The EEOC’s iTutorGroup settlement is cited as the first federal enforcement action over AI in hiring. The figure and the date are right. The technology at the centre of it was a screening rule that rejected applicants for being born before a cut-off, which is worth knowing before you build an argument on the case.
What the EEOC filed
The Equal Employment Opportunity Commission sued iTutorGroup, Inc. and two affiliated entities in the United States District Court for the Eastern District of New York in May 2022. The companies recruited tutors in the United States to teach English to students in China, hiring at volume through an online application process.
The claim was under the Age Discrimination in Employment Act of 1967. The Commission alleged that in 2020 the application software was programmed to reject female applicants aged 55 or older and male applicants aged 60 or older, and that more than two hundred qualified applicants were rejected automatically on that basis. The charge that started it came from an applicant who was rejected, reapplied with a more recent date of birth and nothing else changed, and was offered an interview.
On 9 August 2023 the parties filed a joint notice of settlement, and the case resolved by consent decree. The EEOC’s announcement of the settlement states the terms.
A summary of public filings, not legal advice. Age discrimination exposure in hiring depends on your process, your data and your jurisdiction, and federal, state and local rules differ — take advice on your own screening rather than reasoning from a settled case.
The mechanism actually alleged
The allegation is a configured rule, not a learned model. Applicants supplied a date of birth; the software compared it against a threshold that differed by sex; applications past the threshold were rejected without human review. There is no allegation of a trained classifier, a proxy variable, or a disparate impact emerging from data nobody intended. The protected characteristic was an input, and the rule acted on it directly.
That makes it a disparate treatment case in a straightforward sense. The ADEA prohibits refusing to hire an individual because of age, and software that refuses on the basis of a date of birth does that explicitly. The novelty is not the legal theory, which is sixty years old. It is that the discriminating act was a configuration setting applied at scale in seconds, which is the part worth transferring to a modern stack.
One detail is worth dwelling on, because it explains how these cases get discovered. The evidence that made the claim provable was the natural experiment the applicant ran herself: same person, same qualifications, different date of birth, different outcome. An automated rule is testable in a way a human hiring manager is not, and a rejected applicant with a spare email address can run the test. Systems that make decisions deterministically from an input also make those decisions auditable by anyone who suspects them.
The consent decree terms
The decree, filed in the Eastern District of New York, required a payment of $365,000 distributed to a group of approximately two hundred applicants aged 40 and over who were rejected in the relevant period. Alongside the money it required the usual injunctive apparatus: adoption and distribution of anti-discrimination policies, training on the ADEA for personnel involved in hiring, invitations to the affected applicants to reapply, and reporting to the Commission over the decree’s term.
Two features are standard and both matter to how the case should be cited. The companies did not admit liability — a consent decree resolves a dispute without adjudicating it. And the monetary figure is not a measure of the conduct’s seriousness; it is a settlement number reflecting the class size, the wage loss claimed and the litigation risk on both sides. $365,000 across roughly two hundred people is a small individual recovery, and reading the amount as the price of algorithmic age discrimination would be a mistake.
Why calling it an AI case overstates it
The EEOC’s own framing was about automated screening tools, and the case is genuinely the first the agency resolved involving software making the rejection. But the reason it settled quickly is that the conduct was legally simple: an explicit age cut-off is per se disparate treatment, and no expert testimony about model behaviour was needed to establish it.
The hard problems in algorithmic hiring are the ones this case did not reach. Whether a model that never sees age but proxies it through graduation year or years of experience produces unlawful disparate impact. Who is liable when the tool is a vendor’s and the decision is the employer’s. What a plaintiff has to plead when the selection mechanism is opaque to them. How a validation study defends a model under the business necessity standard. None of these was in issue, and a settlement cannot answer them anyway.
Those questions are being litigated elsewhere. The vendor liability question in particular is the live one in the Mobley v. Workday litigation, which is a very different procedural animal: contested, at scale, and producing actual rulings.
What it does establish
Three things, and they are practical rather than doctrinal.
Automation is not a defence and never was. The ADEA asks whether the applicant was refused because of age. Whether a person or a rule applied the criterion is irrelevant to the elements of the claim, and the EEOC has been consistent that existing anti-discrimination law applies to automated selection procedures without needing new statutory authority — its guidance on Title VII and selection tools takes the same position for disparate impact.
The screening configuration is a compliance artefact.Somebody set those thresholds, and the rule presumably sat in a form builder or an applicant tracking system where it looked like a preference rather than a policy. Any protected characteristic collected during an application — date of birth, sex, graduation year, disability accommodation requests — is a field that must not be available to an automatic filter, and the control for that is a review of the filter configuration itself, not of the hiring policy document.
Rejection at volume creates a class. A rule applied to every applicant produces a defined group of affected people with a shared claim and a machine-readable record proving membership. The reason a small configuration error becomes a two-hundred-person case is that the applicant tracking system kept the list.
Top comments (0)