The Fair Work Act and AI recruitment: you have to prove it wasn’t discrimination
Australian employers evaluating AI screening tools usually start with privacy. The provision that should actually decide the purchase is s.361 of the Fair Work Act, which presumes you discriminated unless you can prove you did not.
- Section 351 makes refusing to employ someone because of a protected attribute unlawful adverse action.
- Section 360 sets no dominance threshold — it is enough that the attribute was among the reasons.
- Section 361 reverses the onus: the reason is presumed, and the employer must prove otherwise.
- Compensation under s.545 is uncapped, and a court can order you to employ the applicant.
- A rejected applicant has six years, not 21 days — so decision records need a six-year floor.
Three sections that compound
Read on their own, the general protections provisions look like ordinary anti-discrimination law. Read together, they produce something considerably sharper.
| Section | What it does | Effect on an AI screen |
|---|---|---|
| s. 351 | Prohibits adverse action — expressly including refusing to employ — because of a protected attribute | Brings pre-employment screening squarely inside the regime, not just decisions about existing staff |
| s. 360 | An action is taken for a reason if the reasons include that reason | A protected attribute contributing even partly to a model’s output is enough; there is no "but for" or dominance test to hide behind |
| s. 361 | The reason is presumed unless the person proves otherwise | You start the case having already lost the factual question, and must affirmatively disprove it |
| s. 544 | Six-year general limitation period (the 21-day clock is dismissal-only) | A rejected applicant can bring a claim about a screen you ran six years ago |
| s. 545 | Remedies, uncapped, including an order to employ | The financial ceiling that caps most employment claims does not apply here |
What a reversed onus means when the decision-maker is a model
In a conventional claim, the employer defends by putting the decision-maker in the box. They explain what they saw, what they weighed, and why the applicant was not progressed. The tribunal assesses whether that account is credible. It is a well-worn path.
Substitute a model for the decision-maker and the path disappears. There is no witness. If the vendor cannot produce, for that specific candidate, what was asked, what evidence was cited for each score and what the system was structurally prevented from considering, then the employer has nothing to lead. And under s.361 having nothing to lead is not neutral — the presumption stands and the applicant wins.
The three defences that do not work
- 01The vendor is certified. Certification is evidence about the product in general. Section 361 asks about this candidate, on this date, on this requisition.
- 02The model does not receive protected attributes. Useful, and necessary, but not sufficient — proxies carry the attribute in without the label, and s.360 does not distinguish.
- 03A human reviewed the shortlist. Only helps if the record shows the human genuinely decided and why. A human who rubber-stamps a ranked list has added a signature, not a decision.
What does work is a contemporaneous, per-decision record. Not a dashboard, not an aggregate fairness score — the specific documentary trail for the specific candidate. That is the artefact the reversed onus is asking for, and it is the reason FirstPanel is built to produce it by construction rather than on request.
The evidence that actually discharges the burden
If you are specifying a screening system for the Australian market, specify the record first and the features second. A defensible file for one candidate contains all of the following, captured at the time and immutable afterwards.
- The rubric and criteria in force when the interview ran, versioned — so a later rubric change cannot be mistaken for the one applied.
- The questions actually asked, including any adaptive follow-ups the system generated.
- For every competency score, the verbatim, timestamped interview evidence that supports it — or an explicit abstention where evidence was insufficient.
- The criteria the system was forbidden to consider, and the rule set that forbade them.
- Model and prompt versions used for the assessment.
- The named human who made the disposition, their stated rationale, and any override of the system’s ranking.
- The adverse-impact position for that requisition at the time the shortlist shipped.
- The candidate-facing disclosure and consent, with the acknowledgement timestamp.
Six years, not twelve months
This is the single most commonly misconfigured setting in Australian hiring stacks. The 21-day time limit that HR teams know from unfair-dismissal matters is in s.366 and applies to dismissals. A person who was never employed was never dismissed, so it does not apply to them.
What applies instead is the general six-year limitation in s.544. A candidate rejected by an automated screen in 2026 can bring a general protections claim in 2032, and the presumption in s.361 will apply then exactly as it would today.
Most ATS retention defaults, and most vendor data-deletion schedules, are set to twelve or twenty-four months — often for good privacy-minimisation reasons in other jurisdictions. In Australia that setting quietly destroys your own defence. The correct posture is a deliberate, documented six-year floor for decision records specifically, with tighter minimisation for everything that is not part of the decision trail.
That distinction matters for the privacy side too: retaining a full video recording for six years is a different proposition from retaining a transcript, a score sheet and an evidence map. Structuring the decision record so the defensible part is small is what makes six-year retention proportionate.
What to do about it this quarter
- 01Audit retention. Find the deletion schedule for candidate assessment data and check it against six years for decision records.
- 02Ask each screening vendor to produce a full per-decision file for one real candidate. Not a sample report — a real one. The gap between the demo and the file is your exposure.
- 03Check whether criteria are free text. If a hiring manager can type an unlawful requirement into your system and it will run, you have a configuration-time hole no amount of downstream monitoring closes.
- 04Confirm scoring inputs in writing. Transcript content only is defensible; audio features and video frames invite a national-extraction and disability argument you will have to answer.
- 05Make the human review real. Give reviewers the evidence, require a written rationale, and record overrides. A reviewer who cannot see why a candidate scored the way they did cannot be your s.361 witness.
- 06Test proxies. Run adverse-impact analysis per requisition, at the four-fifths threshold, before shortlists ship — and keep the alerts and resolutions, because the fact you looked is itself evidence.