In late 2024, the Supreme Court of New South Wales issued Practice Note SC Gen 23, a binding instrument governing how generative AI may be used in civil litigation. The rules took effect on 3 February 2025 and made NSW one of the first Australian jurisdictions to formally regulate AI-assisted court work. The Practice Note prohibits the use of generative AI in preparing affidavits, witness statements, character references, and expert reports, and requires disclosure of AI assistance in certain submissions. Chief Justice Andrew Bell framed the rules as a response to a string of embarrassing incidents in which lawyers — in Australia and overseas — filed submissions citing case authorities that simply did not exist.
The trigger for these rules is real and worth taking seriously. The 2023 New York saga of Mata v. Avianca, in which two lawyers cited six fabricated decisions hallucinated by ChatGPT and were sanctioned US$5,000, has now been echoed in courts from Manchester to Melbourne. In 2024, a Melbourne family lawyer was referred to the Victorian Legal Services Board after filing a list of authorities that included non-existent judgments. Judicial confidence in the integrity of submissions is a load-bearing column of the adversarial system. When that confidence cracks, the consequences for cost, delay, and access to justice are not abstract.
What the Practice Note actually does
SC Gen 23 is more nuanced than the headlines suggest. It does not ban AI from legal practice. Lawyers may continue to use generative AI for legal research, summarisation of public documents, drafting routine correspondence, or non-evidentiary submissions — provided that any output is independently verified. The hard prohibitions are narrowly drawn around evidentiary material: affidavits, statements of witnesses of fact, expert reports, and references. The logic is straightforward — these are the documents whose voice must remain the deponent's own, not a stochastic parrot's.
The disclosure regime is similarly targeted. Where AI is used to prepare written submissions or chronologies, parties must inform the court of the tool used and the nature of the assistance. This is closer to the position taken by the Federal Court of Australia in its own 2024 guidance, and to the approach signalled by the King's Bench Division in England and the Fifth Circuit in the United States.
Where the rules over-shoot
A pro-innovation reading of SC Gen 23 has to be honest about two costs. First, the blanket prohibition on AI use in affidavit drafting is broader than the underlying problem requires. Affidavits are not magical incantations — they are records of fact transcribed and refined under a solicitor's supervision. A junior lawyer who uses an AI tool to tighten the prose of an affidavit, while ensuring every factual sentence is checked against the deponent's instructions, is doing exactly what good drafting has always required: economising on time without economising on truth. Treating that workflow as forbidden risks pushing the use underground rather than disciplining it.
Second, the rules apply uniformly regardless of party resources. Self-represented litigants and small-firm practitioners — the people for whom AI productivity gains matter most — face the same restrictions as a top-tier firm with seven-figure tech budgets. The Productivity Commission has previously found that ~30% of Australians who experience a legal problem do not seek any assistance, with cost the dominant barrier. A rule that closes off the most promising route to cheaper legal drafting in the name of preventing hallucinations should be sized to the real risk, not the worst case.
What a proportionate version would look like
The right answer is not to scrap SC Gen 23 but to refine it. Three principles should guide the next iteration:
- Outcome-based duties, not tool-based bans. The professional obligation of accuracy already exists under rule 19 of the Legal Profession Uniform Conduct Rules. A rule that says "every authority cited must be verified and every factual statement attested" achieves the policy goal without specifying which software is forbidden.
- Calibrated disclosure. Disclosure of AI assistance is sensible for evidentiary material and expert reports, where provenance matters. Mandating it for ordinary correspondence or research memos generates paperwork without protecting any litigant.
- Safe-harbour pilots. The court could authorise specific verified AI tools — for example, those that produce auditable citation logs — for use in defined classes of matters. This is the approach being trialled in the Singapore Courts' AI sandbox and is consistent with the Australian Law Reform Commission's recommendation that regulators experiment before generalising.
The bigger Australian picture
SC Gen 23 sits alongside the Federal Government's voluntary AI Safety Standard released in September 2024, the proposed mandatory guardrails for high-risk AI use cases, and the Privacy Act review's recommendations on automated decision-making. Together they signal an Australian regulatory style that is more cautious than the United States' sectoral approach and less prescriptive than the EU AI Act. That middle path is defensible — but its credibility depends on the rules being calibrated to harm, not to optics.
The NSW Supreme Court deserves credit for moving early and writing a rule that distinguishes evidentiary documents from ordinary practice. The risk is that other Australian courts treat SC Gen 23 as a template rather than a draft. Generative AI is the most significant productivity shift in legal services since the introduction of word processing. A regulatory posture that defaults to prohibition will not protect the courts — it will simply concentrate the gains with those large enough to ignore the costs.