AI in medico-legal reports: what the rules actually say.

Most doctors who write medico-legal reports believe one of two things about AI: that it is not allowed, or that the rules are so unclear that avoiding it entirely is the only safe course. Both beliefs are out of date. Between late 2024 and mid 2026, Australian courts and tribunals built a detailed rulebook for AI in expert evidence, and it prohibits far less than most report writers assume. What the rules require instead is disclosure, verification and human accountability, and they now describe with unusual precision what a compliant tool must be able to do. Every instrument cited below was retrieved and checked against its source in the week this article was published.

No forum bans it, and every rule points the same way

Start with the instrument closest to most IME work: the NSW Personal Injury Commission, the tribunal for workers compensation and motor accidents disputes, whose Procedural Direction PIC13 commenced on 1 January 2026 alongside new Commission rules.¹ PIC13 does not prohibit AI in expert reports. It permits generative AI outright for chronologies, indexes, schedules of earnings and damages, and the summarising or review of documents and transcripts (paragraph 11), and for report content it adopts a disclosure model: under rule 33B, generative AI must not be used to draft the content of a report without that use being disclosed in the report, identifying the parts, the program, its version and the date, with a record of prompts, default values and variables set annexed (paragraphs 22 to 23).¹ It also obliges instructing lawyers to draw these requirements to the attention of experts, expressly including independent medical examiners and treating specialists (paragraph 26), and it bans AI tools, including transcription apps, from being joined to live hearings (paragraphs 27 to 28).¹

The NSW Supreme Court's Practice Note SC Gen 23, which commenced on 3 February 2025, is the strictest instrument in the country, and even it is a permission structure rather than a ban. Its definition of generative AI expressly excludes tools that merely provide transcription or translation, correct spelling or grammar, or assist with formatting (paragraph 6); chronologies, indexes, witness lists and the summarising of documents are expressly permitted uses (paragraph 9B).² The hard rule sits in paragraph 20: generative AI must not be used to draft the content of an expert report without prior leave of the Court, and where leave is granted the expert must disclose which parts used it and annex a record of how, including prompts (paragraph 22). Paragraph 23 extends the leave requirement to work injury damages reports referred to in a pre-filing statement under section 315 of the Workplace Injury Management and Workers Compensation Act 1998, which is how the Supreme Court's regime reaches reports that begin life in the workers compensation system.²

The same architecture has now spread through the rest of the country. The Federal Court's GPN-AI practice note, published on 16 April 2026, requires that where AI is used to summarise, analyse or generate material informing evidence or opinion, the use be disclosed at the start of the document, stating where and how, and it reminds experts that reports must reflect their own reasoning and opinions.³ The Federal Circuit and Family Court followed on 29 May 2026 with a practice direction applying to all court users including private experts, whom instructing lawyers must give a copy of the direction and who must certify they have read it.⁴ Victoria's Supreme Court issued Practice Note SC Gen 25 on 14 May 2026, requiring AI-produced content to be verified with meaningful human control and singling out expert reports for particular caution.⁵ South Australia adopted guidelines across all its courts from 1 January 2026.⁶ Queensland has no civil-proceedings AI practice direction, but its amended criminal expert evidence direction shows where the thinking is: an expert who used generative AI in formulating or expressing an opinion must name the program and annex a complete record of prompts and outputs (Practice Direction 14 of 2024, paragraph 16(l)).⁷ Western Australia, Tasmania and the territories have no dedicated instrument yet; the Law Society of NSW maintains a current tracker of every Australian protocol.⁸ The Victorian Law Reform Commission's February 2026 report, the most thorough Australian review to date, recommended a principles-based framework, not prohibition.⁹

One more fact matters to IME work specifically: the scheme regulators are silent. A review for this article found no AI-specific rules for independent medical examination or permanent impairment reports from SIRA, icare, WorkCover Queensland, WorkSafe Victoria, the TAC, Comcare, ReturnToWorkSA or WorkCover WA as at 2 August 2026. The operative obligations therefore come from the courts and tribunals above, plus the professional layer below. That could change at any time, and it is worth watching.

