Writing from the team building ADMRL: on assessment, evidence, privacy, regulation, and the careful use of AI in Australian psychology practice. No hype and no hand-waving, just the thinking behind the platform.
On the desk this weekAHPRA's AI guidance.
32Insights
6Practice Areas
0Hallucinated Citations
Alix · Reading with you
Showing all 32 insights
Latest insightAI Reliability
Same data, same score. Every time.
A tool that can turn identical raw scores into two different standard scores is a dice cup rather than a scoring engine. Determinism in scoring is the line we will not cross.
Craig Adams · 5 min read
WISC-V · Raw 42 · Norm table locked
run 001→SS 98✓
run 002→SS 98✓
run 003→SS 98✓
scoring engine v2.3.1 · deterministic · audited
Large language models are probabilistic by design. The same property that makes them fluent with language disqualifies them from arithmetic. Ask one to convert a raw score to a scaled score often enough and it will eventually give two different answers to the same question. In a consumer chatbot that is a quirk. In a clinical assessment it is a defect with a client's name attached to it.
The fix is architectural. No prompt turns a probability distribution into a lookup table. Scoring belongs in versioned, tested, deterministic code: lookup tables, norm conversions, confidence intervals, the same input producing the same output on every run, for as long as the software exists. The model's job is to narrate what the numbers mean. It is never allowed to make them.
That split, with a deterministic engine underneath, a probabilistic narrator on top and human judgement above both, is how you get the speed of AI without surrendering the repeatability your registration depends on. Any tool that cannot tell you which of its numbers came from code and which came from a model has not made the distinction, and the question is worth asking before a report goes out under your name.
The ADMRL position
Every score in ADMRL is produced by a locked, versioned scoring engine with a full audit trail. Alix, the platform's assistant, explains scores and never calculates them.
AHPRA & Regulation
AHPRA's AI guidance, read closely
AHPRA's guidance on AI did something more demanding than a ban: it left you accountable for everything the tool produces. Read as workflow requirements, its principles change how an assessment practice runs.
Myra Jennings · 6 min read
→
AHPRA's position on practitioner use of artificial intelligence is short, and every line of it lands on the same point: the tool changes and the accountability does not. You remain responsible for the care, the records and the decisions, whether a sentence was typed by you at the keyboard or drafted by software in seconds. Transparency with clients, human oversight of outputs and continued professional judgement run through the document as recurring themes.
Read as workflow requirements rather than abstract principles, the guidance asks four things of a practising psychologist: to understand what the tool does with client information, to be open with clients about its use, to check every output before adopting it into the record, and to explain the tool's behaviour when asked. A practitioner who could not describe that behaviour to a colleague, or to a panel, is trusting the tool rather than overseeing it.
The uncomfortable implication is that oversight has to be real rather than ceremonial. Scrolling to the bottom of a generated report and pressing approve does not amount to review, and everyone involved can tell the difference, including a future panel reading the record. The practitioners who do well under this guidance will be the ones whose tools let proper review happen quickly, because software that rewards skimming will eventually be used exactly that way.
The ADMRL position
ADMRL enforces section-level approval gates: nothing reaches a signed report without explicit clinician review, and the audit trail records exactly what you saw, changed and approved.
AI Reliability
Agents in the clinic, on a short leash
Agentic AI can run an assessment workflow end to end. The safety question is what it may decide alone, what it must refer to you, and how the boundary is enforced.
Justin Delahunty · 5 min read
→
Alix runs an assessment pipeline end to end. It proposes the battery, sequences the instruments, sends the questionnaire links and chases the missing consent form, and no human scripts each step. Software with that ability is an agent in the technical sense: given a goal, it plans its own route. The property that makes it useful is the same property that should make a clinician cautious, since the system's next step is a step you did not choose.
Capability is worth keeping, so the remedy is a boundary on authority. In our architecture every agent action belongs to one of three classes: actions Alix may take on its own (assemble, score, draft, remind), actions it may only propose (battery changes, interpretive framings), and actions it may never take (diagnosis, sign-off, anything that leaves the platform). The leash is a permission table, enforced in code and audited.
Human-in-the-loop fails when the human becomes a formality. The design test I apply is whether the system would notice if a clinician approved every proposal for a week without amending one. Ours notices: review gates ask for a decision on each flagged item, and wholesale unexamined approval is treated as a warning sign. The mechanical work gets autonomy, the judgement stays with you, and the architecture is built to know the difference.
The ADMRL position
Alix's agentic workflow runs the assessment pipeline, but its authority is bounded by a three-class permission model (do, propose, never), and every action is logged against the class it was taken under.
NDIS & Funding
The NDIS report that gets funded
A diagnosis alone has never funded a support package. The NDIS reports that move carry a visible evidence chain from standardised instrument to observed behaviour to functional statement to support recommendation.
Myra Jennings · 6 min read
→
A diagnosis has never funded a support package. The NDIS runs on functional impact: what the participant can and cannot do, how often, in which settings, and with how much help. Reports written as diagnostic essays make the planner do the translation from clinical findings to funding criteria, and a planner under volume pressure will usually request further information instead, and the participant's plan waits while the file sits in a queue.
The reports that move stand on a visible chain running from standardised instrument to observed behaviour to functional statement to support recommendation. "Struggles with daily living" is an assertion, and a planner can do nothing with it. "Needs physical prompting to start every self-care task, in keeping with a Vineland-3 Daily Living Skills score at the 1st percentile and the parent interview" is evidence a stranger can check. The difference between those sentences is structural rather than stylistic.
