🏥 RecovrSignal Issue 10: The algorithm assessors can't override, and the AI that's finally good enough to trust

Issue 10 · 21 September 2026 · Fortnightly | AI in Healthcare · Australia & Beyond

This fortnight starts with a warning from inside government. Freedom of information emails reveal officials knew, before the new Aged Care Act even commenced, that assessors would have no legal power to override the algorithm deciding a client's Support at Home funding, and that gap is now the subject of a Senate standoff. On the more reassuring side, a validation study finds that AI reading a phone camera can measure hip and knee range of motion about as reliably as you can with a goniometer. A Belgian survey confirms what you've probably already seen in session: people are turning to ChatGPT for mental health support it was never built or regulated to give. And a new AI model reads breast cancer pathology more accurately than the genomic test oncologists have relied on for over a decade.

Here's what you need to know:

Summary

Read time: ~4 min

Top Stories

• FOI emails reveal officials knew before the Aged Care Act began that assessors would have no legal power to overrule the Support at Home algorithm
• A validation study finds AI pose estimation matches manual goniometry for lower-limb flexibility and range of motion
• A Belgian survey finds people are turning to general-purpose AI chatbots for emotional and social support, not just information
• An AI model combining pathology, molecular and clinical data beat the 21-gene Recurrence Score at predicting breast cancer recurrence in the TAILORx cohort

Signals

• The Senate voted to reject the government's plan to cap DVA-funded allied health at $5,000 a year per veteran
• WHO's global AI-in-health forum says governance readiness, not deployment speed, should be the real measure of progress
• At-home care providers are being warned to slow down on agentic AI until data governance catches up
• The TGA has cleared no AI medical scribes, and some products may already be operating illegally

Top Stories

The Algorithm Assessors Can't Override, Even When It's Wrong

🇦🇺 Australia | Aged Care Assessment

The warning landed on 29 October 2025, three days before Support at Home began. An "URGENT" brief to Aged Care Minister Sam Rae, copied to Health Minister Mark Butler, said the Aged Care Rules had been drafted so tightly that the classification algorithm's ongoing assessments could not be legally overridden by a human assessor, whatever the assessor's own clinical judgement found.

Assessment organisations in Victoria, Western Australia, Queensland, New South Wales and the Northern Territory have all since written to the federal government with detailed complaints. Queensland Health alone sent a spreadsheet of more than 100 cases of "unexpected" outcomes, nearly all of them the algorithm under-assessing a person's needs rather than over-assessing them.

The Coalition, the Greens and the crossbench passed a bill in the Senate to restore assessor override. Labor blocked it from even being debated in the lower house. Health Minister Mark Butler has said he remains a "strong supporter" of the framework, and Minister Rae has put the cost of reversing it at $6 billion to $11 billion, while almost 154,000 people, according to a Four Corners investigation, are still waiting for their full home care package.

What it means for clinicians:

• If you complete functional assessments that feed into an IAT or Support at Home classification, document your clinical reasoning in full, in writing, every time. It is currently the only record that a human, rather than the algorithm, looked at the case properly.
• Tell clients and families explicitly that a request for review exists, and that review has produced results: 207 of the roughly 1,000 requests lodged in the tool's first five months succeeded. A wrong classification is not necessarily final.
• Expect this fight to keep running. If a client's package looks wrong against what you're seeing in their home, put it in writing to the assessment organisation now rather than waiting for a legislative fix Labor has already blocked once.

An AI Reading Your Phone Camera Can Now Measure a Hip or Knee About as Well as You Can

🌏 Global | Clinical Assessment Tools

The study measured the Active Knee Extension Test, the Modified Thomas Test and the Weight-Bearing Lunge Test, three assessments any physiotherapist or exercise physiologist runs routinely for hamstring, hip flexor and ankle flexibility. Camera-based AI pose estimation reached intra-rater reliability of 0.90 or higher and concurrent validity between 0.73 and 0.94 against manual measurement, in the same range as an examiner doing it by hand.

HRNet, the most computationally sophisticated of the three, outperformed OpenPose on the Modified Thomas Test specifically, which the authors attribute to it retaining high-resolution detail throughout its processing rather than reconstructing it from a lower-resolution image. MediaPipe and OpenPose still returned respectable measurement error of 1.8 and 1.7 degrees.

The catch is in the detail rather than the headline. The authors found systematic bias and wide limits of agreement on some measures, meaning the tools are reliable for tracking change in the same client over time but not yet a drop-in replacement for a goniometer when you need an exact absolute figure, such as for a funding submission.

What it means for clinicians:

• If you're evaluating a documentation or telehealth platform that advertises automated range of motion measurement, ask specifically which of these three tests it's been validated against, and whether that validation is published, not just claimed.
• Use a tool like this for tracking a client's own trend across sessions, where its reliability is strong, rather than for the single absolute measurement a funding body might scrutinise.
• This is a research validation, not a cleared product on the ARTG. File it as a category to watch over the next 12 months rather than something to add to your kit this week.

