Part of the AI Regulation News hub.
Australia's Administrative Review Tribunal made a practice direction on 20 August 2026 requiring anyone in its proceedings to verify generative AI output, disclose its use, certify that on request, and explain how it was used
The duties are four and they are cumulative: verify, disclose, certify on request, explain on request. The Tribunal states in the document that it adapted this from the Federal Court's own practice note with the Chief Justice's permission, which places it third among Australian forums rather than first. The sharpest line is the responsibility rule, which does not depend on whether AI was used at all.
Bottom line: A practice direction, in force from the day it was signed, 20 August 2026. It applies to every person involved in proceedings across all of the Tribunal's jurisdictional areas, and to applications lodged before as well as after that date.
Who this affects: Parties, legal and other representatives, litigation supporters and anyone else involved in an ART proceeding. Representatives carry the extra exposure: the Tribunal notes it may refer a suspected breach of professional obligation to a regulatory body.
What to do: Be able to name the tool, say how it was used, and say who checked the output. That is the explanation duty in the Tribunal's own definition, and it is a record-keeping problem before it is a legal one.
Primary sources: ART (Use of Generative AI) Practice Direction 2026 (PDF) · ART news item, 24 August 2026
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
- Instrument (EN)
- Administrative Review Tribunal (Use of Generative AI) Practice Direction 2026. The running header and body use that title; the cover footer, the table of contents, the ART news item and the PDF filename use the shorter Administrative Review Tribunal (Generative AI) Practice Direction 2026
- Authority
- Administrative Review Tribunal, Australia. Signed by the Hon Justice Emilios Kyrou AO, President
- Jurisdiction
- Australia, Commonwealth. All jurisdictional areas of the Tribunal except where stated otherwise
- Event type
- Practice direction. Made under section 36(1) of the Administrative Review Tribunal Act 2024
- Date of legal event
- Dated 20 August 2026. It commences and has effect from the date it is signed. The Tribunal announced it on 24 August 2026
- Retrospective reach
- Applies to all applications whether lodged before, on or after that date, and remains in effect until superseded or revoked
- Derivation
- The document states it was adapted from the Federal Court of Australia Use of Generative Artificial Intelligence Practice Note (GPN-AI) dated 16 April 2026, with the kind permission of the Chief Justice of the Court
- Companion documents
- A separate Checklist for using Generative AI in Tribunal proceedings, and the ART Common Procedures and Expert Evidence Practice Directions 2026, both recently updated. The Tribunal says this is its 8th practice direction
- Consequence of non-compliance
- The Tribunal may decline to rely on affected material where its accuracy, source or reliability cannot reasonably be established, and may refer a representative's conduct to an appropriate regulatory body
- Primary source
- https://www.art.gov.au/sites/default/files/2026-08/Administrative%20Review%20Tribunal%20%28Generative%20AI%29%20Practice%20Direction.pdf
The four duties, in the Tribunal's own definitions
The practice direction defines each duty rather than leaving it to inference, which is what makes it usable. Verification means independently checking Generative AI-assisted material against the relevant facts, evidence, legal authorities, or other source material, as appropriate, before relying on it. Disclosure means telling the Tribunal that Generative AI was used.
Explanation means being able to say, if asked, what Generative AI tool was used, how it was used, and how the output was checked and by whom. Certification means a written statement, provided if required, certifying compliance with the disclosure, explanation and verification requirements, and identifying who made it and when.
Read those together and the compliance burden is largely evidentiary. Three of the four duties are satisfied only if someone kept a record at the time: which tool, what it did, who checked it. None of that can be reconstructed convincingly after a challenge.
The responsibility rule is the sharpest sentence in the document. A person who uses Generative AI must verify the material and remains responsible for it, whether or not Generative AI was used to prepare it. Responsibility does not turn on the tool.
What is carved out
The duties do not extend to incidental outputs that result from the carrying out of ordinary research, such as an AI summary that a general search engine produces automatically, if the person does not materially rely on that output in preparing or presenting the material.
That carve-out is narrower than it first reads. It is anchored to material reliance, not to whether the AI was asked for. Seeing an automatic summary is fine; building on it is not covered by the exception.
The direction also yields where it conflicts with the Act, another Act or instrument under which the Tribunal reviews decisions, a regulation, or a rule. It does not set jurisdiction-specific requirements, and the Tribunal may supplement it through directions, initial letters or standard correspondence.
Where the Tribunal drew the line on its own function
Part 2 frames the whole instrument around the Tribunal's statutory objective. Use of Generative AI must not undermine the Tribunal's objective in section 9 of the Act or, where relevant, the making of the correct or preferable decision in a proceeding. The Tribunal also states that it embraces the beneficial use of technology, so this is not framed as a restriction.