The quiet failures are almost always broken links in that chain: a recommendation with no finding behind it, a score with no functional consequence drawn from it, an opinion that generalises across settings from a single informant. A reviewer's job is to notice exactly those gaps. Before a report goes out the door, someone should trace the chain in reverse, following every recommendation back to a finding, and every finding back to data.
The ADMRL position
ADMRL's NDIS templates carry the evidence chain in their structure: every recommendation links to the findings and instrument data behind it, with WHODAS 2.0 and Vineland-3 built into scoring.
Privacy & Data
The Privacy Act grew teeth
A statutory tort, automated-decision transparency and penalties that finally sting have arrived. The reform program lands squarely on AI tools in health, and on the practices that use them.
Sam Hallawell · 6 min read
→
The Privacy Act spent decades as a law you complied with in principle, and the reform program has ended that era. A statutory tort for serious invasions of privacy, transparency obligations around automated decision-making, and a regulator with genuine penalty powers change the risk calculus for anyone holding client records, and health information sits in the most protected tier of all of it.
For a psychologist using AI tools, the questions are now concrete: where pasted text actually goes, whether it is retained, whether it trains someone else's model, and whose courts would hear the dispute if it leaked. A vendor's assurance that everything is fine is not an answer a regulator will accept from you. The answers need to be contractual rather than marketing.
The practical shift is that privacy has become a design constraint rather than a policy document. Tools built after these reforms tend to behave differently, with data minimisation, Australian residency and de-identification by default, because retrofitting privacy into a product that was architected without it rarely survives contact with an audit. The safest position is a tool that assumed the reformed Act from its first design document.
The ADMRL position
ADMRL was designed inside the reformed Act: Australian data residency end to end, no training on client data, and de-identification before any content reaches a model.
Report Quality
How we validate what Alix drafts
Every instrument in ADMRL passes the same gauntlet before a clinician can use it: golden cases scored twice by hand, drafts challenged for overreach, and sign-off that includes external clinicians.
Myra Jennings · 6 min read
→
Every instrument in ADMRL passes through the same gauntlet before a clinician can use it. We build golden cases, de-identified profiles with known scoring, hand-verified by two psychologists working independently with the manual open. The platform's output is then compared against those cases across the full range of the instrument: floor effects, ceiling effects, and the messy middles where prorating rules and age-band edges live.
Scoring turns out to be the easy half. The harder validation is interpretive: whether the drafted narrative stays inside what the data can support, and adds nothing beyond it. We challenge drafts the way a careful supervisor would, looking for the borderline score described as clinically significant or the single informant generalised into a cross-setting pattern. Each failure is logged, the rules on what a draft may claim are tightened, and the cases are run again, because a draft that overstates is a defect of the same kind as a wrong number.
Nothing ships on one signature. Validation ends with external clinicians who take consented, de-identified cases through the platform and separately by hand, until every difference between the two runs is explainable and documented. Where clinical judgement legitimately varies, the platform must present options rather than verdicts, and holding that line is the point of the entire process.
The ADMRL position
Clinical validation at ADMRL is the precondition for release: no instrument reaches clinicians without golden-case verification and named clinician sign-off.
Clinical Practice
The second opinion you didn't ask for
Anchoring bias is strongest exactly when you are busiest. A dissenting voice built into the assessment workflow forces the alternative hypothesis to be argued before the conclusion settles.
Justin Delahunty · 5 min read
→
The first plausible hypothesis in an assessment has gravity. Every later piece of data gets read in its light, and the intake impression quietly becomes the report's conclusion. Psychologists know this literature better than anyone, since anchoring and confirmation bias are home-ground findings, yet knowing it has never been sufficient protection. Time pressure makes the pull stronger, and assessment work runs on time pressure.
The interventions that work are structural: a second reader, a structured challenge, a checklist that forces the alternative hypothesis to be argued before it is dismissed. Awareness and good intentions perform poorly by comparison. In most practices, though, structure of that kind is a luxury. A colleague with a spare hour to interrogate your formulation is a rare event, so the anchor holds by default.
This is one of the few places where AI can add something genuinely new to assessment practice. A tireless devil's advocate can argue the differential you did not pursue, with citations, on every case rather than the occasional interesting one, and it carries none of the social cost of disagreeing with a colleague. The challenge earns its place even when it is wrong, because the checking it forces is the point.
The ADMRL position
Alix's Advisor mode argues the alternative formulation on request, and the disagreement, together with how you resolved it, becomes part of the case record.
Privacy & Data
The infrastructure you'll never notice
Nobody chooses clinical software for its failover strategy until the Tuesday it matters. Health data sets a higher floor for what enterprise-grade has to mean.
Sam Hallawell · 5 min read
→
The most important parts of a clinical platform are the ones nobody demonstrates. They include where backups live and how often restoration is actually rehearsed, what happens when a cloud region degrades mid-assessment, and how a session's autosave behaves when the clinic wifi drops at question 40 of 60. None of that is visible at the sales call, and all of it is visible on the wrong Tuesday.
Health data raises the floor. Encryption at rest and in transit is table stakes, and the real work sits above it: key management, least-privilege access that is enforced rather than promised, audit logs that cannot be edited by the people they audit, and an incident-response plan with the statutory notification timelines built in, because the OAIC does not treat improvisation as a plan.
The risk rule I apply to every feature is to assume the worst day, then design until the worst day is boring. A failed model call should degrade to a saved draft and never to a lost report. A regional outage should mean latency and never downtime. If the infrastructure is doing its job you will never think about it, which is exactly why someone has to, full time.
