Naomi Alefelder.
From 10 December 2026, private and not-for-profit health care providers must disclose the kinds of significant decisions their software makes or substantially shapes.
Size is no exemption: the small business carve-out doesn’t apply to anyone who provides a health service and holds health information, so everyone from a solo GP or small physiotherapy clinic to a large not-for-profit aged-care provider or private hospital is covered.
The real work isn’t writing the disclosure; it’s finding and verifying every significant decision your software already influences. So start with one question:
Where is software making, or substantially shaping, decisions about the people in your care? Not the AI on next year’s roadmap, but the automation already running, switched on in a routine upgrade and labelled a “smart” feature. It never announced itself as AI, and nobody logged it as a decision.
New provisions in the Privacy Act (Australian Privacy Principles 1.7 to 1.9, introduced by the Privacy and Other Legislation Amendment Act 2024) create a mandatory transparency obligation.
In the words of the privacy regulator, the Office of the Australian Information Commissioner (OAIC): “APP entities must include information in their privacy policy where they have arranged for a computer program to make or inform decisions that significantly affect individuals’ rights or interests.”
An “APP entity” is any organisation or agency bound by the Privacy Act. State and territory public hospitals sit outside the Privacy Act, under separate state and territory privacy regimes.
1. It’s already inside your platforms, and “we don’t use AI” is not a defence
Automated decision-making is spread across the sector (hospitals, primary care, aged and disability care), rarely because a board chose it. It arrives bundled into the platforms you run: pushed in a vendor update, switched on by one department and never escalated, or inherited in a merger. Pre-existing software gets no exemption: the obligation applies to decisions it makes from 10 December, however long ago it was bought.
Whether it counts as “AI” is beside the point: the trigger is a computer program. Nothing has to learn or adapt. The regulator reads that term broadly, covering “pre-programmed rule-based processes, artificial intelligence and machine learning processes” alike. A fixed, points-based falls-risk score is a computer program. So is a hard-coded eligibility rule. Any tool however simple, is caught if it makes, or substantially and directly shapes a decision that could significantly affect someone, using their personal information.
The OAIC’s updated Guidelines (30 September 2026) list access to healthcare and aged care services among the interests that would “generally be considered in scope”: who gets care, how soon, and who is turned away.
Where a program’s output is materially different for a vulnerable group, the decision “should be taken to” be significant. The OAIC’s fact sheet makes the point with a spreadsheet, not AI: a provider ranks clients by need so staff can decide whom to contact. That is caught, and so is deciding whom not to contact.
2. “There’s a human in the loop” is the control everyone claims, and the regulator now looks straight at it
Ask how the risk is managed and you’ll hear the same line: a clinician reviews everything. The most cited safeguard in digital health, and the least scrutinised.
Picture a nurse on a late shift, clearing a queue of alerts between observation and medication rounds. Faced with a flood of mostly inconsequential alerts, people become desensitised and miss the real ones: alert fatigue is a recognised patient-safety hazard.
A 2024 meta-analysis found doctors override about 90 per cent of drug-interaction alerts, and nurses face the same pressure. Under real workload, “review” can become a reflex click rather than a judgement, which parks accountability at the bedside, far from the board that answers for the system.
The law catches two kinds of decision: those a computer makes on its own, and those where a program does a thing “substantially and directly related to making” the decision, such as scoring, ranking or recommending. The OAIC spells out what that means: the program is “a key factor in facilitating the human’s decision making” and has “a direct connection with making the decision”.
The regulator is explicit: a decision may be in scope “even where a computer program output does not replace the entire decision-making process or is subject to human review”. A human in the loop doesn’t, by itself, take a tool out of the net.
Machine learning and generative AI used in these decisions generally stay in scope “unless subject to extensive human oversight and control”. Genuine review may take a tool out: staff who interrogate the output, weigh other evidence and record their reasons when they don’t follow it.
A rubber stamp won’t. The Guidelines list the “ability and likelihood of human override” as a factor, and that is measurable: the override rate, the alerts per shift, whether the reviewer can actually disagree.
3. The deadline is closer than the readiness, and the usual exits are closed
There are three ways out an organisation might look for, and this obligation closes all three. First, it’s your duty, not the vendor’s: it binds whoever arranged for the software, and the OAIC’s examples include buying a third-party program, configuring off-the-shelf software or relying on its “advisory outputs”.
Vendors “should provide clear, high-level information about how their software can be used to make decisions”, so ask for it in writing. Second, “proprietary” is no shield: the commercial-in confidence exclusion covers trade secrets and commercially sensitive detail (how a tool weights its inputs), not the fact that a tool decides who is seen first.
Third, silence isn’t safe: from 10 December, a privacy policy that leaves out this information breaches APP 1.7, which the OAIC can act on with an infringement or compliance notice without first going to court.
You don’t need a new privacy policy, just a short new section in your existing one. For tools that are caught, APP 1.8 asks you to list the kinds of:
• personal information these programs use
• decisions they make on their own
• decisions they substantially and directly shape.
The fact sheet gives a model section on decisions made by computer programs. Grouping is allowed if the detail is “meaningful to a reasonable person”, and where health information is used, “this should be clear”.
By 10 December, you need two things: an inventory, built with your vendors’ input, of what each tool decides or shapes and what personal and health information it uses; and an updated privacy policy that, as the OAIC advises, takes “a cautious approach” and includes any tool you’re unsure about. After that, test whether your human review is real, and make contracts clear on who keeps the final decision.
4. Why this is a board and governance problem, not an IT line item
The vendor may have built the system, but your board answers for it. Company directors owe a personal duty of care and diligence under Corporations Act section 180; charity boards carry similar duties under the ACNC governance standards.
The Aged Care Act 2024 adds its own section 180: responsible persons who run or control a provider (its board and senior executives) must exercise due diligence, including understanding the risks its services pose, and can face civil penalties even if the provider isn’t found liable. Two section 180s, one question for every board: what did you do to know?
So a board should be able to answer three questions a regulator, a coroner or a plaintiff’s lawyer may one day ask:
(a) Where is automated decision-making already operating (including the rule-based tools nobody calls AI, and the ones nobody approved), and who signed off on each, against what evidence?
(b) Where a human reviews an automated output, have we tested that the review is real under actual workload, or are we assuming it?
(c) Does our privacy policy describe the kinds of significant decisions our software makes or substantially and directly shapes, and the kinds of personal information, including health information, it uses? If not, who owns closing that gap?
If today’s answer is “we’re not sure,” that is the finding. The first move is smaller than it sounds: name one owner, give them a date well before 10 December, and have them map where software is making, or substantially shaping, decisions about the people in your care.
Due diligence isn’t about slowing technology down; it’s about being able to answer for it, and keeping people safe while it runs. The technology can be bought in. The accountability can’t.
References available here.
Viewpoint articles are the author’s opinion, produced without payment or sponsorship.
Add Comment