AI and Unfair Dismissal in Australia: What the Law Actually Says in 2026

Key Takeaways

• Australia has no standalone AI Act, and the National AI Plan released on 2 December 2025 confirmed the government would not proceed with mandatory guardrails for high-risk AI at this time. AI in the workplace is instead governed through existing laws, above all the Fair Work Act 2009 (Cth).
• The same unfair dismissal test applies whether a human manager, an algorithm, or a chatbot made the call. A dismissal can be found harsh, unjust or unreasonable, and an employer cannot hide behind a tool to avoid procedural fairness.
• From 20 October 2026, the Fair Work Commission's AI Guidance Note requires anyone who uses generative AI to prepare a document lodged in a Commission case to disclose that use, verify every fact and citation by a human, and keep witness statements in the witness's own words.
• Recent cases show both sides of the risk. In Lord v Millet Hospitality [2025] FWC 2740, a ChatGPT-drafted redundancy email helped make a dismissal unfair. In Khan v Aldi [2026] FWC 3144, an AI-assisted claim with no reasonable prospects ended in a $1,230 costs order against the worker.
• From 10 December 2026, larger employers covered by the Privacy Act 1988 (Cth) must also disclose in their privacy policies where computer programs, including AI tools, make or help make decisions that could significantly affect a person.

There Is No AI Dismissal Law — and That Is the Point

Anyone searching for ai unfair dismissal australia regulation is likely expecting a dedicated statute, something like a European-style AI Act with a chapter on employment. That statute does not exist in Australia, and understanding the topic starts with accepting that fact rather than working around it.

The Australian Government released its National AI Plan on 2 December 2025. According to legal analysis of the plan, it confirmed there would be no standalone AI Act and that the proposed mandatory guardrails for high-risk AI would not be introduced. Instead, the government chose to manage AI risk through existing, largely technology-neutral laws, reviewed where gaps appear, with regulators keeping responsibility for harms inside their own areas and a new AI Safety Institute set up to monitor, test and share information about AI capabilities and harms. The September 2024 consultation had proposed ten mandatory guardrails for high-risk settings, covering accountability, risk management, testing, transparency, human oversight and records. Those were proposals under consideration, not an enacted regime, and the December 2025 plan set them aside.

That choice shapes everything that follows. An AI tool that screens candidates, scores performance, drafts a termination letter or deactivates a gig worker does not sit in a legal vacuum, but neither does it sit under one AI employment law. It sits under the Fair Work Act, anti-discrimination legislation, privacy law, work health and safety law, and the rules of the tribunal that hears the dispute. The question in an unfair dismissal case is therefore rarely "did the employer use AI?" It is the older, harder question: "was this dismissal fair, and was the process fair?"

The Unfair Dismissal Test Does Not Change Because AI Was Involved

Under Part 3-2 of the Fair Work Act, a dismissal can be found unfair if it was harsh, unjust or unreasonable. The Fair Work Commission considers, among other things, whether there was a valid reason for the dismissal, whether the employee was notified of that reason, whether they were given an opportunity to respond, and whether proper procedure was followed. Nothing in that test distinguishes between a decision made by a human manager, one recommended by an algorithm, or one generated entirely by an AI system.

That matters in two directions. For workers, it means AI involvement does not remove their protections. If an AI system recommended termination and the employer simply rubber-stamped it without proper investigation or giving the worker a chance to respond, that is likely to be an unfair dismissal, because procedural fairness cannot be outsourced to technology. For employers, it means the opposite error is equally dangerous: assuming that because a system produced a score, a ranking or a draft letter, the decision is somehow objective, documented and defensible. A score is not a reason. A ranking is not a consultation. A generated letter is not a conversation.

The basic gateways to a claim have not moved either. According to a summary by the Fair Work Centre, a claim must be lodged within 21 days of the dismissal taking effect, the worker must have completed the minimum employment period — six months for employers with 15 or more staff, and 12 months for small businesses — and must stay below the high income threshold of $190,100 unless covered by an award or enterprise agreement. Compensation, where awarded, stays capped at the lesser of six months' pay or $95,050. Those figures are reported thresholds and caps, and they should be checked against the Commission's current materials before anyone relies on them, but the structure is the point: AI has changed how claims are prepared and how dismissals happen, not who can claim or how quickly they must act.

