Using AI to Write Psychology Reports Without Falling Foul of AHPRA or the Courts

Ask a psychologist what they like least about the job and very few will say the clients. They will say the reports.

Reports are where the hours disappear. A complex workers compensation or medico-legal report can absorb four to six hours, most of it after the last appointment of the day, and much of it unpaid. So it is no surprise that AI has arrived in this corner of the profession faster than in almost any other. The question is no longer whether psychologists are using it. Plenty are. The question is whether they are using it in a way that would survive a notification to AHPRA or a cross-examination in court.

Those are two different tests, and it is worth understanding both.

What AHPRA actually asks of you

In August 2024, AHPRA published guidance titled Meeting your professional obligations when using Artificial Intelligence in healthcare. It is not a new code. It sits over the top of the obligations you already have, and it is refreshingly short. Five ideas run through it.

You remain accountable. AHPRA is direct about this: you must "apply human judgment to any output of AI" and verify the accuracy of anything AI-generated before it goes into the record. The tool does not carry the risk. You do. AHPRA also notes that TGA approval of a product does not shift that responsibility one inch.

You must understand the tool. Not at an engineering level, but enough to explain what it was built for, what it was trained on, where the data goes, and where it falls down.

You must be transparent. Patients and clients should know you are using AI, and you should take their concerns seriously if they raise them. How much detail you give depends on what the tool does.

You must obtain informed consent where personal information is being entered into a tool. This matters most for anything recording a session.

You must manage the legal and ethical side: confidentiality and privacy obligations, awareness of algorithmic bias across diverse populations, and professional indemnity cover that actually contemplates how you are working.

Read plainly, none of that prohibits using AI in report writing. It asks you to stay in charge of it, to know what you are using, and to be honest about it. Most psychologists I speak to are comfortable with that once they see it written down. What catches people out is the second test.

What the courts say, and why it is stricter

If your report is going to court, the rules are considerably tighter than AHPRA's, and they vary by jurisdiction.

The Supreme Court of New South Wales set the strictest position in the country. Practice Note SC Gen 23, which commenced on 3 February 2025, states at paragraph 20 that generative AI "must not be used to draft or prepare the content of an expert report (or any part of an expert report) without prior leave of the Court." Where leave is granted, the expert must disclose which parts of the report were prepared using it, name the program and version, and annexe a record of how it was used, including the prompts.

Read that carefully, because the reach is wider than most clinicians assume. It is not aimed at fabricated citations. It is aimed at the content of the report itself, on the reasoning that an expert report is supposed to be the expert's own reasoning in the expert's own words.

Victoria has taken a different route. Practice Note SC Gen 25, which commenced on 14 May 2026, does not impose a blanket prohibition. It asks for particular caution where AI has assisted in preparing an expert report, and it requires court users to be able to identify the specific portions of a document produced using AI and to explain how the output was verified.

That difference in approach tells you something useful. NSW has drawn a line at the door. Victoria has said: come in, but be able to show your work. Either way, the practical demand landing on the psychologist is the same one — know exactly what the machine contributed, and be able to account for it.

And note the limits. These are practice notes of particular Supreme Courts. They do not automatically govern a report to an insurer, a treating report to a GP, an NDIS report or a workers compensation tribunal in another state. Before you assume a rule applies, check which forum your report is actually going to. Before you assume it doesn't, ask whether the matter could end up somewhere it does.

The distinction that matters more than the technology

Underneath all of this sits one question: did the AI help you express your reasoning, or did it supply reasoning you did not have?

That is the line, and it is a clinical line rather than a technical one. A tool that structures a report you have already thought through, chases the thread from referral question to opinion, and stops you leaving out the history section at 11pm is doing something quite different from a tool that produces a plausible-sounding formulation you then nod along to. The first makes you faster at your own work. The second quietly outsources the part of the job you are actually registered to do.

This is also why the general-purpose chatbot is the riskiest option on the table, not the safest. It will happily generate a confident paragraph about a client it knows nothing about, it has no idea what a fair and reasonable opinion looks like under a code of conduct for expert witnesses, and it may be storing your client's information offshore. A purpose-built report-writing tool is not automatically safer — but at least you can ask it the right questions.

Where a purpose-built tool can genuinely lift the quality

I should declare an interest before this next part: I built one of these — AI report-writing software for psychologists. Read it with that in mind.

The reason I built it is that most of what goes wrong in a complex report is not a factual error. It is a reasoning gap. The opinion at the end doesn't visibly connect to the findings in the middle. The referral question is acknowledged on page one and then quietly abandoned. Something material comes up at interview and never reappears in the formulation. Reviewers, insurers and cross-examining counsel find those gaps quickly, because they are precisely the seams a report is tested on.

