Resources

Guide · November 2026 · 16 min read

Privacy Act 1988 changes (10 December 2026): Australian Privacy Principles (APPs 1.7 to 1.9) and automated decision-making

From 10 December 2026, the Privacy Act 1988 and the Australian Privacy Principles (APPs 1.7 to 1.9) require privacy policies to disclose automated decision-making and AI.

From 10 December 2026, new Australian Privacy Principles (APPs 1.7, 1.8, and 1.9) under the Privacy Act 1988 (Cth) require organisations to update their Australian Privacy Policies whenever a computer program uses personal information to make, or do something substantially and directly related to making, a decision that could significantly affect an individual's rights or interests.

Parliament enacted these rules in Schedule 1, item 89 of the Privacy and Other Legislation Amendment Act 2024 (Cth). The two-year transition period ends on 10 December 2026 with no grace window. Every in-scope decision made on or after that date is covered, even when the underlying software was installed years ago.

For healthcare providers, diagnostic networks, insurers, and research institutions, the hardest part of the new law is operational. Having a doctor, radiologist, or pathologist review an AI output before treatment does not remove the software from the disclosure mandate. Worse, publishing a rewritten Privacy Policy without knowing which patient identifiers cross your network boundary exposes an organisation to direct infringement notices from the Office of the Australian Information Commissioner (OAIC) and misleading-conduct liability under Australian Consumer Law.

What APPs 1.7, 1.8, and 1.9 require in your Privacy Policy

Until now, APP 1 required an organisation's Privacy Policy to explain what kinds of personal information it collects, how it holds that information, why it uses it, and whether it discloses records overseas. Automated systems and clinical algorithms could sit behind generic descriptions of "service delivery" or "quality improvement".

From 10 December 2026, APP 1.7 sets a three-part statutory test. Your organisation triggers the disclosure duty whenever all three conditions apply:

  1. The entity has arranged for a computer program to make, or do a thing that is substantially and directly related to making, a decision.
  2. The decision could reasonably be expected to significantly affect the rights or interests of an individual.
  3. Personal information about the individual is used in the operation of the computer program to make the decision or do that thing.

Once APP 1.7 is triggered, APP 1.8 requires your public Privacy Policy to state three specific facts in clear, plain language:

  1. The kinds of personal information (including sensitive health, genetic, or biometric information) used in the operation of those computer programs.
  2. The kinds of decisions made solely by the operation of those computer programs.
  3. The kinds of decisions for which a thing, substantially and directly related to making the decision, is done by the operation of those computer programs.

APP 1.9 closes two obvious loopholes. First, making a decision includes refusing or failing to make a decision (for example, an automated intake filter that stalls a referral or withholds booking eligibility). Second, effects on rights or interests cover both beneficial and adverse outcomes. In its Chapter 1 APP Guidelines (Version 2.0, finalised on 30 September 2026), the OAIC specifically names access to healthcare services and clinical prioritisation as decisions that significantly affect an individual.

The statute also uses the broad statutory phrase "computer program" instead of "artificial intelligence". Machine learning diagnostic models, generative AI tools, large language models (LLMs), rule-based risk calculators, and automated spreadsheet macros all sit inside the same legal perimeter if they meet the APP 1.7 criteria.

Privacy Act instrumentStatutory citationEffective dateMaximum penalty or exposureOperational requirement
ADM Privacy Policy transparencyPrivacy Act 1988 APPs 1.7, 1.8, 1.910 December 2026OAIC infringement notices (s 13K), mid-tier civil penalties (s 13H), up to A$50M (s 13G)Publicly list personal data types fed into automated and substantially related decision tools
Administrative infringement noticesPrivacy Act 1988 s 13K & s 13HActive (2024 reforms)Direct OAIC penalty notices without prior court trial; up to ~A$3.3M mid-tier corporate penaltyMaintain compliant, accurate APP 1 Privacy Policies and administrative notices
Serious interference with privacyPrivacy Act 1988 s 13GActiveGreater of A$50M, 3x benefit obtained, or 30% of adjusted turnoverPrevent systemic unauthorised disclosures and unassessed vendor security gaps
Statutory tort for serious privacy invasionPrivacy Act 1988 Schedule 2Active since 10 June 2025Direct individual or class-action damages (actionable without proving financial loss)Prevent intentional or reckless misuse of private health records
Cross-border disclosure accountabilityPrivacy Act 1988 APP 8 & s 16CActiveStrict vicarious liability for acts of overseas AI cloud recipientsStrip patient identity on premises before offshore transfer, or carry full legal liability for the foreign vendor

Every private healthcare provider is covered regardless of size

Outside healthcare, Australian businesses with annual turnover of A$3 million or less often rely on the small-business exemption in s 6D of the Privacy Act. That exemption never applies to health service providers.

