected immediately. About a day later she applied again with an application identical in every respect except a more recent birthdate, and was offered an interview. One changed field, no inside access, roughly twenty-four hours — the lowest-capacity audit anyone can perform on a hiring system, and in this record the only one ever performed.
From there the machinery was slow and conventional. A charge, a failed conciliation, and then on 5 May 2022 a suit in the Eastern District of New York under the Age Discrimination in Employment Act. Then-Chair Charlotte Burrows framed it for the agency's algorithmic-enforcement agenda: "Even when technology automates the discrimination, the employer is still responsible." The parties filed a joint notice of settlement with a proposed consent decree on 9 August 2023, and the court approved the decree in September 2023, announced by the Commission on 11 September. About three and a half years ran between the documented rejections and an enforceable remedy. iTutorGroup admitted no liability and denied the allegations.
The money is the smallest part of it: $365,000, distributed among the more than 200 rejected applicants through a claims process, which averages something near $1,800 if it were spread evenly — the per-claimant amounts were never made public. The rest of the decree is what a governance reader should study, because of where it aimed. It did not correct the rule. It enjoined requesting applicants' birth dates before an offer, which removes from the intake form the one field the rule computed on. Around that sit injunctions against hiring discrimination based on age or sex, a new anti-discrimination policy and an internal memo, multiple anti-discrimination trainings for those involved in hiring tutors, a standing obligation to give the EEOC written notice of discrimination complaints, and invitations to the applicants rejected in March and April 2020 to reapply. Because iTutorGroup had already ceased hiring tutors in the United States, a further obligation — to notify and interview those applicants — was written to trigger only if it resumes. Part of the decree has therefore never had occasion to operate. The Commission monitors compliance "for at least the next five years or longer if iTutorGroup resumes hiring tutors in the United States", which is a floor with a conditional tail rather than a five-year clock.
The case is widely described in the legal press as the EEOC's first workplace artificial-intelligence settlement. The agency's own releases never use the word: they say software, and they say programmed. That distinction is the case's doctrinal value rather than a quibble. Because the rule was authored, the theory was intentional disparate treatment, not disparate impact — which also means that the April 2025 executive order directing agencies to deprioritize disparate-impact enforcement does not reach this decree's theory, and nothing in the public record suggests the decree is threatened. What did change is the channel around it. In January 2025 the Commission removed its Artificial Intelligence and Algorithmic Fairness Initiative pages and its May 2023 Title VII technical assistance; it lacked a quorum from January 2025 until 7 October 2025; and the restored commission's published 2026 priorities do not list artificial intelligence in hiring at all. Removing guidance repeals no statute, and the Age Discrimination in Employment Act and the uniform selection guidelines are untouched. But the proactive federal channel this decree was announced under is measurably quieter than it was, while the decree itself stays court-enforceable.
The sociotechnical reading
Most of the hiring cases in this atlas are arguments about inference. A model learned who was hired rather than who succeeds; a vendor's testing is real but unreachable; an assessment measures only the cohort it selected. Each of those failures needs a paragraph to explain, and each leaves room to argue about whether the system was wrong. This one needs a sentence, and there is nothing to argue about. Somebody wrote a threshold, and the software applied it. The interesting question is therefore not how the failure happened but why it took three and a half years and a stranger's experiment to be noticed.
The answer is structural, and it is about where the automation sat. The rule was at the very top of the funnel, upstream of every human being in the process. That single placement does three things at once. It removes the correction channel, because the people whose judgment was intact — the staff who interview and select tutors — receive only the pool the rule produced and never see what it removed. It removes the complaint channel, because a rejection with no stated reason gives the rejected person nothing to appeal and nothing to name. And it removes the measurement channel, because a hiring outcome is only checkable against what a fair decision would have been, and nobody was in a position to make that comparison. A screen placed there does not need to hide. It is invisible by position.
Which is why detection took the form it did. The one comparison in the whole record was made by the person the system had already removed, from outside, with the only instrument available to her: change one field and resubmit. That is input perturbation, and it is worth naming as the audit method it is, because it required no access, no expertise and no cooperation — and because it worked instantly on a system that had been running for two months. An organisation that had run that test on itself once a quarter would have found this in a week. The gap between how cheap the test was and how long the failure ran is the governance finding here, and it is not a finding about technology.
The remedy is the second thing worth studying, because it did something the atlas rarely records: it acted on the input rather than the output. Courts and regulators usually order a system to be tested, documented, or reviewed. This decree prohibited collecting the field. That is an unusually complete fix for this specific rule — a threshold cannot fire on a birthdate it never receives — and an unusually narrow one, because it fixes this rule and says nothing about the next. Everything else in the decree is an attempt to build, after the fact, the structure the deployment never had: a policy, a memo, training for the people downstream of the gate, and a duty to route complaints to the regulator in writing, which converts a channel that used to end inside the company into one an outsider can read. The board draws that as the second record and the second reviewer, because in this deployment the oversight structure genuinely postdates the harm.
Two boundaries hold. Nothing on this network computes an outcome for any person: the count of more than 200 is a floor attributed to the Commission's complaint with no public denominator, the $365,000 and its distribution live here rather than on the diagram, and the question of who should have been hired is not one this Lab can answer. And the register stays exactly where the documents leave it. This was a rule, not machine learning; the conduct is alleged and was resolved without admission; the decree's obligations and monitoring term are facts. The reason the case belongs in an atlas about AI governance, despite containing no AI, is that it settles the prior question. An automated screen is answerable for what it does regardless of how it was built — and the mechanism that made this one hard to see is the same mechanism that makes a learned one hard to see, which is that it decides before anyone is looking.
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.