TGA finds most AI clinical scribes are likely medical devices

When your GP reaches for a phone or taps an icon before your consultation begins, you may not know what software is about to process your conversation. AI clinical scribes — tools that listen to consultations and generate clinical documentation — are now used by a significant and growing proportion of Australian GPs. The premise is sensible: reduce administrative burden, free the doctor to actually look at you rather than a screen.

This week, the TGA weighed in on a question that has been circling since these products arrived in GP software stacks: are they medical devices?

For most of them, the likely answer is yes.

What the TGA found

The TGA reviewed 12 AI scribe products chosen to reflect market share while also including smaller suppliers. The 12 sit within a larger landscape — the TGA identified approximately 100 AI scribe suppliers in Australia, of whom 40 responded to a voluntary survey. Products reviewed included Heidi Health, Lyrebird Health, Microsoft Dragon Medical One, Abby Health, AskSam, and several others.

Eight of the twelve were found to likely meet the legal definition of a medical device under the Therapeutic Goods Act 1989. Two required more information. Two were unlikely to qualify. None of the twelve had self-identified as medical devices — meaning none had registered with the TGA as such, even where that registration may be legally required.

The TGA did not test the products, assess clinical safety or performance, or examine privacy and data protection practices. The review was about regulatory classification, not clinical safety assessment.

The threshold: when does software become a medical device?

The distinction the TGA is drawing here is important for both clinicians and patients.

Pure transcription — software that converts speech to text and produces a verbatim consultation record — is unlikely to meet the medical device definition. It is doing what a digital recorder would do, faster.

What crosses the threshold is clinical reasoning. When software analyses clinical information and generates findings, clinical conclusions, or recommendations that could affect patient care decisions, it enters the device definition. Newer AI scribes — particularly those using large language models — increasingly do exactly this. They do not just record that you said “my knee has been aching for three weeks.” They may classify the symptom, suggest a clinical framing, or pre-populate items in a care plan template.

The TGA added a point that cuts through a common industry workaround: disclaimers saying the product is “not a medical device” are legally insufficient to determine regulatory status. What the software actually does is the deciding factor — not how the manufacturer chooses to describe it.

Both-and: useful technology and incomplete governance

The honest framing here is both-and.

AI scribes reduce clinician administrative load. Documentation, referral letters, care plans — these consume hours that could otherwise be spent on clinical work. Software that handles that load without sacrificing quality has genuine value, and the early evidence from Australian adopters is broadly positive.

At the same time, none of the major products now used in Australian general practice are registered as medical devices, even for those that likely meet the definition. Registration requires demonstrating safety, performance, and quality. The TGA has confirmed it will take a “proportionate and risk-based” compliance approach — which likely means education and guidance before enforcement, rather than immediate market withdrawal.

The governance concern the TGA named directly is worth sitting with: AI-generated clinical outputs can make it difficult for clinicians to track how information was selected or prioritised. When a clinical note is generated by an algorithm, the reasoning is not visible the way a handwritten note is. What the AI emphasised, what it left out, and why — these become opaque unless the clinician actively cross-checks the output against their own recall.

That is not an argument against the tools. It is an argument for knowing they are there and for clinicians maintaining independent clinical judgment rather than accepting AI-generated summaries uncritically.

2 cents

The TGA finding does not mean AI scribes should be switched off. It means the regulatory framework needs to catch up with the practice reality that is already happening in consulting rooms across Australia — and that both clinicians and patients benefit from a moment of explicit awareness.

If you are curious: it is entirely reasonable to ask your GP which scribe tool they use, whether AI is generating any clinical synthesis during your consultation, and how the outputs are checked before entering your records. Most GPs using these tools have thought about those questions. The asking does not need to be adversarial — it can be part of understanding your own healthcare.

The larger question this story opens — what are Australian patients entitled to know about AI’s role in generating their clinical records? — is one the TGA finding moves forward but does not resolve.

Verdict: maybe — the regulatory finding matters, and proper medical device registration would improve accountability; the clinical practice benefits are real, but the governance framework around these tools remains a work in progress.

---

Sources cited

1. Eight of 12 AI scribes may be medical devices, TGA finds — Medical Republic
2. TGA — About medical devices