The ADMRL position
ADMRL runs on Australian cloud infrastructure with rehearsed recovery, immutable audit logging, and graceful degradation designed feature by feature.
Report Quality
A report a client can rely on
A psychological report is often the most consequential document ever written about a client, and it does its work alone. What gives it that authority is reasoning a stranger can follow.
Myra Jennings · 5 min read
→
A psychological report outlives the appointment that produced it. It is read in rooms its author never enters, by paediatricians, teachers, planners and tribunal members, and it is acted on years after the assessment by people who will never meet the psychologist who signed it. Families keep it. Funding turns on it. Few documents in any profession carry so much decision on their own, and the report can carry it because it was written to stand alone.
That standing is earned on the page. A report holds its authority when a stranger can follow the reasoning from referral question to data to opinion, each step resting on the one before it. The clinical thinking behind a report is almost always sound; the risk is that thinking stays in the clinician's head, with results recorded one evening and opinions another, and the connecting reasoning never quite landing in writing.
The craft that protects a report is specific and learnable: joining sentences that show where each conclusion comes from, clinical impressions labelled as the professional judgement they are, disagreement between informants reasoned through rather than smoothed over, and a colleague's fresh eyes over the reasoning before the document travels. A report built this way deserves the trust that will be placed in it, which is the point of writing one.
The ADMRL position
ADMRL's report sections are structured to carry reasoning as well as results, and section-level approval means every link in the argument was put in front of you before the report was signed.
AI Reliability
How to stop an AI making things up
Hallucination outlives every model upgrade. The fix is retrieval-first design, verification of citations against their sources, and permission for the system to report that it found nothing.
Justin Delahunty · 5 min read
→
Hallucination is intrinsic to how generative models work. A language model continues patterns plausibly, and when the pattern runs out of fact it keeps going anyway, as fluent and confident in error as in truth. Larger models make the failure rarer without changing its character. Waiting for a model too big to invent is a plan with no end point, and changing what the system is allowed to assert works now.
Three design decisions do most of the work. The system answers from documents it retrieved and can show you, so its memory supplies the language while sources supply the claims. Every citation is then verified against the document it names, first for existence and then for whether the passage supports the sentence citing it. Finally, a finding of no reliable evidence is treated as a successful answer. Many tools invert that last decision, because confidence sells in a product demonstration and honest absence does not.
The final layer is workflow rather than technology. Generated content should land in a review queue and never travel straight into a signed document. A hallucination caught at review costs a correction and a few minutes. A hallucination that leaves the practice under your signature can cost a notification. The design of the tool decides which of those outcomes you get.
The ADMRL position
Alix's research answers are retrieval-based, with every claim sourced, dated and linked. Unverifiable claims are flagged, and nothing Alix drafts enters a report without clinician sign-off.
Clinical Practice
Built for Australia first. On purpose.
Global platforms localise last, and it shows. Building in full depth for one profession in one country buys trust that generic software never earns.
Craig Adams · 5 min read
→
I have spent a career building and globalising software platforms, and the pattern repeats: build generic, sell everywhere, localise when the complaints get loud. It works for CRMs. It fails for clinical software, because the local details, Ahpra's expectations, the Privacy Act's treatment of health information, the NDIA's evidence bar, Medicare's item logic, are the product rather than the trim.
So we inverted the model. ADMRL is built for Australian psychologists in full depth before it is built for anyone else: Australian norms in scoring where publishers provide them, Australian templates in reporting, Australian data residency by architecture, and Australian regulatory assumptions in every workflow. The advisory years taught me that a profession gives its trust to platforms that were plainly built for it and withholds it from platforms that were plainly built for somewhere else.
Globalising comes later, and it comes as replication rather than dilution. The architecture travels: deterministic scoring, audit trails, the security model. Everything a profession actually touches gets rebuilt for each jurisdiction: norms, templates, regulatory assumptions, residency. The shortcut does not travel because the shortcut does not exist, and after a career of watching platforms attempt it, I regard that as the point.
The ADMRL position
ADMRL is built to Australian depth by design, across regulation, norms, residency and reporting, on a core architecture intended to replicate that depth in each new jurisdiction rather than dilute it.
Privacy & Data
De-identify first. Then ask the model.
The safest client data in an AI pipeline is the data that never enters it. The order of operations is the whole privacy story.
Sam Hallawell · 5 min read
→
Most privacy debates about AI in health concentrate on whether the model is secure. The load-bearing question sits earlier in the pipeline: what you sent it. A pipeline that redacts after processing has already disclosed. A pipeline that de-identifies before the model ever sees the text has changed the category of the problem, because what crosses the boundary is the clinical picture rather than the person.
Done properly, names, dates of birth, addresses, schools, employers and every other identifying fragment are detected and replaced with placeholder tokens before any AI involvement. Re-identification happens only inside the platform boundary, where the clinical record already lives and access is controlled. The model reasons about the client in the abstract, and it cannot leak what it never received.
The distinction worth demanding from any vendor is architecture versus policy. A policy says the vendor will not misuse your data, which is a promise kept by intention. An architecture ensures identity never reaches the model, which is a promise kept by construction. Only one of those survives a breach, a subpoena or an acquisition, and the difference decides what a bad day costs.
The ADMRL position
ADMRL detects and redacts identifying details before any model call, by construction. Client identity stays inside the platform boundary and is never part of the prompt.