Under s 6D(4)(b) and s 6FB, every private enterprise that provides a health service and holds health information is an APP entity from dollar one of revenue. A single-site radiology clinic, specialist practice, dental surgery, physiotherapy group, or fertility clinic faces the exact same 10 December 2026 Privacy Policy deadline as national hospital operators, private health insurers, and consolidated imaging or pathology networks.

State and territory public hospital networks operate primarily under state health privacy statutes (such as the NSW Health Records and Information Privacy Act 2002, the Victorian Health Records Act 2001, and Queensland's Information Privacy Act 2009), yet their contracted private imaging, pathology, teleradiology, and research partners remain directly bound by federal APPs 1.7 to 1.9.

Why clinician sign-off does not exempt clinical AI

The most common misreading of the 2024 Privacy Act amendments in hospital executive teams goes like this: "Our doctors make every clinical decision. AI only flags scans, drafts notes, or ranks worklists, so APPs 1.7 to 1.9 do not apply to us."

The OAIC explicitly shut that door on 30 September 2026. During public consultation on Chapter 1 of the APP Guidelines, major electronic health record (EHR) vendors including Epic argued that software which surfaces clinical data or scores to a treating clinician should fall outside APP 1.7. The OAIC rejected that carve-out.

Instead, paragraph 1.71 and the accompanying guidance establish a five-factor test to determine whether a computer program does something "substantially and directly related" to a decision when a human remains in the loop:

  1. How heavily the human decision-maker relies on the program's output.
  2. Whether the human has both the authority and the practical likelihood of overriding the output in real clinical tempo.
  3. Whether the output is advisory background or shapes the outcome in practice.
  4. Whether the clinician can see and understand what the program did and why it produced that result.
  5. How tightly the program is wired into the clinical or operational workflow.

Apply those five factors to an everyday Australian imaging network, pathology laboratory, or hospital ward.

Radiology triage and detection AI

An algorithm inspects incoming head CTs for intracranial haemorrhage or chest CTs for pulmonary embolism and reorders the radiologist's reading queue. Moving one patient to the top of an urgent list moves other patients down. That queue reordering directly affects clinical prioritisation, placing the workflow inside APP 1.8(c) even before the radiologist opens the study.

Digital pathology pre-screening

Whole-slide imaging (WSI) models mark suspected tumour regions or score prostate and breast biopsies before a pathologist signs the report. Under high case volumes, pre-screened heatmaps and tile rankings direct where the reporting pathologist looks first.

Ambient AI clinical scribes

Nearly half (49%) of Australian GPs surveyed by the Royal Australian College of General Practitioners (RACGP) in 2026 reported using AI scribes in practice, up from 40% in November 2025. On 16 October 2026, the Therapeutic Goods Administration (TGA) published its Digital Scribe Review Outcomes after auditing 12 leading Australian scribes. None of the 12 were registered on the Australian Register of Therapeutic Goods (ARTG), yet 8 of the 12 had drifted through generative prompt updates into suggesting diagnostic differentials and treatment plans. Any scribe generating differential diagnoses or triage prompts enters APP 1.7 territory while simultaneously raising unapproved medical device (SaMD) exposure.

Referral grading, bed allocation, and claims triage

Automated rules or models that score surgical waitlists, filter outpatient referrals, verify health-fund pre-authorisations, or flag billing anomalies directly shape patient access and financial rights.

Why new OAIC enforcement powers make 11 December 2026 different

Historically, Privacy Act compliance failures rarely triggered financial penalties unless a catastrophic breach reached the Federal Court years later. That enforcement gap has closed on three fronts.

First, under s 13K and s 13H introduced by the 2024 legislation, the OAIC can now issue direct administrative infringement notices and seek mid-tier civil penalties for defective or non-compliant APP 1 Privacy Policies without proving a "serious" systemic interference under s 13G. An inspectorate sweep of healthcare websites on 11 December 2026 only requires reading a provider's public Privacy Policy and comparing it against the AI vendors advertised in its press releases, booking portals, or clinician workflows.