Genuine redundancy has its own test, and AI complicates it rather than simplifying it. Section 389 of the Fair Work Act sets a three-limb test: the job is no longer required to be performed by anyone because of changes in the operational requirements of the enterprise; the employer has complied with any consultation obligation in an applicable modern award or enterprise agreement; and it was not reasonable in the circumstances for the employee to be redeployed. AI may drive the operational change — process automation or AI replacing parts of a role — but it does not dilute the consultation and redeployment limbs. In Helensburgh Coal Pty Ltd v Bartley [2025] HCA 29, the High Court confirmed that when considering redeployment, the Commission can scrutinise how an employer structures its labour model, including reducing reliance on contractors or labour hire, to create or free up roles for otherwise redundant employees. An employer that lets a model decide whose role disappears, then treats consultation as a formality, is taking a risk the statute does not forgive.

The Case That Warns Employers: ChatGPT Wrote the Redundancy Email

The clearest Australian warning to employers so far is Hayley Lord v Millet Hospitality Geelong Pty Ltd [2025] FWC 2740. A hospitality business used ChatGPT to draft an email informing an employee that "we have made the difficult decision to remove the Housekeeping Supervisor position," before offering a "discussion" about alternatives and saying it would "proceed with the removal… as planned" if she did not respond. The email was sent after business hours, and no face-to-face conversation about the redundancy took place.

Commissioner Redford found that the wording showed a final decision had already been made, so any consultation offered was illusory, and that relying on ChatGPT did not excuse the employer's failure "to adhere to basic standards of decency" or to have a face-to-face conversation about redundancy. The Commissioner did not accept that the size of the business, or the absence of human resources expertise, excused the failure to speak to the employee face to face about something as significant as the redundancy of her role. Because the required consultation had not happened before the decision was implemented, the dismissal did not satisfy the meaning of genuine redundancy under the Fair Work Act. The dismissal was ruled unfair, and the employee was awarded compensation for her loss of income, in the form of pay and superannuation equivalent to what she would have received if the proper procedure had been followed.

The lesson is not that employers must never use AI to draft correspondence. It is that AI can help draft, but it cannot own the message. A generated draft that announces a decision as already made, in language the sender did not stop to read, converts a consultation obligation into evidence against the employer. Every AI-drafted employment letter needs a human to read it as the recipient will read it, to check what it actually says about timing and finality, and to ask whether the conversation the law requires has actually happened. Where automation displacement is handled clumsily, employers also face potential work health and safety exposure on top of unfair dismissal, general protections or discrimination claims, so the cost of a careless automated process is not confined to one jurisdiction.

The Case That Warns Workers: An AI Claim That Ended in Costs

AI has also transformed the other side of the tribunal. The Fair Work Commission has reported a marked increase in lodgements, rising from just over 30,000 in 2022–23 to approximately 40,000 in 2023–24, and exceeding 44,000 in 2024–25, with projections of between 50,000 and 55,000 in 2025–26. Commission President Adam Hatcher has linked a 70 per cent rise in the Commission's total workload over three years largely to AI-assisted unfair dismissal and general protections applications, and one law firm tracker reports that applications in 2025–26 reached a record of more than 53,000 cases. The Commission's research has reported instances of AI recommending the wrong type of claim, overlooking eligibility requirements and filing deadlines, inventing cases, and giving inflated predictions of success.

Two 2026 decisions illustrate how differently that can end. In Baker v Macquarie University [2026] FWC 3054, a casual academic who was self-represented used premium AI tools, including ChatGPT Pro and Claude, to prepare and conduct his case, and succeeded. The decision was reported as the first in the country to apply the employee choice casual conversion provisions. Used carefully, in other words, AI can help a self-represented person put a real case.