That is a structural problem, and structure is something software is genuinely good at. A tool built specifically for report writing can hold the thread — carry the referral question through to the opinion, ask where the evidence for a given conclusion sits, notice that a history was taken and never used, and keep the reasoning visible on the page rather than left implied in the clinician's head. It is not manufacturing the clinical judgment. It is declining to let you leave it unstated.

The same applies to consistency, which is quietly one of the larger quality problems in the profession and one we rarely talk about. Most psychologists' reports vary — between a good week and an exhausted one, between the first report of the year and the fortieth, between two clinicians in the same practice writing for the same insurer. In medico-legal work that variation costs you, because a report that departs from your usual form invites questions about why. Consistency of structure and standard is not a cosmetic benefit. It is part of what makes a body of work defensible.

None of this shifts the accountability an inch, and none of it exempts you from the rules above — if your report is heading to the NSW Supreme Court, you still need leave, however the draft was produced. The clinician decides what the opinion is, whether the reasoning holds, what stays and what goes, and whether the report is sent at all. A tool can prompt, structure and hold the thread. It cannot form a professional opinion, and it should never be asked to.

A practical checklist before you use anything

• Know where the data sits. Australian hosting, offshore storage, and what happens to your content afterwards. Ask for it in writing.
• Get consent, and record that you did. Say what the tool does in a sentence a client would understand.
• Never sign anything you have not read line by line. Every name, date, dosage, test score and quote gets checked against the source.
• Watch for confident invention. The most dangerous output is not the obvious error. It is the smooth sentence that sounds like you and isn't true.
• Check the forum first. NSW Supreme Court work needs leave. Victorian work needs a verification trail. Know which you are in.
• Keep your own audit trail. What the tool drafted, what you changed, what you verified. If you are ever asked, you will want it.
• Read it as a reviewer would. Does the opinion connect to the findings? Is the referral question answered? Would the reasoning survive someone hostile reading it slowly?
• Tell your insurer. A quick email now is a great deal cheaper than a conversation later.

The real risk

The profession's exposure here is not psychologists using AI. It is psychologists using it invisibly — no consent, no verification trail, no idea where the data went, and no ability to explain in a witness box what came from them and what came from a model.

The regulators have not banned this technology. They have asked us to remain the author. That is a standard psychologists can meet, and the better tools should make it easier to meet rather than harder — by making the reasoning more visible, not less, and by leaving the judgment where it belongs.

What I think happens next

This is not a passing enthusiasm, and I don't think it is useful to write about it as though it were. So here is a prediction, offered as a prediction rather than a fact.

AI will end up setting the standard rather than threatening it. The rules being written now — verify the output, disclose what the tool did, be able to explain how you checked it — read as restrictions. Look again and they are a floor. For the first time we are being asked to make explicit what a defensible report requires, and to be able to demonstrate it. The profession has never had that written down before. My expectation is that reports get better, not worse, because the standard finally gets said out loud.

It will give psychologists their time back, and some of them will stay because of it. In April 2026 the Department of Health, Disability and Ageing projected that demand for psychological services will outstrip supply by 96.6 per cent by 2038. You do not close a gap of that size by training more graduates alone. One of the few levers actually available is reclaiming the hours that currently disappear into unpaid documentation. And there is a quieter version of the same problem: the reports are the reason a great many capable clinicians stop taking medico-legal work, or reduce their days, or leave the profession entirely. People stay in jobs they enjoy. Very few psychologists came into this work for the paperwork.

It will widen access to assessment. A national secret shopper study published in February 2026, in which researchers contacted more than 700 clinicians, found adults waiting an average of just over ten weeks for an initial ADHD appointment and sometimes up to a year, children averaging nineteen weeks and sometimes up to two years, and total assessment costs averaging close to $1,400 and reaching almost $4,000. Assessment is bottlenecked by clinician hours, and a substantial share of those hours is report writing. Freeing that capacity will not fix access on its own, but it is not a trivial contribution either.

And it will improve the decisions made off the back of our reports. Insurers, courts, tribunals and government agencies make consequential decisions on the strength of what we write — whether someone is fit for particular duties, what support they need, which role they are actually suited to. Those decisions are only ever as good as the reasoning we hand over. A clearer, more consistent, better-evidenced report makes for a better decision at the other end. The decision stays human. The material it rests on gets better.

None of that arrives automatically, and none of it arrives if we use these tools carelessly. It arrives if the profession sets the terms early — which is what the last twelve months have actually been about.

Our name goes on the last page. It should mean something, and it should mean the same thing on every report we write.