Every AI hiring vendor is racing to answer it. The question that will actually decide who is still standing is smaller, harder to dodge — and we built the whole platform around it.
Is the model fair?
Unprovable years later. Contested between expert witnesses. Expires with every model release.
What did the system do to this person?
Has an answer or it doesn't. Doesn't expire. Same answer in a courtroom two years from now.
Nearly ninety per cent of large employers now screen candidates with software. That number is not going down, and it shouldn't — the volume is real, and no human team was ever reading four thousand résumés carefully. What is changing, quickly, is what a buyer is entitled to ask about how that screening worked.
The industry's answer so far has been to talk about fairness. Bias audits, fairness constraints, de-biasing pipelines, certificates on websites. It is a reasonable instinct and it is not sufficient, for three reasons that have nothing to do with anyone's good intentions.
Fairness is extraordinarily hard to prove after the fact. It is a statistical property of a population, argued between opposing experts, about a system that has since been updated twice.
It moves. Research through 2024 found language models preferring white-associated names in résumé screening by very large margins. Work on newer model generations finds the gap narrowed, absent, or reversed. That is good news — but it means a fairness result belongs to a specific model version on a specific date, not to a product. Ship an upgrade and the evidence expires.
And flat can mean blind. A 2025 study of three million comparisons found that some models scoring as statistically unbiased reached that flatness by matching surface keywords rather than reading the substance of the document. The researchers called it the illusion of neutrality — a screener that isn't really reading anything will look beautifully fair.
Against an 80% floor under the 1978 Uniform Guidelines. Roughly eight times below the line.
It is real, and it comes from a model generation since superseded. Treating it as a permanent property of AI screening is the same error in the opposite direction — and it is exactly why fairness evidence needs a version stamp to mean anything.
The evidence that convinced everyone — and the evidence that complicates it
Two separate studies, two different measures. Shown apart because they cannot be read on the same axis.
The direction reverses, then flattens — across eight model providers. Which is the point: a fairness finding is a fact about one version on one date, not a property of the category.
There is a smaller question underneath all of this, and it is the one that decides cases.
What, exactly, did the system do to this one person — on what data, under which version, against which criteria, and who made the final call?
It has an answer or it doesn't. It isn't contested by expert witnesses. It doesn't expire when a model is upgraded. And building to answer it produces a materially better platform, entirely independent of anything legal — because a system that can explain a score is a system whose scores you can improve.
That question is the design brief Intelletto was built to. Not retrofitted to, when the letters started arriving. Built to.
Fairness is a claim. A record is an artifact. One is argued; the other is handed over.
Every vendor in this market can make the claim. Very few can produce the artifact.
Two cases have reframed what a screening vendor is on the hook for, and they attack from opposite directions. This is not legal advice, and both remain undetermined — but the direction of travel is unmistakable.
A programme built for the first does almost nothing for the second. Bias testing is no answer to a claim that nobody was told. Consent architecture is no answer to a disparate outcome.
A complete decision record is the only thing that speaks to both.
Two theories, one vendor in the middle
Both pending. Allegations only, and neither has produced a final ruling.
Mobley v. Workday · N.D. Cal.
A court declined to accept that a screening vendor is merely a software supplier, finding it sufficiently involved in hiring to be treated as an agent of the employers using it. Conditionally certified as a nationwide collective covering applicants over forty. "We only make the tool" stopped working.
Kistler v. Eightfold AI · Cal. Super.
This one doesn't allege the algorithm was biased. It alleges it operated in secret — compiling applicant data from outside sources and scoring people with no notice, no access, and no way to dispute what was collected.
Regulators are converging on the same instinct. California's employment rules now reach any automated system that influences a hiring decision, not only one that makes it. Colorado's replacement law is built around notice, an explanation of adverse outcomes, and meaningful human review. The EU's high-risk obligations for recruitment centre on logging and traceability.
Read those three together and they are not asking whether a model is fair. They are asking for a record — and for it to still exist years later.
The record has to outlive the hire
Every major regime assumes a decision is still explainable long after everyone has moved on.