NDIS & Funding
The template trap
Reviewers can spot a copy-paste report inside one paragraph. A template should structure the thinking while the client's data supplies the substance, and there is a simple editing test for telling the two apart.
Myra Jennings · 5 min read
→
Every NDIS-heavy practice runs on templates, and it should, because coverage, consistency and speed are legitimate clinical values. The trouble is the reader. The NDIA reviews reports in volume, and identical phrasing across different participants is one of the fastest credibility killers in the queue. The moment a report reads as generic, every specific claim inside it starts getting discounted, including the ones that are true and carefully evidenced.
The tension is real. Hand-writing every report from a blank page does not scale, and templating every sentence produces reports that describe nobody. The resolution is to be honest about what each layer is for. Structure belongs to the template: the sections, the evidence chain, the coverage of functional domains. Substance can only come from the data, because it concerns this child, this classroom, this score.
A useful editing test comes from the philosophy of science. Every sentence in the report should be falsifiable for this particular client, which means it could in principle be contradicted by their file. If a sentence could sit unchanged in the next client's report, it is upholstery rather than evidence. Reviewers recognise the difference at a glance, because reading reports is what they do all day.
The ADMRL position
ADMRL drafts each report from the client's actual file data into your template's structure, so the structure comes from the form and the substance comes from the file.
AHPRA & Regulation
Consent when AI is in the room
A consent clause that says "we use software" covers nothing. Clients need three plain facts about the AI involved in their care, and a genuine right to refuse it.
Justin Delahunty · 5 min read
→
Consent to "the use of software" covers nothing, because it means nothing. For consent to be meaningful, a client needs to know three things in plain words: what the AI actually does (it drafts and organises, under your review), what data it sees (a de-identified clinical picture), and who is responsible for the result (you, entirely). That is the whole disclosure, and it fits in four sentences.
Over-disclosure is its own failure mode. Three pages of vendor terms stapled to an intake pack transfers liability and calls it transparency. The research on informed consent has been consistent for decades: comprehension makes consent valid, and volume works against comprehension. Ahpra's 2025 practitioner guidance on AI points the same way, treating transparency about how AI is involved in a client's care as part of the practitioner's obligations.
Consent also has to be able to say no. If a client declines AI involvement, the workflow needs a fully manual path that does not punish them for asking: the same quality of report, in a timeframe you can defend. Consent that cannot be refused is notification. Build the refusal path first, and every yes you collect afterwards starts to mean something.
The ADMRL position
ADMRL ships plain-language consent wording for your intake pack, and every client file supports a no-AI pathway; the workflow runs manually, to the same standard, without penalty to the client.
Clinical Practice
What you do with the three hours
Nobody entered this profession to format score tables at 11 pm. The reclaimed hours go to waitlists, to deeper formulation, or to dinner, and returning that choice to the clinician was always the point.
Craig Adams · 5 min read
→
Ask a psychologist what they would do with three reclaimed hours and you learn what the backlog has been costing. Some name the waitlist and the families who have been holding for months. Some name depth, the formulation they wanted to sit with instead of racing past. A striking number simply say dinner, at the table, without a laptop open beside the plate.
Report writing after hours is the profession's quiet attrition engine. The work that drives people out is the fourth retyping of the same scores into the same table at 11 pm, unpaid, invisible and entirely separate from the clinical thinking that finished hours earlier. Burnout research keeps arriving at the same finding: chronic, unrewarded workload wears clinicians down, and the administrative layer is where that load accumulates.
Automation returns the hours without deciding what they are for. More clients seen sooner, deeper thinking per case, or a week that ends when it should are all legitimate clinical outcomes, and the point is that the choice becomes the clinician's again. A platform's twenty-minute workflow is the mechanism. The three hours it hands back are the product.
The ADMRL position
ADMRL's intake-to-signed-report workflow runs in about twenty minutes. What you do with the rest of the afternoon is the product.
Report Quality
Built to survive cross-examination
The question about AI-drafted reports is coming to Australian courtrooms, and the only good answer is a boring one. The record that makes an answer boring must exist before anyone asks.
Myra Jennings · 5 min read
→
Expert witnesses in Australian proceedings should expect to be asked whether artificial intelligence wrote their reports, and the question will become routine. The good answer is boring: drafting assistance was used, every finding, score and opinion is the practitioner's own, and here is the record of what was reviewed and changed. Boring answers end cross-examinations, and interesting answers extend them, sometimes by days.
What makes the answer boring is preparation that cannot be done retrospectively. It rests on an audit trail that distinguishes what was generated, what was edited and what was approved, on scores that reproduce exactly when the method is re-run against the raw data, and on citations that resolve to real sources saying what the report claims they say. Method gets examined long before opinion does, because method is where reports break.
The interesting answers come from the inverse position. The witness cannot say which sentences they reviewed. A standard score will not regenerate from the raw data. A reference the opposing side's junior spent an evening hunting for turns out not to exist. Each of these is a record-keeping problem rather than an AI problem, and AI has simply made every one of them faster to produce.
The ADMRL position
ADMRL keeps the record cross-examination asks for: per-section review history, deterministic and reproducible scoring, and citations that link to their sources.
AI Reliability
Citations that exist
A reference that does not resolve is worse than no reference at all. Retrieval systems that show their sources earn a place in clinical work, and systems that answer from memory forfeit one.