One week later, in Khan v Aldi [2026] FWC 3144, the outcome was the opposite. A store assistant hired on 29 September 2025 was notified of his dismissal on 26 March 2026, three days short of the six-month minimum employment period under sections 382 and 383 of the Fair Work Act. He lodged an AI-drafted unfair dismissal application and persisted with AI-generated arguments after warnings. The Commission characterised his argument as "just plain wrong" and ordered him to pay $1,230 towards Aldi's costs under section 400A of the Fair Work Act, at the maximum rates under the Fair Work Regulations. Costs orders are uncommon in the Commission, which makes this one notable: it has been widely reported as the first FWC costs order of its kind arising from unreasonable reliance on AI. The contrast between the two cases is the whole story. AI assistance is not itself unreasonable. Failing to check what the AI produced, and pressing on after the error has been pointed out, can be.

The FWC's New AI Guidance Note: Disclose, Verify, Own It

The Commission's response to this surge is now in force. Its AI Guidance Note was published on 24 August 2026, with the new obligations commencing on 20 October 2026, and it applies to any document lodged in a Commission proceeding, irrespective of the type of application. It sets out three principal requirements for any party that uses generative AI to help prepare a document to be lodged, including an application or an employer response.

First, disclosure. The document must state that generative AI was used and how it was used, whether for legal research, organising facts and evidence, transcribing speech, translating languages, or writing or editing content. Second, verification. Every fact, authority, quotation and statutory provision must be checked for accuracy and existence by a person, rather than by another AI tool, and the lodged document must state that verification has occurred. Professional representatives are required to include hyperlinks to all cases referred to in the document, while self-represented litigants are encouraged to do the same. Third, authenticity. Witness statements and declarations must genuinely reflect the witness's own knowledge and own words, and must be true to the best of their knowledge. Generative AI must not be used to create the factual content of witness statements.

The consequences of non-compliance are serious. A document may be given less weight or disregarded, an adverse costs order may be made, or the case may be dismissed altogether. The guidance also cautions against entering personal, confidential or legally privileged information into generative AI tools, and warns of the risk of breaching Commission orders that prohibit or restrict publication of certain information if that information is entered into a tool. Knowingly providing false or misleading information to the Commission is a criminal offence carrying up to 12 months' imprisonment under section 137.1 of the Criminal Code. The Commission has framed the guidance as an initiative to encourage responsible AI use rather than as a punitive measure, and properly disclosed AI use will not, of itself, affect the Commission's assessment of the merits of a case. But the message is unambiguous: responsibility for the accuracy, relevance and legal soundness of anything lodged remains with the person who lodges it.

The guidance arrived alongside procedural reforms. The Workplace Relations Legislation Amendment (Building Cooperative Workplaces No. 1) Act (Cth) commenced on 7 July 2026, introducing changes to the Commission's procedural framework, including broader discretion to determine matters on the basis of written submissions alone, with the consent of the parties. The direction of travel is clear: more claims, more AI-assisted documents, and a tribunal adapting its procedures to manage both.

Automated Dismissal and the Gig Economy: The Uber Eats Ruling

The most striking recent decision on automated decision-making itself is not a traditional unfair dismissal case at all, but a gig worker deactivation. On 10 April 2026, a three-member Full Bench of the Fair Work Commission upheld the appeal of Canberra delivery driver Umair Ayyub against Portier Pacific Pty Ltd, trading as Uber Eats. The Commission ruled his deactivation was unfair, ordered his reactivation, and found a lost pay order appropriate, directing the parties to confer on the amount.

The facts matter because they show what "automated" actually looked like. His deactivation on 8 July 2025 followed automated warnings triggered by his failure to maintain an 85 per cent minimum satisfaction rating, a threshold introduced in late 2024 and not communicated to him until 21 February 2025. The rating system relied on a simple thumbs-up or thumbs-down from customers and merchants, and only 7.2 per cent of all deliveries attracted any rating at all. The Full Bench found that the process up to and including the issuing of the final notice was "entirely automated; no human is involved in any decision-making in respect of the process." The deactivation warnings advised him he could seek support from a person, but failed to specify that this included a union delegate or employee, which the Full Bench found was required under the applicable Code. The decision was also made on 3 June 2025, but the final notice was not sent until 8 July 2025, a gap of over five weeks.

