From 20 October 2026, Fair Work Commission AI disclosure rules require anyone who used generative AI on a filing to say so and confirm they checked every fact and citation. The Commission has already ordered one applicant to pay the other side's costs after AI-drafted submissions ran a hopeless case. For any Australian firm, unverified AI output in front of a tribunal, regulator or client is now a governance failure with a price attached.
The Fair Work Commission (FWC) is Australia's national workplace tribunal. It handles unfair dismissal, general protections (adverse action for exercising a workplace right) and bullying claims. If your firm employs people, you can end up there as a respondent, and small businesses often appear without a lawyer.
What has the Fair Work Commission changed on AI disclosure?
On 24 August 2026 the Commission's President published a guidance note on the use of generative AI in Commission cases. It takes effect on 20 October 2026 and applies to all Commission matters.
The Commission's own plain-language summary sets out three obligations. They apply to anyone who uses a tool such as ChatGPT, Claude, Copilot or Gemini to create, write or edit a document:
Disclose. State in the document that generative AI was used, and explain how.
Verify. Check the document yourself so that every fact, piece of evidence, piece of legislation, case and quote is accurate, relevant and real.
Confirm. State in the document that you have completed those checks.
Witness statements and declarations carry an extra test. The witness must make sure the document is based on their own knowledge, reflects their own words, and is true to the best of their knowledge. Legal practitioners and paid agents must also include hyperlinks to any case law cited in a document prepared with AI.
The same page spells out the consequences. Documents may be given less weight or disregarded. The Commission may order you to pay the other party's costs as well as your own. The application or case may be dismissed. Application and response forms are being updated by 20 October to prompt compliance.
Why is the Commission cracking down now?
Commission-funded research, reported by the Australian Broadcasting Corporation (ABC) on 29 August 2026, found generative AI partly responsible for a 40 per cent surge in cases from 2023-24 to 2024-25. Roughly four in ten surveyed cases involved a litigant using AI. More than three-quarters of those users relied on ChatGPT, and 60 per cent of them were on the free version.
The law firm Kingston Reid notes that President Justice Adam Hatcher put the Commission's normal caseload at around 30,000 matters a year until about 2023. He named generative AI as a significant driver of the increase since. His concern was specific: AI tools have a common tendency to make a claim with no reasonable prospect of success look somewhat plausible.
The same ABC report records the Commission acknowledging that AI is improving access to justice for people with genuine claims. The guidance note is a set of conduct rules, not a ban: it targets unverified output, not AI use itself.
What happened in the Khan v Aldi costs order?
The case that drew national attention was Khan v Aldi Pty Ltd [2026] FWC 3144, decided by Deputy President Michael Easton in Sydney on 19 August 2026. A worker was dismissed three days short of the six-month minimum employment period needed to bring an unfair dismissal claim. His AI-generated submissions repeatedly focused on the wrong dismissal date and raised arguments the Commission found irrelevant. His original application still contained the AI's own drafting note, quoted in the decision: "Here is your final clean copy-paste version".
Deputy President Easton warned him several times that the claim was unlikely to succeed and that continuing could attract a costs order. He kept filing, then discontinued the application at the hearing. The result, as reported by the ABC, was an order to pay A$1,230 towards Aldi's legal fees. The Deputy President called it "the very rare step of ordering him to pay some of the employer's legal costs, because his unreasonable conduct caused his former employer to incur those costs unnecessarily". He also found the applicant had used AI as a "quasi-legal advisor".
Two details matter. First, the decision notes the material was "not complete AI slop": some arguments were more or less correct, though irrelevant. Nobody had checked whether the output answered the actual problem. Second, the costs order flowed from unreasonable conduct after warnings, not from using AI as such.
