When the photocopier turned up, there was no Procedural Direction requiring disclosure that a brief had been duplicated rather than transcribed by hand. The fax machine, the word processor, email, electronic filing – each one quietly revolutionised legal practice, and each time the profession absorbed the change without requiring an entirely new regulatory apparatus. The use of generative AI in Australian courts is the latest test of that adaptability. And as with every tool before it, the question is not whether lawyers will use it, but whether the profession’s existing filters are equal to the risks it introduces.
Artificial intelligence is, in one sense, just the latest entry in that long line of tools. But it has a quality the others lacked. A photocopier never invented a case citation. Similarly, a fax machine never hallucinated a provision of the Corporations Act. AI can, and does, produce material that looks authoritative and reads convincingly, but is completely wrong. That changes the risk profile. The filter – the solicitor – matters more than ever.
The real question for the courts has been how much new regulation this new risk actually demands. On 16 April 2026, the Federal Court of Australia issued its Practice Note on Generative Artificial Intelligence (GPN-AI). Having now read it alongside the NSW Supreme Court’s Practice Note SC Gen 23 and NCAT’s Procedural Direction 7, my view is that the Federal Court has landed in a materially better place.
The Lay of the Land
Let me be direct about what each instrument actually does, because the differences matter. The NSW Supreme Court moved first, and moved hard. SC Gen 23 flatly prohibits the use of generative AI to draft or prepare any part of an expert report without prior leave of the Court. It requires a positive disclosure statement in every affidavit, witness statement and character reference confirming that AI was not used to generate the content. It requires verification of every legal citation, academic authority, case and legislative reference. The overall tone is cautious and prescriptive.
NCAT’s Procedural Direction 7 follows a similar template. It requires disclosure when practitioners have used AI to prepare documents filed with the Tribunal, demands verification of legal authorities, and emphasises the continuing responsibility of the person who files the document.
The Federal Court’s GPN-AI takes a different approach. Rather than leading with prohibitions, it starts from the premise that AI is a tool and that the Court “embraces the beneficial use of technology.” It then builds its expectations around existing legal and professional obligations. There is no blanket ban on AI in expert reports. Instead, the practice note says an expert report ‘should contain that expert’s own opinions based on the expert’s specialised knowledge,’ which is simply restating what rule 23.13 of the Federal Court Rules already requires. It expects disclosure where practitioners have used AI in the preparation of documents filed with the Court, but frames this as enabling the Court to gather information and develop its approach, rather than as a policing mechanism. It sensibly distinguishes between ‘open’ or ‘public’ AI tools (where providers may retain and reuse data) and ‘closed’ or ‘controlled’ environments, recognising that the confidentiality risks differ significantly.
Quick Reference: Comparing the Three Regimes
The following table summarises the key rules across the Federal Court (GPN-AI, April 2026), the NSW Supreme Court (SC Gen 23, February 2025) and NCAT (PD7). It is intended as a quick reference only and should be read together with the full text of each instrument.
| Topic | Federal Court (GPN-AI) | NSW Supreme Court (SC Gen 23) | NCAT (PD7) |
| General stance on AI use | Embraces beneficial use of technology; expectations built on existing legal and professional obligations. | Cautious and prescriptive; detailed mandatory rules overlaid on existing obligations. | Follows NSW Supreme Court model; emphasises continuing responsibility of the filer. |
| Disclosure of AI use | Required where AI used in preparing filed documents. All persons must be able to tell the Court what AI was used, how and for what purpose. | Mandatory disclosure in affidavits, witness statements and character references that AI was not used to generate content. Verification of all citations required. | Required when AI used to prepare documents filed with the Tribunal. Person filing must verify accuracy. |
| Expert reports | No blanket prohibition. Expert reports must contain the expert’s own opinions based on specialised knowledge. Disclosure required under GPN-EXPT if AI used. | AI must not be used to draft or prepare any part of an expert report without prior leave of the Court. Strict prohibition. | Not specifically addressed in PD7; general disclosure and verification obligations apply. |
| Affidavits and witness statements | Must reflect deponent’s own recollection, knowledge or experience. Disclosure required if AI assisted in preparation. | Positive declaration required that AI was not used to generate content. Applies to all affidavits, witness statements and character references. | Disclosure required. Person must confirm content is accurate and not misleading. |
| Verification of citations | Follows from existing obligation not to mislead the Court. Practice note flags risk of hallucinations and fictitious citations. | Explicit mandatory requirement to verify every legal citation, academic authority, case and legislative reference. | Verification required. Filing party responsible for accuracy of all content. |
| Confidentiality and data privacy | Distinguishes between ‘open’ and ‘closed’ AI tools. Specific guidance on protecting confidential, suppressed and private information. Expects ‘rigour’ in preserving integrity of compulsory process information. | General warnings about data input. Practitioners must be aware of risks. No detailed distinction between open and closed tools. | General obligation not to disclose confidential information to AI tools. |
| Self-represented litigants | Explicitly addresses litigants in person. Refers to GPN-LIP and external resources for non-lawyers. | Applies to all persons including unrepresented litigants, but no specific separate guidance. | Applies to all parties. No separate guidance for self-represented parties. |
| Approach to future change | Acknowledges AI is “dynamic and constantly evolving.” Maintains a separate ‘Generative AI Resources’ page for ongoing updates rather than requiring practice note amendments. | Fixed rules in the practice note. Amendments published as needed (amended January 2025). | Fixed procedural direction. Updated as required by the President of NCAT. |
Why the Federal Court’s Approach Works Better
The NSW Supreme Court’s approach, adopted in early 2025, was understandable. The profession was still digesting the implications of ChatGPT, lawyers overseas had been caught citing fake cases, and there was genuine alarm about what might land on a judge’s desk. A firm hand felt appropriate. And to be fair, much of SC Gen 23 simply makes explicit what was already implicit in existing professional obligations.