The ruling crystallises several obligations: performance standards and measurement methods must be communicated before they are applied; a worker's response to a potential termination must be genuinely considered by a human with real authority to affect the outcome; termination decisions must be communicated as soon as reasonably practicable after they are made; and deactivation warnings must expressly state that workers can seek assistance from a union delegate or employee. The case was brought under provisions of the Fair Work Act introduced by the Fair Work Legislation Amendment (Closing Loopholes No. 2) Act 2024, which extended protections against unfair deactivation to gig workers. It overturned a January 2026 ruling by Deputy President Dean and departed from two earlier Commission decisions that had found the 85 per cent threshold reasonable. For any employer running performance-based exits through an automated system, the message is direct: human judgment cannot be replaced by algorithms in performance-based dismissals, and a process with no human in it is not a process the Commission will accept.

Discrimination, Privacy and the Other Laws That Still Apply

Unfair dismissal is only one of the legal frames that can catch an AI-driven employment decision. Australian anti-discrimination law under the Fair Work Act, the Age Discrimination Act 2004, the Disability Discrimination Act 1992, the Racial Discrimination Act 1975 and the Sex Discrimination Act 1984 all prohibit discriminatory practices. If an AI system produces discriminatory outcomes — even unintentionally, for example by screening out candidates or flagging workers on the basis of a proxy for age, disability, race or sex — the employer remains liable. There is no defence that the model did it. The same logic applies to general protections claims: adverse action taken because of a protected attribute or workplace right is unlawful regardless of whether a tool recommended it.

Privacy law is about to add a concrete disclosure duty. From 10 December 2026, new Australian Privacy Principles 1.7 to 1.9, added by the Privacy and Other Legislation Amendment Act 2024, require entities covered by the Privacy Act — broadly, Australian Government agencies and private sector organisations with annual turnover of more than $3 million, plus some smaller businesses such as health service providers — to disclose in their privacy policies where they use a computer program to make, or help make, decisions that could significantly affect a person, using that person's personal information. The Office of the Australian Information Commissioner published final guidance on 30 September 2026, and it reads "computer program" broadly: it covers simple rule-based processes, machine learning, generative AI and chatbots, and everyday software. Human-in-the-loop review does not, on its own, remove the requirement to disclose. The OAIC's examples include recruitment software used to sort candidates and make hiring decisions, and AI-generated reports used to rank employee performance or determine promotions, bonuses or remuneration. The obligation is primarily a transparency requirement rather than a general prohibition on automated decisions, but for an employer already defending a dismissal, a privacy policy that says nothing about the system that recommended it is an avoidable problem.

What Employers Should Do Before an AI Tool Touches a Dismissal

The practical response follows from the cases, not from any single rule. First, keep a human with real authority in the decision. Someone who can change the outcome must consider the worker's response, look at the underlying evidence rather than the score, and record why they agreed or disagreed with the tool. A rubber stamp is not human oversight, and the Uber Eats ruling shows the Commission will look at whether a human was actually involved, not whether one was nominally available.

Second, communicate the standard before applying it. A performance threshold, a rating system or a productivity metric that a worker learns about only when it is used against them is vulnerable on fairness grounds alone. Tell workers what is measured, how it is measured, and what happens if they fall below it, before any of it is used in a decision.

Third, treat AI-drafted correspondence as a first draft, never as a finished letter. Check what it actually says about whether a decision is final, whether consultation is genuine, and whether the tone matches what the law requires. Have the conversation the law requires — face to face where the decision is as significant as a redundancy — rather than sending an email after hours and calling that consultation.

Fourth, audit the tool for discriminatory effects, not just for accuracy. A model that is accurate on average can still systematically disadvantage a group, and the liability for that sits with the employer. Keep records of what data the tool used, what it recommended, what the human decision-maker did with that recommendation, and why. If the matter reaches the Commission, those records are the difference between a defensible process and an unexplained one.

Fifth, prepare for the Commission's AI rules as well as for the dismissal itself. If generative AI helped prepare any document to be lodged — the employer response, a submission, a witness statement — disclose the use and how the tool was used, verify every fact, case and quotation by a person, hyperlink cited cases, and keep witness statements in the witness's own words. Do not put personal, confidential or privileged material into a public AI tool.