Your Clients Are Already Asking ChatGPT for Mental Health Support It Was Never Built to Give

🌏 Global | Patient-Facing AI

The researchers, from Thomas More University of Applied Sciences, University College Dublin and Queen's University Belfast, set out to establish how, why and by whom general-purpose AI chatbots are actually being used for mental health support, rather than assuming use mirrors the purpose-built mental health apps most existing research studies.

The finding that stands out is who is doing it. Use skewed toward people already familiar with professional mental health support, not people avoiding it. That reframes the usual assumption that chatbot use signals an access gap. For a meaningful share of users, it looks more like a supplement running alongside therapy than a replacement for it.

Nothing about ChatGPT is designed, tested or regulated as a mental health support tool, a gap Issue 9 already flagged in the AHPRA and university research context. This study adds the missing piece: real usage data showing the gap is not hypothetical. It is already how a meaningful number of people are using these tools, alongside whatever care you're providing them.

What it means for clinicians:

• Ask directly, as a standard intake question, whether a client is using any AI chatbot for support alongside your sessions. Most won't volunteer it unprompted, and it may be shaping what they bring into the room.
• If a client mentions using ChatGPT this way, don't treat it as a red flag on its own. This data suggests it's more often a supplement from someone already engaged with care than a sign of disengagement.
• Keep a plain-language line ready for when it comes up: general chatbots aren't built or regulated for this, and anything that sounds like a diagnosis or a crisis response from one needs to come back to you, not stay with the bot.

An AI Model Just Out-Predicted the Genomic Test Oncologists Have Used for Over a Decade

🌏 Global | Diagnostics

The model, built by the ECOG-ACRIN Cancer Research Group with Caris Life Sciences, combines a histopathology foundation model reading tissue slide images with molecular and clinical features, rather than relying on gene expression data alone. HR-positive, HER2-negative disease is the most common breast cancer subtype, accounting for around half of all cases.

The 21-gene Recurrence Score has guided treatment decisions, particularly whether a patient needs chemotherapy on top of hormone therapy, since TAILORx reported its practice-changing results in 2018. A model that improves on it changes who gets recommended for more aggressive treatment and who is safely spared it.

This is a research result, not yet a clinical tool available on referral. But it points at where oncology risk stratification is heading: multimodal AI reading tissue images, molecular markers and clinical history together, rather than any single test read in isolation.

What it means for clinicians:

• If you work with clients through breast cancer treatment and survivorship, expect recurrence risk categories to keep shifting over the next few years as tools like this reach clinical use. Build a habit of asking whether a client's risk category has been recently reassessed rather than assuming it's fixed at diagnosis.
• This is a prediction tool, not a diagnosis. A lower predicted recurrence risk doesn't change your own clinical monitoring for the functional and psychosocial effects of treatment.
• File this under evidence to watch rather than something to raise with clients yet. It hasn't reached clinical practice, and Australian oncologists won't have access to it for some time.

Signals

🇦🇺 Australia | DVA Funding: On Monday 8 September, the Senate passed an urgency motion against the $5,000 annual cap on DVA-funded allied health due to start 1 July 2027, with the Coalition and crossbench voting together against the government. The motion doesn't repeal the cap itself, but it's a formal signal the Senate wants it abandoned before it takes effect, alongside the 45.6% fee increase also legislated for that date. If you see DVA clients, this is worth tracking over the next few sittings rather than treated as settled either way. Read →

🌏 Global | AI Regulation: Published 1 September, findings from WHO's first Knowledge Community on responsible AI in health, drawing on 105 countries, conclude that the central obstacle to safe AI in healthcare isn't the technology, it's whether the institutions meant to govern it are ready. Nothing binding follows immediately, but it's the same governance-first framing to expect TGA and AHPRA guidance to keep drawing on. Read →

🌏 Global | Practice Admin: A September briefing for US home and community care providers argues the compliance risk in agentic AI, tools that act on scheduling, documentation or care plans with less human review, sits mainly in data ownership and vendor control rather than the AI itself. The advice translates directly for any Australian practice trialling similar automation: run an internal data review before adoption, not after, and know exactly what a vendor's tool can act on without you checking first. Read →

🇦🇺 Australia | Vendor Claims: As of 15 September, the TGA has approved zero AI medical scribes, and has signalled that products offering clinical suggestions rather than pure transcription may already be operating in breach of the law, with findings from its year-long review due in coming weeks. The bigger risk flagged this fortnight isn't fabrication, the kind Issue 8 covered, but omission: a scribe that quietly drops something you said, with no error message to catch it. Read →