AI Regulation Tracker  /  Regulator guidance

Australia's TGA: AI-based medical software must meet SaMD rules; scope changes need regulatory action

The Therapeutic Goods Administration has set out how it regulates AI-based medical device software, confirming that AI features sit inside the existing Software as a Medical Device framework and do not create a new device class. Manufacturers must not ship changes that alter a product's intended purpose until the change is re-approved.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

The Therapeutic Goods Administration has set out how it regulates AI-based software that functions as a medical device, and the message to manufacturers is that little about the rulebook is new. Software that meets the definition of a medical device is regulated as Software as a Medical Device, or SaMD, under the Therapeutic Goods Act 1989 and the Therapeutic Goods (Medical Devices) Regulations 2002. Adding artificial intelligence or machine learning to a product does not move it into a special category, and it does not open a shorter route to market. The obligations that already applied to medical device software apply to AI products in full.

AI does not create a new device class

The central point of the guidance is one of continuity. Classification is driven by intended purpose, not by the technology inside the product. A tool that informs diagnosis or treatment carries the classification that its clinical role dictates, whether its logic is a fixed rule set or a trained model. The TGA does not treat AI as a reason to raise or lower that classification on its own. For sponsors, this means the entry point is the same question that governs any SaMD, which is what the software is for and what claims are made about it.

The full SaMD obligations apply

Because AI products are SaMD, they carry the standard duties. Manufacturers must define and document the intended purpose, apply the correct classification, and demonstrate compliance with the Essential Principles for safety and performance. They must hold clinical evidence appropriate to the product's risk and claims, and they must operate a quality management system that controls design and change. The recognised standards that manufacturers use to show this include ISO 14971 for the application of risk management to medical devices and IEC 62304 for medical device software lifecycle processes. These standards are the accepted means of meeting the Essential Principles rather than a separate AI checklist. The product must be included in the Australian Register of Therapeutic Goods, or ARTG, before it is supplied, unless a specific exemption applies.

Change control is the sharpest edge

The obligation most likely to catch AI teams is change control. A manufacturer cannot ship a feature update, a retrained model, or an expanded indication that alters the product's intended purpose or performance and then seek approval afterward. Changes of that kind require regulatory action first, which may mean notifying the TGA or updating the device's ARTG inclusion before the change reaches users. This is the point where continuous-delivery habits collide with device law. An adaptive model that drifts beyond its approved intended purpose is, in regulatory terms, a different device. The guidance frames scope creep through updates as a compliance failure, not a product improvement to be regularised later.

The practical effect is that release management becomes a regulatory checkpoint. Teams that plan model retraining, add clinical features, or broaden the patient population need a documented view of the approved intended purpose and a way to classify each change against it. Updates that stay inside that approved purpose can proceed under the quality system in the ordinary way. Updates that reach outside it must wait. The safest posture is to decide the regulatory status of a change before writing the code that ships it, rather than after users have it in hand.

What the guidance does not do

The guidance does not exempt AI from the medical device framework, and it does not lighten the evidence burden for products that use novel methods. It also does not remove the sponsor's ongoing responsibilities once a product is on the market, including adverse-event reporting and post-market monitoring. What it does is close the argument that AI is too new or too different to fit existing rules. The TGA's answer is that the rules already reach these products.

The cross-border angle

For a United States reader, Australia is a separate jurisdiction with its own registration regime, so a SaMD supplied into Australia needs ARTG inclusion regardless of any clearance held elsewhere. The direction matters beyond compliance in one market. The TGA's approach tracks the international pattern of treating AI as SaMD under existing device law and tightening control over post-approval changes, which previews the way regulators including the US Food and Drug Administration are handling model updates and predetermined change control. Firms building AI products for global supply are converging on the same operating standard, and the TGA guidance is one clear statement of it. The guidance follows the TGA's earlier work on AI in healthcare.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.