Domain Atlas / Hiring & employment screening AI
Aon pre-hire assessment suite (vendor's own tables)
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: 2 assumed · 9 published baseline.
Aon Consulting, Inc., the human-capital arm of Aon plc, designs, markets and administers a suite of pre-hire assessments sold into other companies' hiring pipelines. Three of them are at issue: ADEPT-15, a computer-adaptive forced-choice personality test scoring fifteen constructs from item pairs matched for social desirability out of a bank of more than a thousand statements, with no skip option and no agree-with-both option, whose outputs are construct scores, job-fit scores and generated interview scripts for the employer; gridChallenge, a gamified working-memory test of nine items each pairing a timed target-icon memorisation task with a symmetry or shape-judgment distractor task; and vidAssess-AI, an asynchronous video product that transcribes a candidate's spoken answers with a commercial speech-to-text service and scores phrases in that transcript as far-extreme positive or negative indicators of the same ADEPT-15 constructs. On Aon's own marketing, quoted in the ACLU's complaint, the company administers more than 30 million assessments a year in over 40 languages across 90 countries, and third-party market data quoted in the same paragraph places it at the second-largest share of the global pre-hire assessment market. The assessments are marketed to screen out applicants prior to more resource-intensive hurdles such as resume screening and interviews, and Aon's own case-study marketing is cited for one client that screened out 62 percent of applicants on the assessment alone; for the population removed at that step there is no human review, and the assessment is the first and the final decision point. Every scale and screen-out figure here is Aon's own marketing and none is independently audited.[3]
What happened
A candidate applies for a job and is asked, before anyone reads a resume, to take a test. It may be the ADEPT-15 test, a computer-adaptive forced-choice personality assessment that scores fifteen constructs — Awareness, Positivity, Liveliness, Sensitivity, Drive and others — from pairs of statements matched for social desirability, drawn adaptively from a bank of more than a thousand statements running to hundreds of thousands of possible pairings. There is no skip option and no option to agree with both. It may be gridChallenge, a gamified working-memory test of nine items, each pairing a timed memorisation task over highlighted circles or arrows with a symmetry or shape-judgment distractor task. It may be vidAssess-AI, an asynchronous video product: the candidate records spoken answers, a commercial speech-to-text service produces a machine transcript, and a language classifier — a program that sorts text into predefined categories — maps words and phrases in that transcript onto far-extreme positive or negative indicators of the same ADEPT-15 constructs. The results — construct scores, job-fit scores, and interview scripts generated for the recruiter — go to the employer, and the employer decides what they mean.
The scale is the company's own claim: more than 30 million assessments a year, in over 40 languages, across 90 countries, with third-party market data quoted in the same paragraph of the complaint placing Aon at the second-largest share of the global pre-hire assessment market. So is the marketed function. The assessments are sold to screen out applicants prior to more resource-intensive hurdles such as resume screening and interviews, and Aon's own case-study marketing is cited for one client that screened out 62 percent of applicants on the assessment alone. For the applicants that step removes, there is no human review at all: the assessment is the first and the final decision point, and no person reads their file.
Now the part that makes this case different from every other hiring file in this atlas. No model here is alleged to have misbehaved. The contested link is between two sets of documents Aon itself wrote.
In its buyer-facing technical documentation, Aon published what its own studies of its own instruments measured. On ADEPT-15, the reliability table shows ten of fifteen constructs producing two-week test-retest coefficients below the 0.7 floor Aon's own documents tell buyers to demand; the lowest is 0.44, on 'Awareness', which the complaint identifies as the construct overlapping most with autism diagnostic screeners. A two-week test-retest coefficient of 0.44 means the same person retaking the same instrument lands materially elsewhere. On gridChallenge, the gridChallenge and GAME test documentation at Table 43 reports Black assessment-takers scoring on average 0.48 standard deviations below white assessment-takers, with two-or-more-ethnicities at 0.38, Hispanic or Latino at 0.37 and Asian at 0.35. Aon's documentation for sibling cognitive tools reports Black test-takers averaging 0.86 standard deviations lower on scales clues, an information-processing test, and 1.21 lower on switchChallenge, a deductive-reasoning test — 'large' effects under the Cohen's-d heuristic Aon applies in those same documents.
On its sales pages, Aon wrote that the assessments are fair for all with no adverse impact, that they are culturally agnostic, that prejudices are excluded, and that they are scientifically proven not to have bias.
