AI in Hiring

AI hiring bias: how to review a vendor's evidence

Review AI hiring bias claims with a worked evidence checklist covering test scope, deployed versions, candidate coverage and unanswered questions.

An open folder of blank sheets sits beside a microphone and keyboard, with a cobalt-blue binder clip holding the papers.
On this page
  1. Define the claim and the intended screening use
  2. Read this fictional evidence packet
  3. Turn each scope gap into a focused evidence request
  4. Check what an aggregate result actually measures
  5. Keep candidate access inside the evidence scope
  6. Write the worked conclusion without a fairness grade
  7. Record changes before reusing the evidence
  8. Frequently asked questions
  9. Sources

Review an AI hiring bias claim by matching its evidence to the exact screening workflow you intend to use. Check the tested version, inputs, job task, people represented and outcome studied. Record unsupported extensions as questions for the responsible specialist, and keep the claim no broader than the evidence.

This is an evidence-request process for recruiters and hiring managers. It does not teach you to certify a system or determine whether a real candidate should advance. The separate AI hiring tool pilot guide covers workflow fit; a useful demonstration leaves different questions about a fairness claim.

Define the claim and the intended screening use

Start with the supplier’s exact statement and a description of the buyer’s proposed workflow. “Bias-free” leaves too much unstated: which potential harm, which people, which outcome and which use were examined?

Fairlearn’s assessment guidance starts with harms and affected groups, then considers how to quantify and compare relevant effects. That is more specific than treating fairness as a property that a product name carries into every hiring process.

Write down what the screen is supposed to measure. OPM’s structured interview guidance describes job-related competencies, predetermined questions and common evaluation benchmarks. It is a federal assessment-design reference, not an AI fairness certification or a required process for every private employer.

For a backend role, specify the technical task and the decision the report will inform. Use the technical screening handoff worksheet to document those requirements. A common job title alone does not establish that a study examined the same work.

Read this fictional evidence packet

The following example is invented. It contains no real supplier, candidate records or observed study results. Treat the listed documents as the complete supplied packet; anything not supplied remains unknown.

Supplied item What it states or contains
Supplier claim “Our AI technical-screening system is bias-free for your backend hiring process.”
Buyer’s intended-use brief Proposed scoring version B processes live English voice and coding responses about debugging and testing a backend reporting endpoint. It produces a rating for a technical reviewer; a person decides candidate progress.
Report A Describes comparison of scoring version A with human ratings on the same backend reporting task. Inputs are human-corrected English text transcripts from completed interviews. Audio processing and coding responses were excluded.
Configuration sheet Identifies scoring versions A and B. It provides no evidence showing that the version and input changes preserve the relevant measurement properties.
Available study detail No compared-group definitions or counts, subgroup results, uncertainty analysis or independent job-performance outcome is supplied.
Candidate-process material No analysis of incomplete or failed sessions, accessibility evaluation or alternative-route results is supplied.

The stated task matches, but the tested inputs and proposed inputs differ. Report A also describes a human-rating comparison rather than an independently verified job-performance outcome. Nothing in the packet establishes whether those human ratings were an appropriate reference for this purpose.

The initial finding is narrow: the supplied packet does not substantiate the broad claim for version B’s proposed use. It does not establish that version A is fair, that version B is biased, or that any particular group would be harmed.

Turn each scope gap into a focused evidence request

Keep the artifact, limitation and next question together. The NIST AI RMF Core addresses documented evaluation methods, deployment conditions, generalization limits and identified fairness risks. The cited framework is voluntary, dates from 2023 and is under revision; it does not certify a supplier.

Use the record below as a proposed working template. The owners are roles to assign within the buying team, not people who have reviewed this fictional system.

Gap in the supplied packet Evidence request Owner to answer or coordinate
Report A tests version A transcripts; proposed version B adds voice and coding Identify the changed components and provide evidence relevant to the proposed inputs. Explain which earlier findings remain applicable and why. Product or engineering lead, with a qualified assessment reviewer
The reference is human interview ratings Explain what the ratings measure, how the criteria relate to the role and how the reference ratings were established. Assessment reviewer and hiring manager
Group definitions, counts and subgroup results are absent Provide the study population, relevant comparison definitions, denominators and scoped results through an approved documentation or expert-access process. Qualified analyst, with privacy and legal review as needed
Uncertainty is undocumented Explain how sampling, missing observations and uncertainty affect interpretation. Identify conclusions the evidence cannot support. Qualified analyst
Only completed interviews enter Report A Account for relevant process stages, unfinished sessions and technical failures. Explain what the study can say about people excluded before a usable transcript exists. Research or assessment lead and recruiting operations
Accessibility and alternative routes are unexamined in the packet Supply the applicable evaluation scope, findings, unresolved barriers and evidence for any proposed alternative route. Accessibility reviewer and product owner

Request appropriate documentation or controlled expert access. A recruiter does not need unrestricted copies of candidate recordings to maintain this record. Agree on the data-governance route before requesting sensitive information, and do not infer demographic attributes from names, voices or images.

The UK DSIT recruitment-AI guide provides buyer questions about intended purpose, model and data, performance, limitations and supplier evidence. It is UK guidance published in 2024, not a statement of US legal obligations. The worksheet applies those procurement questions to a hypothetical technical screen.

