Audit the rules that remove candidates, including thresholds, missing answers and the evidence behind each requirement.
A screening rule can make a decision before anyone reads the application. If you require a particular graduation year, reject an unexplained career gap or turn an empty field into a failed answer, start your review with that rule. Ask what work it helps you assess and what happens to candidates it removes.
Looking only for a field labelled age or disability won't answer those questions. Natalie Sheard's Australian study of AI hiring systems describes how proxies and the way recruiters use filters can create discrimination risks. It is qualitative research, based on interviews conducted in 2022 and 2023, rather than a measurement of your software or today's market. Its useful challenge is to examine how a requirement becomes an exclusion. Read the original study.
Use the audit below with the person who knows your screening configuration. You need the rule as the system applies it, not just the wording in the job brief.
Follow the field through to the rejection
A proxy is something used in place of, or associated with, another characteristic. In screening, the concern is that a criterion can stand in for a protected attribute even when that attribute isn't explicitly requested. That makes a field worth examining; it doesn't establish that every use of it is unlawful.
The legal question depends on the applicable law and circumstances. For example, Anti-Discrimination NSW describes indirect disability discrimination through a common rule that unfairly affects people with disability and is not reasonable in the circumstances. The reasonableness qualification matters. Get appropriate advice about a disputed requirement rather than deciding its legal status from a correlation alone.
Begin with one active vacancy. Find out which fields affect the screen, including rules supplied by a client or inherited from an earlier search. For each one, ask how it is used. Is it a search aid, a weighted score, a threshold or an automatic rejection? Does a consultant see excluded applications? What happens when the data is missing?
The Australian Human Rights Commission's recruitment guide recommends focusing on essential job requirements and avoiding irrelevant personal information. Establish those requirements with the client first. Our guide to essential requirements and client preferences covers that conversation. This audit follows the agreed requirement into the screening rule.
A screening-rule audit you can copy
Complete the first columns with the person responsible for the vacancy, then have the system owner explain the actual behaviour. Leave an unanswered question open. A field that looks optional on a form still needs checking if its absence changes whether the candidate progresses.
| Field or rule | Job-related reason and supporting evidence | How the screen uses it | Missing or conflicting data | Review and next decision |
|---|---|---|---|---|
| [Field, inferred feature or client instruction] | [Task being assessed; why this input helps assess it] | [Search, weight, threshold or rejection; exact setting] | [Blank, unknown, parser failure and conflicting records] | [Who sees exclusions; evidence needed; authorised decision] |
Add the configuration version and date. For a weighted input, ask how much influence it has and what the supplier can explain about that influence. For a hard rejection, ask whether one answer can end the application without another assessment.
The Merit Protection Commissioner's guidance advises APS agencies to minimise screening candidates out on a single AI-assisted assessment. That guidance is specific to the Australian Public Service, not a private-agency legal rule. It is a useful question to put to the owner of a hard screen: why should this result settle the decision on its own?
Keep the candidate's supplied information distinguishable from the system's interpretation. A CV with no parsed graduation date is different from a candidate saying they don't meet a qualification requirement. Decide what the reviewer should check before treating either as a failed criterion.
Worked example: a career-gap threshold
Consider this hypothetical rule: reject an application if an employment gap exceeds twelve months. It is an illustration of an audit, not a report about a customer or a finding that this threshold is unlawful.
The hiring manager's stated concern is whether applicants have current knowledge of a particular work process. The audit should investigate whether the gap rule actually assesses that knowledge, or whether another assessment would address the concern more directly.
| Audit question | What the team needs to establish |
|---|---|
| Why twelve months? | The role-related reason for the threshold and evidence supporting it |
| What counts as a gap? | Whether dates come from the candidate or a parser; how overlapping jobs, study and missing months are handled |
| What does rejection mean here? | Whether the application disappears from consideration or reaches a reviewer with the question unresolved |
| What alternative evidence could help? | A relevant work example, recent training or another agreed way to assess current knowledge |
| What will change? | An authorised decision about the rule, the assessment and any affected applications needing another look |
Don't ask candidates to disclose a diagnosis or family circumstances simply to help you categorise the gap. Ask for the evidence needed to assess the work. Escalate any question about collecting sensitive information or investigating group effects to the appropriate privacy and legal advisers.
The same method can be applied to other inputs. A location rule needs an explanation of the attendance requirement it is meant to assess. A graduation-year filter needs a reason for using that date rather than the relevant qualification or capability. These are questions to investigate, not a list of automatically unlawful fields.
What a changed result does, and doesn't, tell you
If an authorised test removes one input and the shortlist changes, record the change. You have learned that the result is sensitive to that input in the tested configuration. You have not, from that observation alone, established discrimination or proved that the revised screen is fair.
Keep the applicant pool and other settings consistent for a comparison where that is feasible. Record what could not be held steady, including a model or configuration change. Have a suitable reviewer inspect the affected applications against the job requirements, rather than assuming the new order is better because it differs.
Do not infer individual protected characteristics from a name, postcode or school to fill an audit spreadsheet. If a specialist assessment needs demographic information, have its collection and use reviewed before obtaining it. The Fair Work Ombudsman's discrimination guidance explains protections for employees and prospective employees, including decisions not to hire. It does not turn a homemade correlation check into a legal determination.
Close the audit with an instruction the recruiting team can use: retain the rule with its reason, change it through the authorised process, or stop relying on it while a specific question is reviewed. Record who will revisit applications affected by a correction. A changed setting alone leaves that work unfinished.
Correct the record before preparing the next change
An audit may reveal a factual error in a recruitment record as well as a problem with a rule. Keep those tasks distinct. The people responsible for selection criteria must decide how candidates are assessed; an AI-prepared record edit cannot settle that decision for them.
FluxHire's AI assistant prepares proposed changes to existing records and holds them for approval. During early access, record edits are made by people holding the records editor permission. An authorised reviewer can inspect the proposal before approving a change.
If you want to evaluate that assistance after a record error has been identified, bring the proposed correction and ask which records and fields are supported for your team. Check the source fact, the proposed change and the human review step. Do not assume FluxHire provides candidate ranking or this screening audit.
Request early access to discuss AI-prepared record changes your authorised team can review. Start with one correction you need to make, and the source material the reviewer should check.