In late 2023 the ACLU, with co-counsel Winston Cooks, LLC, filed class-wide charges of discrimination with the Equal Employment Opportunity Commission against Aon and against a client employer, under the Americans with Disabilities Act and Title VII, on behalf of a biracial Black and white autistic job applicant with mental-health disabilities. The employer is publicly identified only as a mid-sized company headquartered in the United States, and it is not named here or anywhere. On 30 May 2024 the ACLU filed a complaint and request for investigation with the Federal Trade Commission, alleging that marketing the assessments as bias-free is a deceptive act under Section 5 and that selling them while failing to assess disability harms is an unfair practice injuring workers and employers alike. Keep the mapping straight, because it is easy to get wrong and the whole shape of the case depends on it: the EEOC charges cover ADEPT-15 and gridChallenge; vidAssess-AI appears in the FTC complaint and does not appear in the charges.
Everything characterizing these assessments as discriminatory or as deceptively marketed is an allegation. Aon disputes it, and its response through trade press in June 2024 is the only substantive statement it has made on the record: its assessment solutions 'follow industry best practices as well as the Equal Employment Opportunity Commission, legal and professional guidelines' and are 'intended to be fair to everyone'. In August 2025 an Aon spokesperson declined further comment.
The honest qualification on the numbers is load-bearing and the complaint makes it itself. An average score difference is not an adjudicated adverse impact. Whether a gap becomes a selection outcome depends on the cut-off score the employer sets and on what the assessment is combined with — a decision Aon does not control, conceded at paragraph 62 of the complaint. The complaint cites the selection-statistics literature for the general proposition that even small average differences can produce four-fifths-rule problems at commonly occurring selection ratios, and that is a statement about selection arithmetic rather than a measurement of any Aon client's pipeline. No such measurement exists in this record.
The disability theory runs the same way: through overlap, not through anything anyone counted. Statements the charging party recalled encountering on ADEPT-15 — 'I tend to avoid large groups of people', 'I have difficulty determining how someone feels by looking at their face' — are set in the complaint's Table 1 beside near-parallel items on the clinical autism screening test AQ and the RAADS-R scale. The wordings are the charging party's recollections and the precise wording may differ. Aon's own interpretation guides flag extreme-end scorers as potential problems for employers. Meta-analyses cited in the complaint show autistic people scoring significantly lower on working-memory measures of the kind gridChallenge uses. And then the design record, which is the part that carries the argument: across a 166-page ADEPT-15 technical documentation, the only disability-bias measure the complaint identifies anywhere is an outside-attorney review of item wording that flagged physical-disability language such as 'seeing', 'speaking' and 'hearing'. No clinical or neurodiversity sensitivity review. No study on disabled populations. What is alleged is a high risk of screening out autistic and disabled applicants, plus one charging party's individual account; no adjudicated or statistically demonstrated screen-out of disabled applicants exists here.
There is one more channel, and it closes a loop. Aon's test-taker guide tells candidates that 'there is no possibility of bias — for or against specific candidates — either during the test or during the evaluation of answers', and discusses accommodations only for physical disabilities. The ACLU's argument is that a neurodivergent candidate who reads that does not make the accommodation request that would generate the only disability-relevant signal anywhere in the system — and that the absence of the signal then reads as evidence that nothing was needed. Invoking the request costs the candidate a disclosure of disability. The quoted text is Aon's; the alleged suppressing effect is the ACLU's.
The vidAssess-AI argument is a compounding one and it carries the thinnest provenance in the file, which the complaint itself flags. Two inference stages sit on top of a personality model already in dispute: a commercial speech-to-text stage, then a language classifier, a program that sorts text into predefined categories, tagging transcript phrases as far-extreme indicators of the constructs. The identification of which commercial services those stages are comes from a developer's published dissertation and a patent rather than from any Aon disclosure. The independent research the complaint cites — roughly double the speech-recognition error rate for Black speakers — tested a system from the same provider, not necessarily the deployed version. Nothing in this file names a provider, and nothing rests on the identification.
Then watch what happened when scrutiny arrived, because the record is two sentences long. At some point after the EEOC charges were filed, Aon disabled the public link to the assessment technical documentation on its Norwegian support portal — the documents whose tables anchor the complaint. The ACLU states it knows of no other public source and cites archived copies. The marketing did not change.
Meanwhile the ground moved under both filings. In January 2025 the EEOC removed its artificial-intelligence and algorithmic-hiring technical-assistance guidance following the administration change. On 23 April 2025 an executive order directed federal agencies to deprioritize disparate-impact liability enforcement — the exact theory underlying the race claims. Legal-press analysis in March 2025 recorded the retreat, listed the Aon matters among pending ones with no indication of closure, and predicted that AI-bias enforcement would shift to states and private plaintiffs. In August 2025 Bloomberg Law reported the EEOC charge as filed and unresolved.
