Doctor in, doctor out: where AI belongs in medico-legal drafting, and where it must not.

"Doctor in, doctor out" is easy to say, and on its own it settles nothing. Every vendor of medico-legal drafting software claims a human stays in the loop, and the phrase holds equally whether the doctor rewrites the draft line by line or signs it after a skim. The useful question is narrower: which tasks inside report production can be handed to software at all, and what property decides it.

The property that decides it has little to do with model capability, and it does not move when the models improve. It turns on what the signing doctor can check, and on how much that check costs.

Three different jobs hide inside "writing the report"

Report production reads as one activity and behaves as three. There is assembly: collating the brief, ordering the documents, building a dated chronology, listing what was reviewed, lifting a figure off a page. There is compression: turning forty pages of physiotherapy notes into a paragraph. And there is judgement: diagnosis, causation, prognosis, capacity, impairment, and the reasoning that connects them.

The three differ in a way that matters more than their difficulty. Assembly has an external referent, because every output points at a document and a page that either says it or does not. Compression has a referent too, but a diffuse one, since checking the paragraph means reading the forty pages it replaced. Judgement has no external referent at all, and the only thing a reader can test it against is the reasoning the doctor sets out.

Read down the third column and the delegation rule falls out of it. The tasks that are safe to hand over are the ones where checking is a lookup. The tasks that are not are the ones where checking means redoing the work, or where there is nothing to look up.

The test is whether checking is cheaper than doing

The human factors literature is more useful here than anything written about AI specifically. A systematic review by Lyell and Coiera set out to test the prevailing view that automation bias happens mainly when people are multitasking, and that view did not survive the evidence. Of the forty studies meeting inclusion, automation bias appeared in single tasks, typically involving diagnosis rather than monitoring, and in tasks rated as having high verification complexity.¹

Their conclusion is the design-relevant part. Automation bias appears to track the cognitive load of the decision task rather than multitasking, and strategies to reduce it should focus on reducing that load.¹ Verification that is expensive does not get done carefully, and the more finished the output looks, the less it invites the expense.

An earlier review by Goddard, Roudsari and Wyatt reached compatible conclusions about what helps, finding that mitigators of automation bias included implementation factors such as training and emphasising user accountability, alongside design factors in the decision support itself.² A signature supplies the accountability. It cannot supply the attention, which is what the design has to buy.

So the rule that follows for medico-legal drafting is a demanding one. Keeping a human in the loop counts for nothing unless the check in front of that human is cheap enough to be performed every time, on the twentieth report of the week as reliably as on the first.

The rules landed on the same boundary, which is a good sign

The Australian instruments were not written from a human factors argument, and they arrive at more or less the same place. The NSW Supreme Court's practice note permits generative AI for the generation of chronologies, indexes and witness lists and for the summarising or review of documents and transcripts.³ The Personal Injury Commission's procedural direction permits the same class of work and adds schedules of earnings and damages.⁴

What both prohibit is the mirror image: generating the content of material intended to reflect a person's evidence or opinion, because statements should contain and reflect a person's own knowledge.³⁴ Both then require an expert report to state the opinion and the reasoning process,³⁴ which is also what the harmonised expert witness code demands when it requires the assumptions and material facts behind each opinion, and the reasons for it.⁵

The line those instruments draw sits exactly where the external referent runs out. We have written separately on what the rules actually say, forum by forum, including the leave and disclosure obligations that attach when generative AI touches report content. The point here is narrower: a practice that designs to the verification test will land inside those rules without having to consult them, which is usually a sign that the rule is tracking something real.

Selection is judgement wearing assembly's clothes

The category that causes trouble is the one that looks mechanical and is not. Deciding which entries out of nine hundred pages belong in a chronology, or which of four inconsistent accounts of an injury deserves to appear, is a judgement about materiality, and materiality is the expert's standard rather than a property of the document.

A chronology that silently drops entries has produced an opinion about relevance while presenting itself as a transcript of the file, with none of the reasoning exposed. It is the most dangerous output in the whole pipeline, because it looks like assembly and audits like assembly, and the thing it got wrong is invisible by construction.

The safe construction is to make omission visible. Include by default, mark anything the pipeline could not read or could not date rather than quietly leaving it out, and keep the count of source documents reconcilable against what arrived. A reviewer can then check a boundary decision in seconds, which moves the task back onto the cheap side of the line.

What "doctor out" has to mean in the software

If the boundary is a verification-cost problem, the review surface is where a product either honours it or defeats it. The failure to design against is a draft that reads as finished, which is the hazard the practice notes describe in their own terms as output that is apparently plausible, authoritative and coherent while being inaccurate or fictitious.³

In practice that turns into a short list of properties, every one of them about cost:

• every factual line carries a page reference that opens the source image, so checking a claim is a lookup rather than a re-read;
• uncertainty is surfaced where it happened, so poor recognition and illegible handwriting arrive flagged instead of smoothed into fluent prose;
• omissions and unreadable pages are listed, so the reviewer works through a short exception list rather than the whole file;
• the record of how each page was read accumulates as a by-product, because that record is what a disclosure annexure is made of;³⁴
• the opinion sections stay empty until the doctor writes them, since a drafted opinion is one the doctor is being invited to adopt.

The professional obligation sits underneath all of it. The Medical Board's code requires a doctor to take reasonable steps to verify the content before signing a report, and to make clear the limits of their knowledge rather than giving opinion beyond them.⁶ Neither duty can be discharged against a document the doctor cannot cheaply check.

Trenthos IME is built to that shape. It reads the brief and builds the dated chronology, with a page reference behind every entry, while the opinion stays where the doctor writes it. What the software could not read is flagged rather than filled in. It is a chronology and document-indexing tool, not a decision-maker.

The boundary is a design constraint, not a disclaimer

The reason to draw this line precisely is that the alternative is drawing it in the report's tone. A system with no boundary produces a document that reads as complete, and the doctor's review then becomes an act of trust in prose rather than an inspection of sources. Everything that makes an expert report defensible runs the other way.

Doctor in, doctor out earns its meaning when the doctor's part is the part only a doctor can do, and when checking the rest costs seconds rather than an evening. With anything less behind it, the phrase is a slogan attached to whatever the software happened to produce.