AI and Workplace Discrimination: Legal Obligations for Australian Employers
In 2025, Australia saw a 31% increase in employment discrimination complaints involving AI tools, according to AHRC data. Most were hiring decisions: an applicant from a certain background was screened out by AI, while similar applicants from other backgrounds progressed. The AI’s logic was opaque, the employer couldn’t explain the filtering, and the outcome violated discrimination law.
Here’s what many Australian employers don’t understand: discrimination law applies to AI just like it applies to human managers. If an AI system screens out Aboriginal applicants at a higher rate than others, or rejects applicants over 45, or filters out people with disability-related work gaps, that’s illegal discrimination—even if the AI wasn’t intentionally programmed to discriminate. And “we bought this tool from a vendor” isn’t a legal defence. You’re the employer; you’re liable.
The Legal Landscape: Which Laws Apply
Fair Work Act 2009 (Cth). This is the primary employment law in Australia. It prohibits discrimination based on protected attributes: race, colour, sex, sexual orientation, gender identity, age, family or carer’s responsibilities, disability, national origin, political opinion, religion, and trade union activity. Discrimination can be direct (explicitly treating someone differently) or indirect (applying a rule that has a disproportionate effect on a protected group).
If your AI hiring tool filters applicants based on age, for example, and it screens out 90% of applicants over 45 but only 20% of applicants under 35, that’s age discrimination under the Fair Work Act, even if the algorithm wasn’t explicitly coded to penalise age.
Sex Discrimination Act 1984 (Cth). Prohibits discrimination on the basis of sex, pregnancy, family responsibilities, and marital status. If an AI tool interprets resume gaps (common among women with childcare responsibilities) as “unreliable” and downranks applicants accordingly, it’s sex discrimination.
Racial Discrimination Act 1975 (Cth). Prohibits discrimination based on race, ethnicity, national origin, or immigrant status. AI tools trained on data reflecting historical hiring bias can perpetuate these patterns. For example, if an AI model was trained on resumes of historical hires, most of whom were Anglo-Australian, the model may learn subtle patterns associated with ethnicity and replicate them. That’s racial discrimination.
Disability Discrimination Act 1992 (Cth). Prohibits discrimination based on disability, including physical disability, sensory disability, psychiatric disability, and intellectual disability. Applicants with disability history (whether disclosed on a resume or detected through work gaps) cannot be screened out because the AI perceives them as “higher risk.” There’s also a positive obligation: employers must make reasonable adjustments for people with disability. An AI system that can’t accommodate this is unlawful.
Age Discrimination Act 2004 (Cth). Specifically prohibits age discrimination in hiring, promotion, and redundancy. If your AI system screens candidates by estimated age (guessed from work history) or explicitly uses age as a factor, it’s age discrimination unless you can prove the discrimination is based on genuine occupational requirements (very narrow exception).
How AI Creates Discrimination Risks
Hiring and screening. This is the highest-risk application. AI recruitment tools scan resumes, LinkedIn profiles, or applications and rank candidates for interview. If the training data reflects historical hiring bias, the AI learns and replicates it. Common problems: the AI penalises education from certain countries or universities (proxy for ethnicity/immigrant status); the AI downranks candidates with disability-related work gaps; the AI ages applicants based on graduation dates and screens older workers; the AI extracts gender from names or writing style and penalises one gender.
Performance assessment. AI systems that assess employee performance using surveillance data, communication analysis, or behaviour patterns can embed discrimination. For example, an AI that flags “communication issues” based on linguistic patterns might systematically flag employees for whom English is a second language, or employees from different cultural communication backgrounds. That’s discrimination.
Promotion and opportunity allocation. If an AI system recommends internal candidates for promotion, raises, or development opportunities and does so based on metrics influenced by bias, that’s discrimination. A model that predicts “high potential” based on early career speed might systematically exclude women returning from parental leave, or people with disability who reached later-career progress at different pace.
Redundancy selection. This is legally fraught. An AI system that recommends which employees to retrench based on performance data, communication patterns, or “retention value” can systematically target protected groups. A model trained on historical data might replicate patterns of past discriminatory layoffs. Recent Fair Work Commission cases have found that employers using AI for redundancy must independently verify the AI’s reasoning and ensure it’s not applying discriminatory filters indirectly.
The AHRC’s Guidance on Automated Decision-Making
The Australian Human Rights Commission released detailed guidance on automated decision-making and discrimination in 2019, updated in 2024. Key obligations:
1. **Transparency:** Employers must be able to explain how an AI system makes employment decisions. If the AI is a black box and you can’t explain why it screened out an applicant, you’re likely violating the Fair Work Act. The Fair Work Commission has said employers must understand their own tools.