What Workers Should Do If AI Was Involved in Their Dismissal

Workers have a symmetrical checklist, and the first item is time. An unfair dismissal application must generally be lodged within 21 days of the dismissal taking effect, and that clock does not pause while a worker waits to see whether an internal appeal or a conversation with the employer resolves things. If AI was involved, the worker should ask, in writing, what system was used, what data it relied on, what it recommended, and who made the final decision and on what basis. Those questions go to the heart of the Commission's test: valid reason, notification, opportunity to respond, and proper procedure.

Workers should also be careful about the AI they use themselves. The Khan v Aldi costs order is the cautionary tale: an AI tool can draft an application quickly, but it can also invent a case, miss a jurisdictional requirement such as the minimum employment period, or give an inflated prediction of success. The Commission's research has documented all of those failures. Anyone using AI to prepare a claim should verify every citation against a reliable source, check the eligibility requirements and the deadline personally, and get independent legal advice before lodging, particularly before persisting with an argument after it has been challenged. Generative AI is not a substitute for legal advice, and the Commission has said so expressly.

It is also worth remembering that AI can help a genuine claim. The Baker decision shows a self-represented applicant using AI tools effectively to prepare and conduct a case that succeeded. The difference between that outcome and the costs order in Khan was not the use of AI. It was verification, judgment and knowing when to stop.

Where This Is Heading

The broader context explains why searches for ai unfair dismissal australia have spiked. The volume of claims is rising sharply, the tribunal has responded with binding disclosure and verification duties, the gig economy now has unfair deactivation protections, and a privacy transparency duty for automated decisions takes effect on 10 December 2026. At the same time, the national policy choice has been to regulate AI through existing laws rather than through a dedicated AI statute, which means the Fair Work Act will go on doing the heavy lifting, case by case.

That is likely to mean more decisions like Lord, Khan and Ayyub — decisions that turn less on the technology than on whether a human took responsibility for it. Employers who treat AI as a decision-maker will keep losing. Employers who treat it as a drafting and analysis aid, supervised by a person who can explain and defend the outcome, will be in a far stronger position. Workers who treat AI as a lawyer will keep being disappointed. Workers who use it as a research aid, verify everything, and meet the 21-day deadline will be able to make their case heard.

For readers tracking how AI regulation is developing beyond the workplace, the UK has taken a different institutional route, creating a dedicated unit for AI agents that go wrong, as covered in UK Budget 2026: A £40 Million Unit for AI Agents That Go Wrong. The employment effects of AI are also visible in other markets, including the UK figures examined in AI Impact on UK Employment: 2026 Data and Job Outlook and the bank restructuring covered in NatWest AI Job Cuts Explained: 1,000 Roles Gone, New AI Roles Created. The Australian story is distinct, but the underlying question is the same everywhere: when a machine recommends that someone loses their job, who answers for it?

The Australian answer, as the law stands in November 2026, is that a person answers for it — the employer who relied on the tool, and the party who lodged the document it produced. That is the rule to plan around, whether you are drafting the next redundancy letter, preparing the next claim, or deciding whether to let a system make a decision that changes someone's livelihood.

This article is general information about Australian law, not legal advice. Employment disputes turn on their specific facts, and anyone facing a dismissal or a claim should seek independent legal advice promptly, particularly given the 21-day time limit for unfair dismissal applications.

Sources

• Fair Work Commission, Guidance Note on the use of generative AI in Commission cases, published 24 August 2026, effective 20 October 2026, as reported by Clayton Utz and MinterEllison.
• Sadnan Khan v Aldi Pty Ltd [2026] FWC 3144 and Gregory Baker v Macquarie University [2026] FWC 3054, as reported by MinterEllison and HR Director.
• Hayley Lord v Millet Hospitality Geelong Pty Ltd [2025] FWC 2740, as reported by McCullough Robertson and CG Law.
• Uber Eats deactivation ruling, Full Bench, 10 April 2026, as reported by HR Director.
• National AI Plan, 2 December 2025, as analysed by Piper Alderman via Legal500.
• Privacy Act automated decision-making requirements, APP 1.7 to 1.9, effective 10 December 2026, as reported by Russell Kennedy.