As of 28 August 2026 no public FTC action, no public EEOC determination, settlement or follow-on lawsuit, and no product withdrawal has been located. That statement has to be read exactly as it is written. Charge proceedings are confidential, so the charge could have been resolved, withdrawn or dismissed without public trace; 'pending' is not a positive fact anyone can assert. The Commission never docketed a public matter, so 'the FTC has not acted' means no public action, and an unannounced staff inquiry can be neither ruled in nor ruled out. The contracting enforcement environment explains the silence without implying anything about the merits, and this file reads it in neither direction.
What is on the record, checked on the run date: Aon's pre-hire talent assessment page advertised 'Fair and unbiased assessments' and asserted that 'Assessment tools that are scientifically proven not to have bias are essential'. Twenty-seven months after those sentences were formally alleged to be deceptive, they were still selling the product.
The sociotechnical reading
Most files in this atlas turn on what a system did. This one turns on what an organisation published, and on what it published beside it.
Start with the fact that the case would not exist without an act of disclosure. Aon ran the studies. Aon computed the two-week test-retest coefficients across fifteen constructs and found ten of them below the 0.7 floor its own documents tell buyers to demand. Aon computed the group score differences and put them in a table. Aon supplied the Cohen's-d heuristic by which a reader can grade 0.86 and 1.21 as large. Then Aon put all of it in buyer-facing manuals on a public support portal. That is more than most vendors in this atlas disclose about anything, and it is the entire reason a forty-eight-page filing quoting at table and page level was possible. The uncomfortable reading follows immediately: the only reason this deployment is documentable is that its own vendor documented it, and the moment the documentation became a liability, the link went dark. A regime that relies on voluntary vendor disclosure has no answer to that, and this record is what the absence of an answer looks like.
Now the structural point, which is what the diagram is drawn to show. In almost every other case here the contested object is a model's behaviour: an error rate, a false-positive flood, a decision that went wrong. Here the contested object is a set of sentences. The evidence store and the sales channel belong to the same organisation, they were written by the same function, and nothing carried the first into the second. That is the whole deception theory, and on the board it is two pathways at zero running in opposite directions between the same two stores: nothing replicating the manual's findings into the marketing, and nothing reconciling the marketing's claims against the manual. The case is the gap, and a gap is exactly what a governance diagram is good at drawing.
Three parties, and one of them is standing in two places at once. Aon faces a deception theory and a discrimination theory simultaneously. The client employer is a co-respondent on the discrimination charges and — on the unfairness theory in the very same filing — one of the parties the deception injures, because the marketing is alleged to lull the party that actually bears adverse-impact liability out of vetting the tool it is answerable for. That contradiction is not rhetorical. It is a real feature of a market where the party that builds the instrument holds the evidence about it while the party that buys the instrument holds the legal exposure and the one decision that matters. Where the cut-off falls is what converts an average score difference into a selection outcome, and the complaint concedes as much; the party that sets it is the party the same complaint says was kept from the information it needed.
Look at what the assessed person can do, because on this board it is almost nothing and the reasons are documented rather than inferred. The instruments sit before any human step, which is the marketed feature and not an implementation accident: they are sold to screen out applicants prior to more resource-intensive hurdles. For the population removed there, no file is read by anyone. The one instrument a candidate holds is an accommodation request — and the guide they read first states that bias is not possible and frames accommodations as a physical-disability matter, so the request that would generate the system's only disability-relevant signal is the one a neurodivergent candidate is least likely to know to make. Then the absence of requests becomes its own evidence. This is the tightest loop in the file, and its two ends are a page of text and a screen-out nobody sees.
The forums are the last thing, and they are worth reading as a measurement rather than as a verdict. Two federal agencies received two different theories about one product suite at almost the same time: consumer protection reaching the marketing, and civil rights reaching the selection. Between them they can compel production, enjoin claims and order relief, which is more power than anything else in this file has. Neither has publicly used it. During the pendency the employment agency withdrew its own artificial-intelligence hiring guidance and an executive order told agencies to deprioritize the exact liability theory the race claims rest on. Read the silence carefully: charge proceedings are confidential and the consumer agency never docketed anything publicly, so the record supports 'no public action located' and nothing beyond it — not that the filings lacked merit, and not that they had it. What the record does support is narrower and more useful. A complaint that can compel nothing, filed in two forums at once, against a company still selling the product and still running the sentence, is the ordinary case rather than the exception; and the one observable response it produced was a reduction in what the public can read.
Two boundaries hold here and they are not decoration. Assessed people are not modelled: no hiring decision, screen-out, score or employment outcome for any person is computed from anything on this diagram, and the group score differences are recorded external observations — mean differences in standard deviations, reported by the vendor about its own study samples and quoted by a complainant. They are not selection-rate ratios, not error rates, and not findings of discrimination against anyone. And the client employer named as a co-respondent is publicly unnamed; Aon's marquee customers appear in the complaint as evidence of reach and are never the respondent.
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.