Check what an aggregate result actually measures

Ask for the outcome and unit behind a reported result before interpreting its headline. A rate per completed interview answers a different question from a rate per invited applicant. A human-rating agreement measure answers a different question from subsequent job performance.

In this example, Report A’s completed transcripts exclude the earlier voice-capture process. The buyer cannot infer how version B handles someone whose audio never produced a usable transcript. That limitation follows from the fictional inputs; it is not a claim that voice interviews necessarily disadvantage a particular group.

Fairlearn’s assessment example examines disaggregated metrics and asks whether the underlying definition of a qualified applicant is appropriate. Its fairness metrics reference explains that metrics embody particular assumptions and definitions. A number needs its population, outcome and context before it can support a conclusion.

Keep uncertainty in the request too. Fairlearn’s confidence-interval guidance explains that resampling does not repair an unrepresentative sample. Ask a qualified analyst how the study’s coverage and uncertainty constrain its conclusions. Do not turn an absent table, an overall average or a single cutoff into your own fairness verdict.

For Report A, no subgroup results or uncertainty estimates are supplied. The record should say exactly that. It should not invent rates, assume the sample represents every applicant, or describe an unreported difference as zero.

Keep candidate access inside the evidence scope

The point at which someone enters the analyzed sample affects what the evidence describes. Ask what happened before a completed interview became a transcript: invitation, device permissions, session start, interruptions and any route for help.

The UK DSIT guide includes accessibility, user feedback and iterative assurance in recruitment-AI evaluation. For this proposed use, request evidence about those parts of the process as well as the final rating. A record of successful completions alone leaves the experiences of other participants unexamined.

In the fictional packet, no failure analysis or alternative-route results are supplied. Assign those questions to recruiting operations, product and accessibility reviewers. Record a failed recording as a process event requiring investigation; do not silently reinterpret it as evidence about a person’s engineering ability.

Human review also needs a defined role. A person looking at a report is not, by itself, evidence that the underlying system or process has been assessed for fairness. Identify what the reviewer can inspect, challenge and escalate, and which separate questions need assessment, privacy, accessibility or legal expertise.

Write the worked conclusion without a fairness grade

A useful completed record for this packet could read:

Claim reviewed: The supplier’s broad bias-free claim for the proposed backend screening use.

Evidence supplied: Report A describes human-rating comparison for version A using corrected English transcripts from completed backend interviews. The intended-use brief proposes version B with live voice and coding inputs. The configuration sheet lists versions A and B.

Applicability finding: The supplied material does not establish how the changed version and inputs affect the claim. Group-level evidence, uncertainty, incomplete-session analysis and accessibility or alternative-route findings are not supplied.

Next action: Product documents the changes. The assessment reviewer and analyst define the additional evidence needed for the intended use. Recruiting operations and accessibility reviewers address the unexamined candidate process. Privacy and legal specialists determine the appropriate data and review requirements.

Status: Needs evidence for this use. No fairness or compliance conclusion recorded.

That status describes the buying team’s documentation. It is not an audit result or a recommendation about a real candidate. Attach the artifacts, requests and subsequent responses so another reviewer can follow the reasoning.

Record changes before reusing the evidence

Keep a change record beside the evidence request: the component or configuration that changed, the earlier artifact it could affect, the person reviewing applicability and the next evidence needed. In this example, version B’s voice and coding inputs are explicit changes, not details to overlook because the supplier name stayed the same.

NIST’s framework includes ongoing monitoring and expert input. Apply that principle by assigning responsibility for reviewing relevant changes and newly reported problems. Set the review plan for the actual use; this article supplies no universal interval or release threshold.

Prepin runs AI technical interviews by voice and live coding and gives recruiters a scored report for human review. Apply the same evidence questions to that workflow: See how screening works.

Frequently asked questions

Does missing bias evidence prove that a tool is biased?

No. It means the supplied material does not establish the particular claim being reviewed. Record what remains unknown and request evidence appropriate to the proposed use; do not infer a group’s outcomes from a missing appendix.

Does a human-rating comparison establish job-performance validity?

It describes a comparison with the chosen ratings. Whether those ratings are an appropriate reference, and what conclusions the study supports, need qualified assessment review. In the fictional packet, no independent job-performance outcome is supplied.

What if the supplier cannot share candidate-level data?

Ask which privacy-appropriate documentation, aggregate results or controlled expert review can answer the question. Agree on suitable access and handling with the responsible specialists. Do not collect sensitive attributes ad hoc or infer them from candidate content.

No. It is an evidence-request record, not a statistical audit, validation study or compliance certificate. Use qualified assessment and accessibility expertise for their respective questions, and talk to employment counsel about obligations for the intended use.

Sources

Prepared with AI assistance; the official sources below were checked on November 19, 2026. The fictional packet and worksheet are original teaching examples, not a reported supplier evaluation.

Oleh Savchuk, founder of Prepin

Oleh Savchuk

Founder, Prepin

Oleh Savchuk is the founder of Prepin, an AI hiring marketplace where recruiters screen engineers with AI-run technical interviews and candidates practice for theirs. He is a software engineer in Philadelphia and writes about technical hiring, AI in recruiting and interview preparation.

Screen engineers before the first call

Send candidates one link. Prepin runs the technical voice interview and live coding, then gives you a scored report.

See how screening works →

Practice the interview out loud

Run a realistic AI mock interview for your target role and get feedback on every answer.

Start a mock interview →