The design decision underneath the whole platform is simple to state and considerable to implement: the auditable record is not a report generated when someone asks for it. It is what the pipeline produces on the way to every decision.
Twelve orchestrated stages, and every one of them writes down what it did.
There is no red band. Intelletto cannot reject anyone. Green, blue and pool are recommendations a human has to act on — which keeps the accountability where the law puts it.
Configuration is hashed. Because fairness evidence is only valid for the exact setup tested, the exact setup is recorded — permanently, on the score itself.
What every scored candidate carries with them
Each line exists because a specific question can be asked about it years later.
Most pipelines have a single empty state. When a field comes back blank, the scoring engine treats it as an absence — as a fact about the candidate.
But a blank can mean two entirely different things. It can mean this person genuinely doesn't hold that qualification, which is a real signal. Or it can mean the reader didn't manage to parse it, which is a fact about the software and says nothing whatsoever about the person.
Parsing difficulty is not randomly distributed. It tracks with document format — dense layouts, unusual section headings, résumés written in a style that was standard twenty years ago, qualifications described the way a different education system describes them. Which is to say it tracks with age, career vintage and country of origin.
Collapse those two meanings into one blank and a parsing limitation becomes a scoring penalty that lands unevenly across protected groups, with no discriminatory rule anywhere in the code. Intelletto keeps them separate, by design, all the way through the pipeline.
A system should never let its own limitations be scored as the candidate's.
Where a field that matters cannot be read with confidence, the run halts rather than scores. We would rather return nothing than return a confident guess.
One blank, two entirely different meanings
What most pipelines record, against what Intelletto records.
Read successfully, has a value. The score can rely on it.
Document read properly, and they genuinely don't hold it. A real signal about a person.
Could not be read with confidence. Says nothing about the candidate — and never scores as a zero.
If you are evaluating screening software this year — or defending the software you already run — these are the questions we would want a buyer to put to us, and to everyone else on the shortlist.
"Bias-free" is now itself a contested claim — the language has been challenged before the FTC. Ask for the dated test on the named model version. A vendor who can produce one is telling you something. A vendor who can't, also is.
Can you show me one candidate's full decision record?
Not a summary or a dashboard export. The criteria, the version, the inputs, the score, the evidence, the human step — for one named applicant, on the day it happened.
INTELLETTOProduced automatically for every scored candidate, as a by-product of the pipeline.
Which exact configuration produced this score?
Model version, weights, and every setting that could have changed the outcome — as it stood at scoring time, not as it stands today.
INTELLETTOAn immutable configuration hash is stamped onto the score artifact itself.
Where does each scoring weight come from?
If the answer is "our model learned it," ask what it learned from. Weights derived from historical hiring inherit historical outcomes.
INTELLETTOEvery dimension cites a published external authority. Twelve of them, in a maintained register.
What happens when the system can't read something?
The most consequential question on this list, and the one least often asked. A silent blank scored as an absence is a disparate-impact mechanism with no intent behind it.
INTELLETTOThree explicit states. A failed read never scores as a zero.
Can the system reject anyone on its own?
Regulators now reach any tool that influences a decision, not only one that makes it. A vendor whose product can eliminate a candidate unattended has taken on the employer's exposure and passed you the bill.
INTELLETTONo rejection state exists. Green, blue and pool are recommendations for a person to act on.
How long will that record still exist?
California expects four years. Colorado three. The EU will require logging from December 2027. A record you can regenerate is not the same as a record you kept.
INTELLETTORetention is designed to the strictest regime in the deployment footprint, not the most convenient.
Today's procurement conversation is about accuracy and throughput. Within about two years I expect it to be a general counsel asking something far simpler: show me what your system did to this specific applicant, on the day it did it.
Vendors who built the record will hand over a file. Vendors who didn't will hand over a promise — and they will be assembling that promise under deadline, from logs that were never designed to answer the question.
We would rather already have the file. That is the whole reason Intelletto is built the way it is.
We will run a live requisition through the pipeline and hand you the complete decision record for every candidate in it — the criteria, the versions, the evidence and the human step.
Fairness will keep moving.
What happened doesn't.