Justin Delahunty · 5 min read
→
A fabricated reference is worse than no reference. An absent citation leaves a gap in your argument; an invented one is evidence about your method, and once a single citation in a report fails to resolve, every other line in it gets read differently. Courts in several jurisdictions, including Australia, have already dealt with professionals who filed authorities that did not exist. Health practitioners should expect the same scrutiny.
The design answer is retrieval rather than memory. A system that searches the literature, quotes what it actually found, dates it and links it can be checked in seconds. A system that answers from its training data is asking to be believed, and in clinical work a claim you must take on trust has no standing as a citation. The difference between the two architectures is visible in the output: one shows its sources, the other performs confidence.
Currency matters as much as existence. The evidence base moves, and a 2019 answer to a 2026 question should announce its age instead of posing as current. Every retrieved claim should carry three tags, sourced, dated and linked, and when a claim cannot carry all three the system should say so plainly. "No reliable evidence found" is a professional answer. A confident invention is not an answer at all.
The ADMRL position
Alix's Research mode returns evidence that is sourced, dated and linked, and when the literature is silent Alix reports the absence instead of improvising.
Privacy & Data
Where your client data sleeps
Jurisdiction has become a clinical governance question, and a claim of Australian hosting has to cover every processing step, including AI inference, before it settles anything.
Sam Hallawell · 5 min read
→
The question of where client data is stored used to belong to IT. It now belongs to clinical governance, and "Australia" has to cover more than the database. Plenty of tools keep records onshore while sending every AI request overseas for processing, which means the clinical content of a session note has visited another jurisdiction and returned without anyone in the room knowing.
Residency decides practical things: which privacy regime applies, whose courts hear a dispute, what a foreign order can reach, and how breach notification works when the breach happened in a region you did not know your data visited. Your client asked where their file lives, and they believed the answer covered everything. If processing happens offshore, "mostly here" is the accurate version of what they were told.
So the useful vendor question has changed. Storage gets a rehearsed answer, so ask where every processing step happens, including AI model inference. If the reply names regions and providers in a sentence, the architecture is probably sound. If it runs to a paragraph, or leans on the word "generally", you have learned what you needed to know.
The ADMRL position
ADMRL runs end to end in Australian data centres, with storage, document handling and AI inference all onshore. The answer to the jurisdiction question fits in one sentence.
AHPRA & Regulation
The audit trail is the point
A clinical record should show what you decided and what you declined. Oversight you can prove beats oversight you assert, and a decision log is how the proof gets made.
Craig Adams · 5 min read
→
Record-keeping obligations were written for a world in which every sentence in the file was put there by a person, so authorship and decision were the same act. AI changes both the volume and the provenance of what enters a record. The old question asked what the file says. The modern record has to answer a second question as well: who decided it should say that.
An audit trail that captures what was generated, what you edited, what you rejected and what you approved does something subtle: it converts the sentence 'I review everything' from a claim into a record. When a complaint arrives eighteen months later, memory supplies a general account of your usual practice. The trail supplies the specific decision, with a timestamp attached.
The same trail protects you from the opposite accusation, the rubber stamp. A log full of your edits, your rejections and your overrides is documentary evidence that judgement was exercised on this report, on this date. The clinician using well-instrumented AI can end up holding better proof of independent thinking than the clinician who typed every word by hand and kept no history at all.
The ADMRL position
Every ADMRL report carries a decision log of what was generated, edited, rejected and approved, so your oversight exists as a record rather than a recollection.
Clinical Practice
Meehl was right, and it still stings
Seventy years of evidence show that structured prediction beats unstructured judgement. AI finally makes Meehl's method affordable in ordinary practice, and every clinical judgement call stays human.
Justin Delahunty · 6 min read
→
Paul Meehl published applied psychology's most uncomfortable book in 1954. Across one prediction task after another, simple actuarial combination of data matched or beat expert clinical intuition, and meta-analyses have replicated the result for seventy years, most thoroughly Grove and colleagues in 2000. The profession absorbed the finding the way professions absorb uncomfortable findings: it cited the book, taught it to students, and carried on combining data by judgement.
The finding is routinely misread as a claim that machines beat clinicians. It says something narrower and more useful: explicit rules, consistently applied, outperform impressionistic judgement at specifiable prediction tasks, whoever applies them. The operative variable throughout was method, structured combination against unstructured, and that distinction survives every generation of technology, including this one.
AI's genuine contribution is making structure affordable. The actuarial virtues, every score converted the same way, every domain covered, nothing skipped because it is 4:45 on a Friday, once cost hours of disciplined table work. They now run at the speed of conversation, with the clinician retaining every judgement call. Meehl practised as a clinician himself; his argument was with a method of combining data. Software built to his standard picks the same fight.
The ADMRL position
ADMRL operationalises the structured method, with consistent scoring, complete coverage and explicit reasoning, while every clinical judgement stays with you.
Report Quality
The blank page was never the problem
Clinicians do not struggle to think; they struggle to type the same findings a fourth time. Drafting and deciding are different work, and the order between them matters.
Craig Adams · 6 min read
→
The argument for AI in report writing is usually made in hours saved. The better argument concerns what the hours were spent on. Composition, the assembly of known findings into readable prose, is mechanical work. Formulation, the decision about what those findings mean, is clinical work. The blank page forces a psychologist to do both at once, in the same sitting, and that is precisely why report writing is miserable.
Separating the two changes the job. When a competent draft of the mechanical layer already exists, attention moves to the questions clinical training equips you to answer: whether a claim is accurate, whether the data support it, and whether you would sign the sentence as written. Editing engages critical judgement in a way that composing from nothing never did, because disagreement with a written sentence surfaces your actual position faster than a blinking cursor does.