The line every instrument draws

Read together, the instruments draw one consistent line through the work of report writing, and it is worth being precise about it because the whole question turns on which side of the line a given use falls. Tools that capture, locate and verify sit outside or inside an express carve-out everywhere: transcription is excluded from the definition in SC Gen 23 and outside the operative rules in every other forum, and source-document chronologies, indexes and summaries are expressly permitted even under the strictest instrument.²¹ Generating the expert's opinion sits on the far side in every jurisdiction, because every expert witness code of conduct requires the opinion to be the expert's own, a duty that predates AI and is unchanged by it.¹⁰ Between those poles is the regulated middle: generative drafting of report content, which needs leave in the NSW Supreme Court and disclosure with a prompts annexure in the Personal Injury Commission, the Federal Court and Queensland's criminal jurisdiction.

The professional layer, and what the cases punish

Above the forum rules sits professional regulation, and it says the same thing in clinical language. Ahpra and the National Boards' guidance on meeting professional obligations when using AI in healthcare keeps the practitioner accountable for outputs: understand the tool, be transparent, obtain consent where personal information is entered, and apply human judgement to anything AI produces.¹¹ Avant's guidance for expert report writers is blunter, advising against using general-purpose generative AI such as ChatGPT to draft reports at all, both for accuracy and because entering identifiable health information into a third-party tool raises privacy problems.¹² Note what the insurer is warning against: general-purpose, unverifiable tools, not the category.

The case law lands the same way. In the United States, where no rule prohibits AI in expert work, a Minnesota federal court excluded an expert declaration that contained AI-hallucinated citations the expert had failed to check, while expressly not faulting the use of AI itself; months later a New York federal court admitted an expert who wrote his report from his own expertise and used a language model only to double-check his conclusions.¹³ In England, the Civil Justice Council's February 2026 consultation proposes that experts declare, within the statement of truth, what AI tools they used and for what purposes, with transcription expressly excepted, which is a proposal and not yet law but confirms the direction.¹⁴ Nowhere in any of this is a court punishing careful, disclosed, verified use. What gets punished, everywhere, is unverified output presented as the expert's own work.

The rules are a design brief

Put the instruments side by side and something useful happens: they stop reading as restrictions and start reading as a specification. A tool fit for this work must ground every factual claim to a source page, so that verification never depends on trusting the AI. It must keep the clinical opinion where every code of conduct requires it to live, with the doctor, and be able to show that it did. It must hold the file in a closed environment on Australian infrastructure, with case material not used to train models, because a claim file is sensitive health information inside a legal proceeding. And it must be able to produce, on demand, the record the rules ask for: which parts of a draft were AI-assembled and which were the doctor's, the program, the version, the dates, and the prompt records behind them, in a form that can be annexed to a report or a leave application.

A doctor dictating case material into a general-purpose chatbot can satisfy none of that. There is no source-page trail, no separation between what the model wrote and what the doctor decided, no closed data environment, and no way to reconstruct the prompts annexure that PIC13 paragraph 23 demands. This is the gap Trenthos IME is built for: every factual line in a draft carries a reference to the source page it came from or a visible flag for review, the opinion is supplied only by the examiner because no code path can generate it, files are processed onshore in ap-southeast-2 in a closed environment and not used to train models, and the system compiles its own audit trail into a one-click disclosure annexure recording the model, its version, the dates, the doctor's instructions and settings for the matter, and which sections were AI-assembled versus doctor-supplied, designed to satisfy the disclosure requirements SC Gen 23 and PIC13 describe. Trenthos IME is being built as the compliance infrastructure for using AI in medico-legal work. Early access enquiries: support@trenthos.com.

What a prudent report writer should do today

The practical position, forum by forum, reduces to six habits. First, transcription tools sit outside the rules everywhere, provided the output is verbatim and you verify it. Second, source-document chronologies and indexes are expressly permitted, even under the strictest instrument in the country. Third, if generative AI touches report content: in the NSW Supreme Court, seek leave first; in the Personal Injury Commission, disclose in the report and annex the prompts and settings; federally, disclose at the start of the document. Fourth, never let any tool generate or shape the opinion, because every code of conduct requires the opinion to be yours and the case law punishes unverified output, not AI use. Fifth, keep records of program, version, date, prompts and settings; if your tools cannot produce that record, they are not fit for this work. Sixth, ask your instructing solicitor which forum the report may end up in, because the rules differ, and under PIC13 paragraph 26 the solicitor is now obliged to raise this with you anyway.

Current as at 2 August 2026; last reviewed 2 August 2026. These instruments are being amended rapidly. This article is general information, not legal advice: confirm the requirements of the specific forum with your instructing party before relying on any tool, including ours.