Keeping clinical judgement defensible when AI is in the workflow.
A medico-legal report is not just a clinical document. It is evidence. It can be tendered, cross-examined and read aloud in a hearing, and the specialist who signed it may have to defend every line of it under oath. That is the test any AI in the workflow has to be measured against. Not "does it save time", but "does what it produces still hold up when a barrister goes through it slowly".
The good news is that the line is clear. AI can legitimately organise the evidence and help you draft. What it cannot do, without putting the report at risk, is form the opinion or invent the material the opinion rests on. Keeping those two things apart is what keeps a report defensible.
Defensibility is already defined, and it is strict
Long before generative AI, the courts set out what an expert report has to contain. Under the Expert Witness Code of Conduct in Schedule 7 of the Uniform Civil Procedure Rules 2005 (NSW), and the equivalent harmonised code annexed to the Federal Court's expert evidence practice note, an expert's paramount duty is to the court, not to the party paying for the report. The expert is not an advocate. The report must state the reasons for each opinion, and the facts and assumptions the opinion is based on. Where an opinion is not fully concluded, the expert has to say so.
Read that list again with AI in mind. Every requirement is about your reasoning being visible and owned: your reasons, your assumptions, your stated limits. A report whose reasoning is really a model's, dressed in your name, does not meet that standard. The duty has not changed. AI just makes it easier to breach without noticing.
The new failure mode: confident invention
The specific danger with generative AI is not that it is unhelpful. It is that it is fluent when it is wrong. It will produce a clean chronology, a plausible citation, a well-phrased history, and some of it may simply be made up. This is not a fringe risk. When Stanford's RegLab tested the purpose-built, retrieval-backed legal research tools that market themselves as hallucination-free, it still found them fabricating material in a meaningful share of answers. If tools engineered specifically for legal accuracy invent content, a general model summarising a referral file will too.
Courts have already seen where that leads. In the United States, the now-notorious matter of Mata v Avianca ended in sanctions after lawyers filed a brief citing cases that ChatGPT had invented. Australia has its own examples: in Valu v Minister for Immigration and Multicultural Affairs (No 2) [2025] FedCFamC2G 95, submissions were filed containing numerous non-existent authorities generated by AI, and the practitioner was referred to the legal regulator. Swap "fabricated case" for "fabricated history, date or citation to the file" and the exposure lands squarely on the medical expert instead of the lawyer.
A hallucinated fact does not announce itself. It reads exactly like a real one, which is why it survives all the way to the witness box.
Automation bias: the quieter risk
The obvious failure is a fabricated fact you did not catch. The subtler one is documented in the clinical literature as automation bias: the measured tendency of trained professionals to defer to a confident machine, even against their own better read. A systematic review in the Journal of the American Medical Informatics Association found that clinicians can be moved from a correct decision to an incorrect one when a decision-support system advises them wrongly. The point is not that experts are careless. It is that the pull towards the confident output is real and has to be consciously resisted, especially at the end of a long file when a tidy AI summary is a relief to accept.
In a medico-legal setting, automation bias is not just a clinical safety issue. It is an evidentiary one. If your opinion drifted towards what the tool suggested rather than what the evidence supports, that is precisely the soft spot a skilled cross-examination is built to find.
The rules now name AI directly
Regulators have stopped speaking in generalities. In New South Wales, the Supreme Court's practice note on the use of generative artificial intelligence, in force from 2025, restricts how the technology may be used in litigation. As a general position it provides that generative AI is not to be used to produce the content of an expert report without the court's leave, and that where it is used, its use must be disclosed. Whatever the precise wording in your jurisdiction, the direction is unmistakable: an expert who quietly lets a model write the substance of a report is now on the wrong side of an explicit rule.
The medical regulators point the same way, and they put the responsibility on the practitioner. Ahpra's guidance is that practitioners remain responsible for any AI used in the course of their practice and must apply their own judgement to its output. The RACGP and medical defence organisations such as Avant say the same thing about AI-generated notes: they are drafts to be checked, the clinician stays liable for what ends up in the record, and clinical decision-making stays with the doctor. None of that liability transfers to the software.
Where the line actually sits
All of this points to a workable division of labour rather than a reason to avoid AI. The tasks that are safe to hand over are the mechanical ones around the opinion. The tasks that must stay with you are the opinion itself and the verification of anything the opinion relies on.
- Fair game for AI. Reading a large file, building a first-pass chronology, pulling the clinically relevant history forward, transcribing the assessment, and turning your own dictation into a first draft of the answers.
- Never the AI's. The clinical opinion, the reasoning behind it, and the sign-off that each fact the opinion relies on is real and traceable to its source.
The practical safeguard is source-linking. If every line the tool produces points back to the exact page it came from, verification stops being an act of faith and becomes a quick check you can actually perform. You are no longer trusting a summary; you are confirming it against the record, one citation at a time. That is the difference between an AI that structures evidence and an AI that manufactures it.
How Medilee is built for this
This is the line Medilee is designed around. It reads the referral files and builds a structured, source-linked chronology, so every fact is one click from the page it came from and nothing has to be taken on trust. It turns your dictation into a first draft of the answers to the referrer's questions, which you then review, correct and own. It does not form a clinical view, and it does not write your conclusion. The design goal is faithfulness over fluency: a draft that is easy to verify, not one that sounds impressive and hides its sources. You form the opinion. Medilee handles the paperwork around it, in a way you can stand behind.
A note on sources
The rules referenced here differ by jurisdiction and are updated regularly, and court practice on AI is moving quickly. The links above point to the primary bodies (the relevant court, Ahpra, the colleges and the cited studies). Before relying on a specific obligation, check the current version that applies to your jurisdiction and matter. The underlying principle is stable regardless of the detail: the expert owns the opinion and the responsibility, and no AI in the workflow changes that.
