Job descriptions and screening

Create an approved role brief independently of any candidate. It should describe:

• the outcomes expected in the first six to twelve months;
• essential tasks and working conditions;
• skills or knowledge genuinely required on entry;
• capabilities that can be learned after hiring;
• lawful location, schedule, travel, or certification constraints;
• how each criterion will be assessed;
• reasonable alternatives to conventional credentials or career paths.

Separate essential criteria from preferences. If a degree is not necessary to perform the work, treating it as essential may exclude capable applicants. If a task requires writing clear incident reports, assess that capability directly instead of using school prestige as a proxy.

Criteria must be stable enough for consistent treatment, yet revisable when evidence shows they are poor. Record who approved them and when. Changing criteria after seeing a favored applicant undermines fairness whether AI is involved or not.

AI can turn the brief into candidate-facing text. A useful prompt might say:

Draft a plain-language advertisement using only the approved role brief. Separate essential criteria from capabilities that can be learned. Preserve the stated working conditions and salary range. Avoid age-coded, gender-coded, ableist, or inflated language. Do not invent benefits, qualifications, or flexibility. Mark any missing information with [CONFIRM].

Review the draft for more than grammar. Check that the title reflects the work, responsibilities are realistic, required criteria match the brief, accessibility information is present, and claims about culture or benefits are approved. Ask people familiar with the role and inclusive hiring practice to review.

An automated language checker can flag terms for consideration, but it cannot declare an advertisement inclusive. Context matters, and exclusion can arise from structure, distribution channels, application design, or inflexible requirements that no word substitution fixes.

For screening support, translate each approved criterion into an evidence question. For example:

• Criterion: can plan work across competing deadlines.
• Evidence question: what example shows the applicant identifying priorities and adapting a plan?
• Acceptable sources: application response, résumé project description, portfolio, or later structured interview.
• Non-evidence: résumé layout, writing style unrelated to the role, name, address, or inferred personality.

Use simple categories such as evidence present, evidence unclear, and not evidenced in submitted material. “Not evidenced” is not the same as “incapable.” An application may omit information that an interview or work sample can reveal. Avoid pseudo-precise scores such as 87.4 unless there is a validated measurement model and a meaningful interpretation.

A bounded screening prompt can request:

For each published criterion, quote or locate relevant evidence from this application. Do not infer protected characteristics, personality, commitment, or capability beyond the text. If evidence is absent or ambiguous, say so. Do not rank, recommend, reject, or compare applicants.

The recruiter then verifies every quotation in the original application. Models can misattribute achievements, overlook evidence in unusual formats, and interpret a team result as an individual claim. Applications containing instructions aimed at the model can also manipulate an automated workflow. Treat applicant content as untrusted input and isolate it from system instructions.

Names, photographs, addresses, graduation years, disability disclosures, family information, and other sensitive details should not become convenient proxies. Removing a field does not guarantee fairness: schools, employers, activities, and language patterns may still correlate with protected or socioeconomic characteristics.

Do not ask a model to infer age, ethnicity, gender, health, religion, sexual orientation, pregnancy, nationality, or personality. Do not treat gaps in employment as evidence of low commitment. A gap can reflect caregiving, illness, migration, education, economic conditions, or a choice unrelated to job performance.

Check local law and organizational policy before processing candidate data or deploying automated employment tools. Notice, assessment, consultation, recordkeeping, and audit obligations vary by place and context.

Two reviewers do not automatically produce fairness. Define what they review, provide the original evidence, train them on the criteria, and require reasons for consequential judgments. Sample both included and excluded applications for quality review. Track disagreement and investigate whether particular criteria or groups experience higher reversal rates.

Candidates should receive an accurate explanation of how AI assists the process where required or appropriate. Provide an accessible way to request accommodation, correct data, or raise a concern. A meaningful review cannot be performed by merely asking the same model to reconsider its output.