AI transparency in hiring: what a candidate is entitled to know

A disclosure line and a transparent process are not the same thing

The typical AI notice reads like a cookie banner: this process may use artificial intelligence. A candidate who reads it learns that something happened, not what. They cannot tell whether AI drafted the questions, transcribed their spoken answers, summarised their session for a reviewer, proposed a score, or produced an outcome. Those are five very different exposures, with five different things a person might reasonably want to ask about, and the sentence collapses all of them into one shrug.

The distinction matters because the candidate's next question is answerable only with specifics. If they want to contest an outcome, "AI was used" gives them nothing to contest. If they want to prepare, it gives them nothing to prepare for. If they have an accessibility need that interacts with a transcription step, the notice has told them nothing that helps. A disclosure that cannot be acted on is a legal artefact, not information.

There is an internal reason too, and it is usually the one that decides whether a process is genuinely transparent. An organisation that has published only a sentence has usually decided only a sentence. The teams able to describe their AI surfaces precisely to a candidate are the teams that inventoried them first, and that inventory is what makes the process governable at all — it is the list you check when someone proposes adding a summariser next quarter.

What a candidate is entitled to be told, before they start

The floor is short, and none of it is controversial once it is written down. What makes it useful is that every item is a fact about the process rather than a reassurance about the technology — something that either happened or did not, and that a candidate could ask a follow-up question about.

That is the test worth applying line by line. "Our AI is fair and unbiased" is not disclosure; it is marketing wearing disclosure's clothes, and a candidate has no way to check it. "A model drafts a score for each rubric criterion and a named evaluator confirms or overrides it before anything is recorded" is disclosure, because it describes a mechanism. Read your own notice sentence by sentence and ask what a candidate could do with each one. Sentences that support a question — which step, what output, whose name, what recording — are doing work. Sentences that describe your intentions are not, and intentions belong somewhere else.

Statutory regimes are converging on roughly this content. New York City's Local Law 144 requires notice to candidates before an automated employment decision tool is used, alongside a published summary of a bias audit. Colorado's SB 24-205 requires deployers of high-risk AI systems to notify the consumer where the system is a substantial factor in a consequential decision, and to offer human review of an adverse decision where technically feasible. Obligations vary by jurisdiction and change over time; what follows is a description of a framework rather than legal advice, and no claim of compliance is being made here for any product.

• That AI is used at all — before the assessment begins, not in the email carrying the result.
• Which surfaces it touches: question drafting, transcription, summarisation, draft scoring, integrity signals. Named individually, not folded into "the process".
• What each surface produces: a summary, a suggested criterion score, a flag for review. Never a status, an outcome, or a rejection.
• Who decides: a named person who can explain the decision, rather than "the system".
• What is recorded, and for how long — especially where camera, screen, or microphone capture is involved.
• How to raise a concern or ask for a review, and to whom.

Provenance is the part that makes disclosure meaningful

Disclosure is a promise about what will happen. Provenance is the evidence that the promise was kept. Concretely, it is a record attached to each AI-assisted artefact answering four things: which surface produced it, which model and version ran, what input it was given, and which named person confirmed or overrode the result — each with a timestamp.

Without those fields, "AI-assisted" quietly degrades into "AI-decided, unverifiably". You cannot tell whether two candidates' draft scores came from the same model version, so you cannot say their starting points were comparable. You cannot tell whether a summary was generated before or after the rubric changed. You cannot tell whether the evaluator whose name sits on a score had the raw response in front of them. With the fields present, an auditor, a candidate, or your own team six months later can reconstruct exactly where the machine's contribution ended and a person's judgement began.