That is why 'AI writes your reports' describes the right idea badly. For most clinicians the hard part of a report was always the typing rather than the thinking. Drafting and deciding are separate acts, the order between them matters, and the signature at the bottom should mean, verifiably, that a clinician did the deciding. A draft you can argue with is a better starting point for judgement than a page you have to fill.
The ADMRL position
ADMRL drafts the composition and you do the deciding, section by section, with an approval gate between every draft and your signature.
NDIS & Funding
Functional impact, in plain words
WHODAS scores tell the planner what is impaired, and the participant's Tuesday tells them why it matters. A fundable report needs both, in words a non-clinician can act on.
Myra Jennings · 5 min read
→
The strongest functional evidence in an NDIS report has two textures. The standardised kind, WHODAS 2.0 or Vineland-3 scores with percentile ranks, tells the planner what is impaired, in numbers that survive scrutiny and comparison. The lived kind, a plain account of what Tuesday morning looks like when the routine breaks, tells them why it matters for this participant. A report that carries only one of these textures is half a report.
Each texture fails alone. Numbers without narrative come back marked insufficient evidence of functional impact, because the planner cannot see the person in the percentiles. Narrative without numbers comes back as unsubstantiated, because sympathy is not evidence and a delegate's decision needs reasons that survive review. The craft sits in the pairing: every instrument score anchored to a concrete, observable consequence in the participant's actual week, stated at a frequency and level of support a stranger could verify.
The pairing has to arrive in plain words. The planner reading the report may have no clinical training, and clinician-speak forces them to guess at what was meant, which goes badly for participants. 'Cannot start self-care tasks without physical prompting at every step' does more work in a review queue than 'demonstrates deficits in adaptive functioning', and the first sentence is also the more precise clinical statement of the two.
The ADMRL position
ADMRL pairs instrument data with structured functional narrative in every NDIS report section, so the numbers and the lived detail sit side by side.
AHPRA & Regulation
Can the model see the test?
Test items are licensed commercial property, and item secrecy is an ethical obligation. What protects both the items and the client is where assessment documents go, and consumer AI tools are the wrong place.
Craig Adams · 6 min read
→
A photographed record form pasted into a consumer chatbot feels like efficiency. It is also a reproduction of copyrighted test items outside any publisher licence, and a breach of the test security that standardised administration depends on. The interpretation that comes back may even be sensible. The two professional problems created on the way have nothing to do with whether it is.
The obligations are older than the technology. Instruments such as the WISC-V are the commercial property of their publishers, licensed for administration and scoring within defined channels, and the profession's ethical standards require psychologists to keep test materials secure because norms assume examinees meet items cold. Content that reaches a consumer AI service may also be retained and used for training, and material that enters a training corpus cannot be recalled from it.
The way through is a governed boundary rather than a chat window. ADMRL works from the administration record the clinician already holds: documents are processed as text inside the platform, scoring runs in deterministic code, and the platform holds no item banks or stimulus materials, so it cannot reproduce what it does not have. Ask any vendor where uploaded documents go, who retains them, and what their models learn from them.
The ADMRL position
ADMRL works from the clinician's own administration record inside a contained Australian processing boundary, scores in deterministic code, and holds no publisher item banks or stimulus materials.
Clinical Practice
Base rates still apply
A positive result on a good instrument can be wrong more often than it is right. Predictive value collapses at low prevalence, and the workflow rather than memory should carry that arithmetic.
Justin Delahunty · 5 min read
→
A screening instrument with 90 per cent sensitivity and 90 per cent specificity sounds close to trustworthy, yet for a condition with 2 per cent prevalence it will produce roughly five false alarms for every correct identification. The deciding number is prevalence, and no score report prints it. Meehl and Rosen set out the argument in 1955, and assessment practice has been citing it and neglecting it ever since.
Walk the numbers once and the effect stops feeling like a technicality. Screening 1,000 children yields 116 positive results, and only 18 of them are correct. Move the same screener into a paediatric developmental clinic where referral has concentrated prevalence to 30 per cent, and a positive result is now right about four times in five. The instrument has not changed and the cutting score has not changed. The setting has, and an interpretation that ignores its setting is borrowing numbers it never examined.
Nobody recomputes predictive values at the end of a long testing day, which is the argument for building the denominator into the workflow instead of leaving it to memory. A prevalence estimate can sit next to the score while the interpretation is being made, and a drafted conclusion that outruns the numbers can be flagged before it hardens into a report. Base-rate neglect survives training. It does not have to survive design.
The ADMRL position
Alix presents base-rate context beside each drafted hypothesis, and flags draft conclusions that claim more certainty than the numbers support.
Report Quality
The error in every score
An obtained FSIQ of 98 is a band of plausible scores, and reporting it as a point claims precision the instrument cannot deliver. Measurement error belongs in the narrative as well as the table.
Myra Jennings · 6 min read
→
An obtained score is the visible half of a measurement. Classical test theory's plainest lesson supplies the hidden half: every score is a true score plus error, and the error is quantifiable. For a composite as reliable as a Wechsler Full Scale IQ, the standard error of measurement is about three points, so an obtained 98 is better read as a range spanning the low 90s to the low 100s. A report that hides the range claims knowledge the test never produced.
The practical consequences stack up quickly. Near the mean, four standard-score points swing ten percentile ranks, so measurement noise reads to families as movement between halves of the class. Differences between index scores carry the error of both numbers, which is why the manuals print critical values and base-rate tables. Retest scores arrive inflated by practice effects, with gains concentrated in performance and speed tasks, so a modest rise on reassessment is usually familiarity rather than growth.