But there is a cost to over-prescription. When you ban AI from expert reports outright – subject to leave – you signal that the technology is inherently suspect rather than a tool whose output depends entirely on how it is used. An expert who uses an AI tool to check whether they have correctly cited a particular dataset, or to help organise the structure of a complex report, is not doing anything that compromises the integrity of their opinion. Therefore, they are doing what experts have always done: using the best available tools to present their work clearly. The blanket prohibition forces a leave application that may discourage perfectly sensible uses.
The Federal Court sidesteps this problem by anchoring its expectations in substance rather than process. The question is not “did you use AI?” but “is the work product yours, is it accurate, and does it comply with your existing obligations?” That is a much more durable framework, because it does not depend on the specific technology. It would have worked for the calculator, the word processor and the search engine, and it will work for whatever comes after large language models.
The distinction between open and closed AI environments is another sensible feature. SC Gen 23 warns practitioners about data risks in general terms, but GPN-AI actually provides a workable framework for thinking about confidentiality. Entering case material into a standard ChatGPT session (where OpenAI may use your prompts for training) is a very different proposition from using a ringfenced, enterprise AI tool that sits behind your firm’s firewall. The Federal Court recognises this and calibrates its expectations accordingly, while still warning that even closed tools carry risks if practitioners later use outputs for different purposes.
The Filter Has Always Been the Lawyer
Here is my central point, and it is not a new one. The role of a solicitor in the administration of justice has always involved filtering. A client walks in with a story. Some of it is relevant, some of it is not. Parts of it will be admissible; other parts are hearsay. And some of it is true, while the rest is, let us say, creatively recalled. The solicitor’s job is to take that raw material, apply professional judgment, and present the Court with something that is accurate, properly supported and genuinely useful.
AI has not changed that job. It has changed the volume and the sophistication of the raw material. A client can now walk in with a 30-page submission drafted by ChatGPT, complete with fabricated case citations, hallucinated statutory provisions and confident assertions about legal principles that do not exist. The material looks polished. It reads well. And it is, to use the technical term that has entered the profession’s vocabulary, slop.
Filtering AI slop is not fundamentally different from filtering a client’s excited recollection of what a witness said, or a director’s optimistic interpretation of their company’s solvency. The stakes are the same: if inaccurate material reaches the Court, it compromises the administration of justice, and the solicitor has failed in their most basic duty. The only difference is that AI can produce slop in much greater quantities and with much greater apparent authority than anything a client could previously manage on their own.
This is why, in my view, the rush to create entirely new regulatory frameworks for AI in litigation is somewhat misplaced. We already have the rules. Solicitors must not mislead the Court. Barristers must not put forward submissions they know to be wrong. Documents filed must be accurate. Evidence must be what it purports to be. The Legal Profession Uniform Law has not suddenly become insufficient because the tool that produced the misleading material happens to run on a neural network rather than a ballpoint pen.
That said, I am not arguing that the practice notes serve no purpose. They serve an important signalling function – they tell the profession and the public that the courts are aware of the technology and its risks, and they provide concrete guidance on disclosure. The Federal Court’s approach, in particular, does this without overstepping into territory that will quickly become dated or unworkable as the technology evolves.
What This Means in Practice
For practitioners, the message across all three jurisdictions is consistent on the essentials. If you use AI, you must understand what it can and cannot do. You must verify every output. You must not feed confidential information into tools that do not protect it. And you must be prepared to tell the Court what you used and how you used it.
Where the instruments diverge is on how much trust they place in practitioners to exercise their own judgment. The Federal Court places more. The NSW Supreme Court places less. NCAT sits broadly with NSW. Reasonable minds can differ on which approach is right. My own view, for what it is worth, is that the Federal Court has the better of the argument, because the most effective regulation is regulation that will still make sense in five years, and prescriptive rules about a technology that is changing monthly will not.
The solicitor’s role as a filter is not diminished by AI. If anything, it is elevated. Clients now have access to tools that can produce vast quantities of plausible-sounding legal material at near-zero cost. The value that a solicitor adds – the judgment, the verification, the ethical and professional framework – is precisely the thing that separates a competent filing from a court’s waste of time.
The Federal Court appears to understand this. The practice note trusts the profession to do the job it has always done, while providing enough structure to keep that trust well-founded. That is the right balance. And it is a balance that will age well, which is more than most attempts to regulate technology mid-flight can claim.
Jake McKinley notes that this article is written for the purpose of providing generalised information and not to provide specialised legal advice. If you require qualified legal advice on anything mentioned in this article, our experienced team of solicitors at Jake McKinley are here to help. Please get in touch with us on 02 9232 8033 today to make an enquiry.
Article Written by Hayden Nelson, Solicitor
References:
Federal Court of Australia, Use of Generative Artificial Intelligence Practice Note (GPN-AI), 16 April 2026
Supreme Court of New South Wales, Practice Note SC Gen 23: Use of Generative Artificial Intelligence, 28 January 2025
NSW Civil and Administrative Tribunal, Procedural Direction 7: Use of Generative AI
