Fairness Review for Job Descriptions
Review a job description for job-related language, access, proxy risk, evidence and accountable human decisions before sourcing starts.

A fairness review for a job description is a record made before the role is published or used as a search brief. It tests whether the language describes the work, whether applicants can access the process, whether criteria create proxy risks and whether a person owns the call. This is educational information, not legal advice; apply relevant local rules.
What should a job description fairness review check first?
Name the role, version, decision owner and audience. Describe outcomes, working conditions and constraints. Avoid coded requests such as "native speaker", "high energy", "culture fit" or "recent graduate" unless the team can explain a necessary, job-related meaning and fair assessment. The OPM Assessment Decision Guide, issued 6 June 2007, recommends connecting selection content to job analysis and giving applicants a realistic job preview. It is US federal guidance, not universal law.
Separate essential from preferred criteria. An essential criterion maps to a task that must be performed at the required level, with reasonable accommodation where applicable. A preferred criterion should not quietly become a rejection rule. Write acceptable equivalent evidence: "five years in a venture-backed company" may proxy for ambiguity, while "can describe owning a production change and its trade-off" asks about the work.
How to check job description language, access and proxies
Read the advert as an applicant using a screen reader, keyboard, zoom or a different first language. Use clear headings, plain instructions, a contact route and an explanation of stages. Check fields, documents and screening tasks against the organisation's accessibility route. WCAG 2.2 became a W3C Recommendation on 12 December 2024; it provides web-content criteria, not proof that hiring is fair or every accommodation is available.
List possible proxies beside each criterion. Names, photos, postcode, school prestige, accent, employment gaps, salary history, age-coded words and unpaid availability can stand in for protected or irrelevant traits. A proxy is a risk, not proof of discrimination. Remove it, narrow it to the work, or document necessity and a less exclusionary alternative.
The EEOC's Employment Tests and Selection Procedures guidance, issued 1 December 2007, explains that a neutral procedure can have disparate impact and should be job-related and consistent with business necessity in its US scope. Its 12 May 2022 disability and AI warning highlights accommodation and screening-out risk. These sources inform questions; they do not decide your legal position.
What goes in a job description evidence record?
For every criterion, record the task, essential or preferred label, observable evidence, acceptable equivalent, assessment stage, owner and open question. Keep source wording separate from interpretation. Silence in a CV or profile is not a negative. Give comparable candidates the same verification opportunity.
For later outcomes, keep the denominator visible: eligible applicants, people reaching the stage and people selected are different populations. A percentage without counts, inclusion rule, time window and missing records is not a fairness conclusion. Small groups may be unstable. The NIST AI RMF 1.0, published 26 January 2023, is voluntary, rights-preserving guidance for managing AI risk. Use it to assign owners and tests, not claim validation.
Fictional example: one platform-engineer role
The platform-engineer example below is fictional, not a benchmark. A four-person team is hiring a platform engineer. Its draft says "rock-star engineer with 5+ years at a top-tier technology company, native English, and always-on availability." The owner replaces it with two essential outcomes: operate a production service and explain a reliability trade-off. Preferred evidence is infrastructure-as-code and incident learning; public projects or a different service scale remain eligible.
The owner flags "top-tier", "rock-star" and "native English" as proxy or access risks. The advert describes the communication task, hours and on-call requirement, with an accommodation route. A reviewer writes one evidence question per outcome and records source, date, interpretation and action. The team does not infer ability from a name, school, accent or career gap.
The release record uses four states:
| State | Meaning and action |
|---|---|
| Evidence | A dated, relevant source supports a narrow job-related statement; retain the source and interpretation. |
| Unknown | The source is silent, stale or ambiguous; ask a comparable verification question. |
| HOLD | An owner, accommodation route, denominator, purpose or retention decision is missing; pause until the release condition is met. |
| STOP | The draft relies on an irrelevant or protected-trait proxy, seeks unnecessary sensitive data or treats a tool output as a final decision; remove the line and escalate. |
Here, "native English" is STOP. On-call is HOLD until hours, rotation and accommodation handling are specified. Repository evidence is Unknown until ownership and recency are checked. The hiring manager owns criteria, the privacy owner approves data handling, and a named reviewer can challenge the draft.
How to keep a job description revision log
Do not overwrite the first draft. Record version, date, change, reason, owner and release state. For example (fictional): v1 / 2026-09-02 / replaced "native English" with a communication outcome / accessibility risk / hiring manager / HOLD; then v2 / 2026-09-03 / added hours, accommodation route and equivalent evidence / condition met / hiring manager / Evidence. Keep candidate wording aligned with the brief. Explain what is assessed, how to ask questions and how to request an alternative.
The ICO's 6 November 2024 recruitment-AI recommendations emphasise fairness, transparency, minimisation and retention. Collect only needed fields, restrict access, record a deletion or review date and never ask AI to infer sensitive traits. A named human owns corrections and escalation.
Where Talent Summoner fits
Talent Summoner is our product for AI candidate sourcing, CV ranking and outreach you confirm before it sends (checked 28 September 2026). Candidate sourcing starts from a brief and searches LinkedIn, GitHub and other public professional sources, returning a ranked shortlist with reasoning. Candidate ranking organises supplied CVs. Our product does not determine legal fairness, provide accommodations, establish denominators, verify every claim, manage an application pipeline, automatically reject candidates or make the decision. Your team owns the brief and controls.
Does a fair job description guarantee fair hiring?
No. It reduces ambiguity at the entry point. Later sourcing, assessment, accessibility, data and decision stages need their own review.
Should every preferred criterion be removed?
No. Keep one when it connects to useful work, is assessed consistently and cannot become an essential screen. Record the reason and alternatives.
Can I conclude there is adverse impact from one percentage?
No. Define population, denominator, stage, counts, missing records and uncertainty first. A signal is not a legal conclusion.
Does Talent Summoner rewrite a job description for compliance?
No. It supports discovery and supplied-CV ranking. Your team reviews language, access, privacy, evidence and decisions.
For the next role, assign hiring, privacy and accessibility owners, resolve every HOLD or STOP, and use candidate sourcing for discovery and candidate ranking for supplied CVs. Keep human review and candidate communication in your record.


