# How do I screen a job ad for discrimination risk?

> A pre-publication check of a job ad covering the ad text, the targeting settings on every platform, and a record of who approved publication, run under the positive duty.

## The trigger

A hiring manager sends through a draft ad for [ROLE_TITLE] and the recruiter is ready to publish. This check runs before the ad goes live, because discrimination risk is locked in at publication in two places: the ad wording and the settings that decide [who is actually shown the ad](/hr/ai-job-ad-targeting-who-sees-the-ad). Both sit inside the positive duty to take reasonable and proportionate measures against discrimination.

## The inputs

- The draft ad text for [ROLE_TITLE].
- The position description, to trace every requirement to a genuine need.
- The targeting and audience settings for each platform, including any saved or lookalike audiences.
- The delivered-audience report from the last similar campaign, if any.
- [ORGANISATION]'s job ad check record template.

## The steps

1. Compare the draft ad against the position description and list every requirement that is not genuinely needed for the job.
2. Run the prompt below so the AI flags coded language, unnecessary criteria and wording that could indicate discriminatory intent.
3. Open the targeting settings on each platform and confirm the audience is broad and role-based, no lookalike audience built from current staff is attached, and the restricted employment-ad category is on where available.
4. Resolve every flag with the named approver: amend the ad, or record why the flagged item is genuinely required.
5. Complete the check record showing what was checked, the flags and their resolution, the targeting settings, and who approved publication, then publish.
6. Diary a review of the delivered-audience report after the campaign runs, so any delivery skew feeds into the next one.

## The prompts

Paste the ad text and position description with this prompt.

```text
You are reviewing a draft job advertisement for [ROLE_TITLE] at
[ORGANISATION] before publication. Using only the ad text and
position description provided, flag for human review:

1. Any requirement in the ad not supported by the position
   description.
2. Any wording that could reasonably be understood as indicating
   an intention to discriminate on sex, age, disability or another
   protected attribute, including coded or gendered language.
3. Any criterion broader or stricter than the job genuinely needs.

For each flag, quote the exact wording and explain the concern in
one sentence. Do not clear the ad or state that no issues exist.
Your output is a list of flags for a human reviewer, who makes
the publication decision.
```

## The checks

- Every requirement left in the ad traces to the position description.
- Each AI flag has been read against the ad text by a person; "no issues found" from the model is treated as an unreviewed ad, not a cleared one.
- The targeting settings have been sighted by a person on every platform.
- No lookalike or workforce-shaped audience is attached.
- The check record is complete before publication.

## Governance controls

- A named person in [TEAM], typically the talent acquisition lead, approves publication and owns the decision. The AI flags risk; it never clears an ad.
- The check, its flags, the targeting settings and the approver are recorded, because [the positive duty](/hr/ai-positive-duty-hiring-performance) is discharged by reasonable and proportionate measures that can be evidenced, not by clean wording alone.
- Targeting settings are reviewed by a person on every platform, and the delivered-audience report is reviewed after the campaign, so delivery skew has an owner.
- No candidate or employee personal data goes into the prompt; the check uses the ad text, position description and placeholders only. For the statutory framing, work through [the LM-H07 positive duty module](/learning-hub/modules/lm-h07-sex-discrimination-act-1984-cth-and-the-respect-work-positive-duty).

## The output

A published job ad with screened wording and targeting, plus a check record filed against the requisition showing what was checked, what was flagged, and who approved publication.

## Time it takes

About 20 minutes on a first run, most of it tracing requirements and walking the platform settings. A practised run takes about 10.

---

Source: https://theaicommand.com/recipes/screen-a-job-ad-for-discrimination-risk
Publisher: TheAICommand (theaicommand.com)
Licence: see https://theaicommand.com/terms