The harm concentrates where scores gatekeep money and legal status. The DSM-5 builds measurement error into its own criterion, putting the relevant IQ range for intellectual disability at roughly 65 to 75, and a report that treats a 72 as exact invites a funding decision the data cannot support. Band language, a stated confidence level and adaptive evidence are what a defensible report carries. The honest sentence names the range and lets clinical judgement do its work in the open.
The ADMRL position
ADMRL reports every score with its confidence interval by default, and drafted narrative language is constrained to band descriptions that respect measurement error.
Privacy & Data
Small practice, big target
Health providers top the OAIC's breach reports, and the attacks that succeed are automated and indiscriminate. The strongest defences are cheap habits plus one architectural choice about where client documents live.
Sam Hallawell · 5 min read
→
Health service providers have topped the OAIC's notifiable data breach reports since the scheme began, and the pattern includes plenty of small practices. Attackers do not select victims; automated sweeps try leaked passwords and phishing lures against every account they can find, and a three-clinician practice gets exactly the same traffic as a hospital group, with a fraction of the defences.
The defences that work are unglamorous: unique passwords in a manager, multi-factor authentication starting with email, an account per staff member, prompt updates, deliberate disposal of old hardware. The habit that matters most is treating the inbox as a corridor rather than an archive, because years of referral PDFs sitting in a mailbox are a breach waiting to be scoped.
Hygiene lowers the odds of a breach, and architecture lowers the cost of one. Every copy of a client document is a separate thing to defend, so a practice where documents live in one governed system, behind access control and an audit log, has less to defend and far less to explain when the OAIC's 30-day assessment clock starts.
The ADMRL position
ADMRL keeps assessment documents inside the platform boundary, behind per-user access control and audit logging, instead of circulating them as email attachments.
AHPRA & Regulation
Supervision with AI in the room
A polished AI draft shows a supervisor very little about a registrar's competence. Supervision keeps its meaning when tool use is disclosed, drafts are compared with finals, and some work is done unassisted.
Myra Jennings · 6 min read
→
Supervision has always leaned on a quiet inference: if the trainee's reports are sound, the trainee is competent, because sound reports could only have come from competent hands. Generative drafting tools break the inference. A registrar can now submit polished work that reads as evidence while revealing very little about the reasoning underneath it, and a progress report signed on ambiguous evidence puts weight on a signature the Board treats as an instrument of public protection.
The workable response lives inside supervision itself. Tool use goes into the supervision agreement the way observation schedules and record keeping already do, machine drafts are kept, and the formative exercise becomes the comparison between draft and submitted final. Each change the trainee made is a clinical decision on display, and each unexamined acceptance is a finding in its own right. Alongside the comparisons, some competencies are demonstrated unassisted by design: live interpretation, formulation in the room, the occasional report written from a blank page.
Handled this way, AI can make supervision richer. A supervisor who reads finished reports sees conclusions; one who reads the trainee's decisions watches judgement form, which is the thing supervision exists to shape. The condition is a record showing what the trainee accepted, rewrote and rejected. Where that record exists, the supervisor's signature keeps its meaning, and the profession keeps its claim that a psychologist it registers can do the work without the tool.
The ADMRL position
ADMRL keeps a per-section decision log showing what a registrar accepted, rewrote and rejected, which gives a supervisor a reviewable record of the trainee's actual decisions.
Clinical Practice
Remote, but not diminished
Remote assessment has a real evidence base, and the evidence is instrument-specific. Competence means knowing what transfers under which controls, and declining in writing when the method cannot hold.
Justin Delahunty · 5 min read
→
For much of Australia the nearest assessment service is a long drive and a longer waitlist away, and video is often the difference between assessment this term and assessment next year. The pandemic made remote delivery general practice almost overnight, and it left the profession split between two positions: remote administration as a degraded copy of the real thing, and remote administration as the standard procedure with a webcam. The equivalence evidence supports neither position.
Equivalence studies predate 2020 and have multiplied since, and their pattern is consistent. Verbally mediated tasks, rating scales and structured interviews transfer well under good conditions. Tasks built on physical manipulatives or fine motor timing, Block Design being the obvious WISC-V example, do not. Equivalence attaches to a particular instrument, client and set of conditions, and treating it as a property of the technology is where remote assessment goes wrong.
The method includes its refusals. Camera position, third parties, connection quality and breaks are protocol variables that can be controlled and recorded, and administrations that cannot meet those controls should be declined and documented rather than improvised. Distance rations assessment in this country, and deferral has developmental costs of its own. A remote administration run as method, with conditions stated and failures declined, brings the standard assessment to families who live nowhere near a testing room.
The ADMRL position
ADMRL's eAssessment delivers instruments validated for remote administration, with integrity checks on timing and completion patterns and a flag on any protocol that warrants cautious interpretation.
AI Reliability
A scribe is not a platform
Ambient scribes and workflow platforms carry different risk surfaces, and each deserves its own evaluation. A scribe's risks follow the microphone, while a platform's follow provenance and review.
Sam Hallawell · 5 min read
→
Ambient scribes and clinical workflow platforms are both sold as AI for psychologists, and they share almost nothing underneath. A scribe captures everything said in a room and condenses it; a platform operates on structured assessment data under defined rules. Both categories have legitimate uses, and buyers keep getting them wrong by evaluating one with the other's checklist.