NIST's AI Risk Management Framework names "accountable and transparent" among the characteristics of a trustworthy AI system, and treats documentation as the mechanism rather than the paperwork. That is the right way round: the record is not a tax levied on the process after the fact, it is the part of the process that makes everything else checkable. A practical way to test yours is to take a decision from three months ago and answer these without asking a colleague:

• Which AI output influenced this decision, and what exactly did it say? Not "a summary was generated" — the summary itself, as it stood at the time.
• What was it working from: the full response, a transcript, a truncated excerpt? An output is only as trustworthy as its input, and the input is the field people forget to store.
• Who turned it into a score, and did they change anything? An override is signal. A record showing no overrides anywhere is a rubber stamp with good posture.

How a true disclosure quietly becomes false

If any of those three questions is unanswerable, the disclosure shown to the candidate has become unverifiable. It was true when it was written; it simply can no longer be demonstrated. That is the failure mode worth designing against, because it is silent — nothing breaks, no error is thrown, and nobody notices until the day someone asks.

The usual cause is scope drift. A new AI surface is added — a transcription step, a summariser over interview transcripts, an integrity heuristic — and the notice written for the previous configuration is never revisited. Nobody lied. The sentence just stopped describing the process, and the gap widens one small feature at a time. The second cause is versioning: prompts, rubrics and models all change, and a record that stores "a large language model" rather than a version, or "scored against the rubric" rather than which revision of it, answers the question in a way that feels complete and is not.

Two habits prevent both. Treat the disclosure as a versioned artefact owned by a named person, with the AI surface inventory attached, so that adding a surface has an obvious place to fail if the notice is not updated. And bind each candidate's record to the version they actually saw: someone who accepted the January notice should have the January notice in their file, not today's. That second habit is more work than it sounds and it is worth it, because when an outcome is contested the question is never "what does your notice say" but "what was this person told" — and version-bound consent turns that from an argument into a lookup.

What transparency does not oblige you to publish

Transparency about AI does not mean handing over the assessment. Item content, answer keys, and scoring thresholds stay confidential, because publishing them destroys the measurement for everyone who takes the assessment afterwards — including, eventually, the candidate asking. The same holds for prompt text and model internals, which are neither useful to a candidate nor safe to expose.

The line runs between the mechanics of the process and the content of the instrument. A candidate is entitled to know that their written response was summarised by a model and scored against a rubric by a named evaluator. They are not entitled to the rubric's anchors before they sit down. Keeping the two straight matters in both directions: conflating them lets an organisation refuse the first by pointing at the second, and lets a vendor claim openness while publishing nothing anyone can use.

One category sits in between and deserves a deliberate decision rather than a default: aggregate results. Selection rates by group, reliability figures where there is enough data to compute them, bias-audit summaries where they are required — these describe how the instrument behaves rather than what it contains, and publishing them costs nothing in test security. If your transparency posture is genuinely about the candidate rather than about liability, this is the category where it shows.

The version we hold ourselves to

We build to a short list, written to be checkable rather than reassuring. AI in SkillCort drafts tasks and rubrics, summarises long responses and transcripts, proposes criterion-level scores against a rubric a person wrote, and surfaces integrity signals for review. It does not produce a status and it does not advance or reject anyone. Every draft score sits pending until a named evaluator confirms or overrides it, and overriding is exactly as easy as confirming — if disagreeing with the model costs more than agreeing, you have built an AI decision-maker with extra steps.

Each assisted output carries its provenance into the candidate's evidence file: which surface, which model and version, when it ran, what it was given, and whose name is on the confirmation. Integrity signals go to a person, never to an automatic outcome. Identity photos are compared by a named reviewer — there is no automatic face matching and no biometric face template is stored. And no candidate carries a score or a reputation from one employer to another, ever.

None of that is a claim that the technology is good. It is a claim about where the technology is allowed to sit, and about what remains on the record afterwards. That is the kind of claim a candidate can hold you to, which is the only kind worth publishing.

Sources & further reading

• NIST AI Risk Management Framework (AI RMF 1.0) — National Institute of Standards and Technology
• Automated Employment Decision Tools (Local Law 144 of 2021) — NYC Department of Consumer and Worker Protection
• Colorado SB 24-205, Consumer Protections for Artificial Intelligence — Colorado General Assembly