The expectations are cumulative and start low: any person who uses Generative AI will have a basic understanding of its capabilities, limitations and risks, and any use must not adversely affect the administration of justice.
The consequence provision is drafted around reliability rather than punishment. The Tribunal may decline to rely on affected material where its accuracy, source or reliability cannot reasonably be established. Separately, if a representative's conduct may involve a breach of a professional obligation, the Tribunal may refer the matter to an appropriate regulatory body.
The third Australian forum, not the first
The document says plainly that it was adapted from the Federal Court of Australia Use of Generative Artificial Intelligence Practice Note (GPN-AI) dated 16 April 2026, with the kind permission of the Chief Justice of the Court. Anyone reporting this as a first-of-kind Australian instrument is reporting something the source contradicts on its first page.
Australia now has at least three separate forums with their own generative-AI instruments: the Federal Court through GPN-AI, the Fair Work Commission through its disclosure requirement, and now the ART. We have covered the first two; the interest in this one is that a merits-review tribunal, which handles high volumes of self-represented applicants, has taken the Federal Court's litigation-shaped rules and applied them across every one of its jurisdictional areas.
That population difference is the part worth watching. A practice direction written for represented parties in a superior court now reaches applicants who may have no representative at all, which is why the separate plain-language checklist matters more here than it would in a court.
What we did not verify
Opened and read: the practice direction PDF in full, all 10 pages, extracted as text in the browser and read directly, plus the ART news item of 24 August 2026. Every quotation and definition in this article was taken from that extracted text.
A retrieval note that matters for reproducibility: art.gov.au returns HTTP 000 to this machine's command-line fetcher on every path including the site root, while Google, the Turkish Resmi Gazete and China's SAMR all returned 200 in the same minute. The 000 is specific to this machine's path to that host and is not evidence about the document. The PDF was retrieved through a browser session instead.
Not opened: the Checklist for using Generative AI in Tribunal proceedings, the Common Procedures and Expert Evidence Practice Directions as recently updated, the Federal Court GPN-AI of 16 April 2026 as a document, and sections 9 and 36(1) of the Administrative Review Tribunal Act 2024 as distinct from the practice direction's references to them.
We refuse to claim: that this is legislation, that it creates a sanction, or that it is the first Australian instrument of its kind, which the document itself contradicts. We state no penalty, because the direction provides none; its consequence provisions concern whether the Tribunal will rely on material and whether a representative is referred onward. We make no claim about how the Tribunal has applied it, because it was signed ten days ago.
The obligation this creates is a record, not a disclaimer. Three of the four duties, disclosure aside, can only be discharged by someone who noted at the time which tool was used, how, and who checked the output. And because responsibility attaches whether or not Generative AI was used to prepare the material, a party gains nothing by staying vague about it. If you appear in the ART, the practical change is that AI use now needs a contemporaneous file note.
Source File
Open the practice direction PDF from art.gov.au and confirm four things. First, the signature block reads Dated: 20 August 2026 with the Hon Justice Emilios Kyrou AO as President. Second, paragraph 1.1 makes it under section 36(1) of the Administrative Review Tribunal Act 2024. Third, footnote 1 on the first page records that it was adapted from the Federal Court GPN-AI dated 16 April 2026 with the Chief Justice's permission. Fourth, the Definitions part gives separate meanings for verification, disclosure, explanation and certification. Note that the running header says Use of Generative AI while the cover and the ART news page say Generative AI.
verification means independently checking Generative AI-assisted material against the relevant facts, evidence, legal authorities, or other source material, as appropriate, before relying on it. ยท ART (Use of Generative AI) Practice Direction 2026, Definitions
FAQ
Does this create a penalty for using AI badly?
No. The direction sets no penalty. Its consequence provisions are about reliance and referral: the Tribunal may decline to rely on affected material where accuracy, source or reliability cannot reasonably be established, and where a representative's conduct may involve a breach of a professional obligation it may refer the matter to an appropriate regulatory body.
Do I have to disclose every use of AI?
Not every use. The duties do not extend to incidental outputs from ordinary research, such as an AI summary a general search engine produces automatically, provided you do not materially rely on that output in preparing or presenting the material. The test is material reliance, not whether you intended to use AI.
Is this Australia's first tribunal or court rule on generative AI?
No, and the document says so. Its first page records that it was adapted from the Federal Court of Australia GPN-AI dated 16 April 2026. The Fair Work Commission also has its own generative-AI disclosure instrument. The ART is a third Australian forum, not the first.
What is the difference between disclosure, explanation and certification here?
They are defined separately. Disclosure is telling the Tribunal that Generative AI was used. Explanation is being able to say, if asked, what tool was used, how it was used and how the output was checked and by whom. Certification is a written statement, provided if required, certifying compliance with all three of disclosure, explanation and verification, and identifying who certified it and when.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.