A scribe's risk surface follows the microphone: consent for recording, retention of raw audio and transcripts, processing jurisdiction, and generative summaries whose errors read fluently enough to pass a tired skim before entering the record. A platform's risk surface is provenance: whether each claim in a draft traces to an instrument response, an informant and a date, and whether a review gate stands between generation and the file.
The buying advice is to identify which machine is in front of you before asking anything else. For a scribe, ask about consent capture, audio retention and whether review is enforced. For a platform, ask where each generated sentence gets its facts and who approves it. Both sets of questions are short, and a vendor who cannot answer the set that matches their own product has answered a larger question.
The ADMRL position
ADMRL is workflow-native: structured assessment data with provenance, generation constrained by it, and review gates before anything enters the record. Transcription, where used, is session-scoped and consented rather than ambient.
Report Quality
When informants disagree
Cross-informant agreement in child assessment averages about .28, and the disagreement is systematic. A parent-teacher gap often carries the clinical story, and averaging it away discards signal.
Justin Delahunty · 6 min read
→
A parent rates a child well into the clinical range and the teacher rates the same child unremarkable, on the same scale, in the same fortnight. Assessment workflows tend to treat this as a problem to resolve, usually by arithmetic. The literature says otherwise. Achenbach, McConaughy and Howell put average cross-informant agreement near .28 in 1987, and decades of replication have barely moved it. Disagreement between home and school is the expected finding, and it is measurement working.
De Los Reyes and colleagues have shown the disagreement is systematic. Behaviour is context-dependent, and each informant is a valid observer of the setting in front of them. A gap on the SDQ or the BASC-3 often carries the clinical story: home-specific stressors, classroom-specific demands, or the masking that holds a school day together at a cost the classroom never sees. An average of discrepant informants describes a child nobody observed, and it dissolves the context effect the two ratings were jointly measuring.
The defensible report keeps observations separate and attributed, and it argues hypotheses about the gap instead of quietly crowning one informant accurate. The same standard applies to software, because a pipeline that blends informants into one number has already decided what the disagreement means, and no clinician made the decision. Multi-informant assessment promises coverage across the contexts of a child's life rather than consensus. When honest observers disagree, the assessment is usually working.
The ADMRL position
ADMRL presents cross-informant results side by side with discrepancies surfaced as findings; nothing in the pipeline silently averages informants together.
NDIS & Funding
Notes that survive a Medicare audit
Medicare audits are decided on what the record substantiates. Notes that capture who, when, duration and clinical content at the point of care carry the claim, and recollection carries nothing.
Craig Adams · 5 min read
→
A compliance letter asking you to substantiate twenty services from eighteen months ago is answered by one thing: the record. Medicare compliance activity has expanded in recent years, Better Access items are within its scope, and the practitioners most exposed are usually those whose work was sound but whose notes cannot prove it. At that distance, memory is testimony without corroboration, and the file does the talking.
The compliance standard is that records be adequate and contemporaneous: the note identifies who, when and for how long, with clinical content consistent with the item claimed, written at or near the time of service. The recurring failures are identical template notes across sessions, missing durations against time-defined items, notes reconstructed weeks later, and content that does not match the descriptor. Each one invites the conclusion that the record is an artefact of billing rather than of care.
Structure lowers the risk without adding administration. A note captured during or immediately after the session, prompted for duration, modality, interventions and plan, is audit-ready as a by-product of being accurate. This is governance rather than legal advice, and the governance is simple: the record is the only witness an audit will hear, so it should be written while the facts are still at hand.
The ADMRL position
ADMRL generates session records at the point of care, time-stamped and complete, so the note that substantiates the item is written while the work is being done.
Report Quality
Reports teachers actually read
Teachers open psychological reports more than anyone else, and most reports are written past them. A report earns classroom use with a plain first page and recommendations a school can implement, review and document.
Myra Jennings · 5 min read
→
Count the readers of a school-age assessment report and the classroom usually wins: the teacher and the learning support coordinator return to it more often than the paediatrician, the planner or the parents. Most reports are still written past them, with a jargon wall where the summary should be and recommendations pitched at a school that does not exist. What survives into the classroom is a sentence or two, and the child's adjustments stall with them.
The translation is a structural job rather than a stylistic one. A plain-language first page states what was found, what it looks like in a classroom and what helps, written plainly enough for a relief teacher to act on without a briefing. Recommendations pass a four-part test: specific enough to observe, realistic under current staffing, reviewable by a named date, and documentable as an adjustment. A working memory score at the 9th percentile is inert; the same finding restated as one instruction at a time, with a written backup, hands the teacher Monday's plan.
Schools must also evidence what they provide. Under the Disability Standards for Education 2005 and the Nationally Consistent Collection of Data, a school documents assessed needs and the adjustments made in response, and the psychologist's report is often the assessed-needs evidence itself. Recommendations a teacher can implement and review feed that record directly. The barrier has always been double handling, because the referrer needs the clinical document, the school needs the translation, and few practices have hours to write both.
The ADMRL position
ADMRL's audience-specific templates draft a clinician-grade report and a school-usable version from the same assessment data, so the translation costs a review rather than a second document.
No notes match the current filters.
Reading is one thing. Using it is another.
Everything argued above is built into the platform: deterministic scoring, review gates, de-identification, and evidence that cites itself. Start a free trial and put it to work on your next assessment.
Craig Adams · 5 min read
Myra Jennings · 6 min read
Justin Delahunty · 5 min read
Sam Hallawell · 6 min read