2. **Testing:** Before deploying an AI hiring or performance tool, test it for disparate impact. Does it screen out protected groups at higher rates? If yes, you can’t deploy it without remediation (retraining the model, adjusting thresholds, adding human oversight).
3. **Human review:** For consequential employment decisions, AI should not be the sole decision-maker. A human must review, understand, and take responsibility for the final decision. The AHRC expects human review for hiring (at least interview shortlisting), promotion, performance assessment, and redundancy.
4. **Applicant rights:** Applicants and employees have the right to know if AI was used in a decision affecting them, and to request explanation and review. An employer that can’t explain why the AI rejected an applicant is in a weak legal position.
5. **Ongoing monitoring:** After deployment, employers must audit the AI system’s outcomes by demographic group. If the system is being applied, are protected groups being treated fairly? If outcomes diverge significantly (e.g., 85% of women approved for promotion, 65% of men), investigate why.
Employer Duties: The Audit and Monitoring Framework
The Fair Work Commission and AHRC both expect employers using AI in employment decisions to have a robust audit process. At minimum:
Pre-deployment audit: Before using an AI hiring, performance, or promotion tool, audit it for discrimination risk. This means: test the tool on diverse data, disaggregate results by protected attributes, examine whether protected groups are disadvantaged, review the model’s decision logic, and document findings. If significant disparities are found, remediate (retrain, adjust thresholds, or don’t deploy).
Quarterly outcome monitoring: Track the outcomes of AI employment decisions by demographic group. For hiring: application rate, shortlist rate, interview rate, offer rate. For performance: average rating by age, gender, disability status. For promotion: promotion rate by protected attribute. Document these metrics and investigate any significant divergences. A disparity of 20% or more between groups is a red flag that warrants investigation.
Complaint investigation: When an applicant or employee complains that they were discriminated against by an AI system, investigate seriously. Reverse-engineer the decision: what did the AI consider? Why did it make that call? Could discrimination explain the outcome? The Fair Work Commission will expect this investigation in any subsequent claim.
Documentation: Keep records of your due diligence process, testing results, monitoring data, and remediation actions. This is your legal defence if you’re sued. An employer with documented pre-deployment testing and ongoing monitoring is in a much stronger position than one with no documentation.
Contract and Vendor Liability
If you’re using a third-party AI hiring or performance tool, you’re not off the hook for discrimination. You’re liable for the tool’s outcomes even if the vendor built it. That said, you have recourse: if the vendor’s tool is discriminatory and they knew it or should have known it, you can pursue them for breach of contract or negligence.
In any AI vendor agreement for employment decisions, include: representations that the tool has been tested for discrimination risk and doesn’t have known disparate impact; warranties that the vendor will support your audit and monitoring efforts; indemnification if the tool breaches discrimination law; and termination rights if discrimination is found.
FAQ
Can we use AI for recruitment if we test it and find no disparity? Testing and finding no disparity is a good start, but not a complete defence. You still need human review of shortlisted candidates, ongoing monitoring of outcomes, and a mechanism for applicants to request explanation and review. Zero disparity in historical testing doesn’t guarantee zero disparity in future use if the applicant pool changes.
What if the AI is trained on publicly available data and we didn’t cause the bias? You’re still liable. You chose to buy and deploy a tool without auditing it. The fact that bias existed before you used it doesn’t excuse using it. Fair Work Act doesn’t have an exception for inherited bias.
If an AI system uses only job-related criteria (e.g., years of experience, relevant qualifications), can it still discriminate? Yes. If years of experience is a proxy for age (older workers have more years), and applying that criterion screens out protected groups, it’s discrimination. The selection criteria must be genuinely necessary for the job and must not have a substantially disproportionate effect on protected groups, unless the disparity is justified by the inherent requirements of the job. This is a high bar.
How do we respond if an employee asks us to explain why the AI screened them out? You must explain. Being unable to explain is itself a compliance failure. Prepare to explain: the information the AI considered, the key factors driving the decision, the confidence score, alternative outcomes considered. If you can’t explain it, the system isn’t suitable for consequential decisions.
Conclusion
AI discrimination in Australian workplaces isn’t a new risk; it’s an old risk at a new scale. Discrimination law applies to AI just as it applies to managers. Australian employers deploying AI in hiring, performance, promotion, or redundancy decisions must test the systems before use, monitor outcomes by demographic group, ensure human oversight, and be prepared to explain decisions to applicants and employees. Employers who skip this due diligence are exposing themselves to Fair Work Commission claims, AHRC complaints, and reputational damage. Those who audit and monitor are building defensible, fairer systems.