An earlier decision, Reece Hoverd v M & J D Pty Ltd [2026] FWC 1013, decided on 25 March 2026, shows the credibility risk. The applicant relied on AI-drafted submissions that quoted contract clauses which did not exist, and kept relying on them after being corrected. The Commission upheld the employer's jurisdictional objection and recommended a costs application. Its response, as summarised by South Geldard Lawyers, was blunt: "Ignorance of the law is one thing. But repeatedly saying the written contract says something which it does not is not mere ignorance; it is a deliberate misrepresentation."
| Case | What the AI output got wrong | Consequence |
|---|---|---|
| Khan v Aldi Pty Ltd [2026] FWC 3144 | Argued the wrong dismissal date; ignored the six-month eligibility problem despite warnings | Application discontinued; A$1,230 costs order |
| Hoverd v M & J D Pty Ltd [2026] FWC 1013 | Quoted contract clauses that did not exist, and kept relying on them | Objection upheld; finding of deliberate misrepresentation; costs application recommended |
Does this apply to employers, or only to sacked workers?
It applies to both. The guidance note covers all Commission matters, and the Commission is updating response forms as well as application forms. A firm that pastes an unfair dismissal claim into a chatbot, asks for a response, and files the result unchecked is in the same position as the applicants above.
Kingston Reid also advises employers to learn the hallmarks of AI-generated claims. Watch for formulaic structure, generic references to statutory criteria, and case law that does not exist or does not support the proposition. From 20 October, an undisclosed AI-drafted claim showing those hallmarks can be raised with the Commission.
Is the Commission the only body demanding this?
No, and that is the point for firms outside employment disputes. The Federal Court published its Use of Generative Artificial Intelligence Practice Note (GPN-AI) on 16 April 2026. Where AI has been used, the Court expects the responsible person to have confirmed that cited authorities exist and support the stated proposition. Cited evidence must exist and be likely to be admissible, and pleaded facts must be reasonably provable.
Australian courts have already referred practitioners to their regulators over unverified AI citations. The law firm Hicksons summarises Dayal [2024] FedCFamC2F 1166 and Mertz & Mertz (No 3) [2025] FedCFamC1A 222. Solicitors and barristers were referred to their regulators after filing lists of authorities containing cases that did not exist.
The same logic extends beyond tribunals. It covers a letter to the Australian Taxation Office, a response to the Australian Securities and Investments Commission (ASIC), or advice to a client. The person who signs the document owns its accuracy, whether the draft came from a chatbot, a junior staff member or an AI automation partner in Sydney. An accountant who sends a client an AI-drafted tax summary with a fabricated ruling number has the same credibility problem as the applicants above.
What does a verification and disclosure workflow look like?
A workable process for a firm of 5 to 50 people has six parts.
Decide which tools are approved and for which tasks. Drafting a first-pass chronology is low risk. Generating legal authorities is high risk. Write the distinction down.
Adopt a standard disclosure line. Something like: "This document was prepared with the assistance of generative AI (tool name) for drafting and structure. All facts, dates, quotes and legal references have been checked against source documents by (name) on (date)."
Run a fixed verification checklist. Every name, date, quote, number, attachment reference and legal proposition is checked against the original record. Every case is opened, read and hyperlinked using the Commission's own document search or the Australasian Legal Information Institute (AustLII).
Keep the prompt and the output. If a filing is later questioned, showing what was asked and what was checked is the difference between an honest error and a credibility finding.
Require a named human sign-off. The person whose name appears on the document does the final read. Witness statements are written in the witness's own words, then tidied, never generated from a summary.
Build it into the workflow, not the memo. In 30 years running transformation programmes at Nokia, Orange, GE and National Australia Bank (NAB), I never saw a control survive that lived only in a policy document. Firms that put the checklist into the template, the document management system, or an AIOS (AI Operating System) that already holds their client and matter data get compliance by default. Hyperlinked, verified authorities belong in the drafting workflow of a law firm, not in a training slide.
The Commission has given every Australian employer eight weeks' notice of a standard that courts and regulators are converging on anyway. Firms that treat verification and disclosure as part of how documents get made, rather than as a rule to remember under pressure, will find 20 October 2026 arrives without incident.
Drafted with generative AI assistance for structure and first draft; every date, figure, quote and citation was checked against the linked sources by Phil Bishop on 5 September 2026.
If you want a second pair of eyes on how AI-drafted documents move through your firm before they reach a tribunal, regulator or client, book a free initial AI consultation. We will map these checks against your current workflow.