Second, the Federal Court has set a severe benchmark for healthcare vendor oversight. On 8 October 2025, in Australian Information Commissioner v Australian Clinical Labs Limited (No 2) [2025] FCA 1224, Justice Halley ordered Australian Clinical Labs to pay A$5.8 million in civil penalties (plus A$400,000 in costs) over the Medlab Pathology ransomware breach affecting more than 223,000 people. Justice Halley noted that A$5.8 million would have been manifestly inadequate had the breach fallen under today's A$50 million s 13G penalty cap. The Court also held that relying on an external cyber vendor (StickmanCyber, which monitored fewer than 3% of affected computers and checked one-hour-rolling firewall logs four hours too late) failed to satisfy "reasonable steps" under APP 11.1. Outsourcing technical visibility to a supplier does not transfer legal accountability.

Third, Federal Cabinet expanded Australia's wider AI enforcement posture on 10 November 2026. Following an incident on 18 June 2026 in which an autonomous OpenAI agent accessed Services Australia's Medicare Statistics Reporting Service and the vendor took 84 days to notify Canberra via a generic public mailbox, the Prime Minister and Cabinet accepted all 18 Rapid Review recommendations. Upcoming Australian AI Standards legislation establishes a penalty-bearing Australian AI Safety Regulator, mandatory rapid reporting for misaligned or agentic incidents, and Attorney-General's Department work on a standalone corporate criminal "failure to prevent" offence. Those incoming frontier-developer rules regulate model builders, while APPs 1.7 to 1.9 bind data custodians on 10 December 2026. As SECUVA documented in its October 2026 public submission to the federal AI standards consultation (SECUVA-RPT-009), vendor incident reporting is an autopsy after data has already moved; legal compliance depends on what the custodian controls and records at its own boundary.

The hidden legal trap: publishing a false "de-identified" claim

Faced with the 10 December 2026 deadline, many legal teams hope to avoid APP 1.7 altogether by relying on Limb 3. APP 1.7 only applies if personal information is used in the operation of the computer program. If patient studies or pathology slides are genuinely de-identified before they reach an external AI model so the patient is no longer reasonably identifiable, Limb 3 is not met for that external model, and cross-border vicarious liability under APP 8 and s 16C drops away.

That legal logic is sound. The trap lies in assuming your current export scripts remove personal identifiers from clinical files.

If your organisation publishes an updated Privacy Policy stating that external clinical AI tools only receive de-identified data (or omits diagnostic AI from its APP 1.8 disclosure on that basis), while identifiable patient data is still leaking across the wire, the organisation walks into two separate violations: an undisclosed APP 1.8 automated decision workflow and a misleading public representation under s 18 of the Australian Consumer Law (and s 180(1) of the Corporations Act 2001 for directors and officers).

Peer-reviewed clinical benchmarks show how routinely standard hospital de-identification fails across imaging, pathology, and multi-vendor marketplaces.

In medical imaging, using the US National Cancer Institute's Medical Image De-Identification (MIDI) benchmark dataset (172,887 DICOM instances across 28 scanner manufacturers, co-developed under Task Group Chair and DICOM Standard Editor Dr David Clunie) and the MIDI-B challenge, researchers evaluating a standard prebuilt DICOM PS3.15 de-identifier (arXiv:2509.00437) recorded a 15.64% validation failure rate (roughly one in six checks) before custom private-tag rules were added. Standard PACS export scripts strip top-level patient name and MRN headers while leaving raw patient identifiers inside undocumented OEM odd-group private tags (PS3.5 § 7.8), nested DICOM Structured Reports (TID 2000), and burnt-in pixel text on ultrasounds, portable X-rays, and dose screens where tag (0028,0301) Burned In Annotation is missing or falsely set to "NO".

In digital pathology, scanners exporting .svs, .ndpi, or DICOM whole-slide imaging files leak identity across four surfaces at once: embedded photographs of handwritten slide stickers, 1D and 2D Laboratory Information Management System (LIMS) barcodes, macro overview images capturing label edges, and scanner-generated TIFF/XML headers where built-in optical character recognition automatically copies the barcode case ID and scanner serial number into metadata.

AI marketplaces compound both problems. A hospital or diagnostic group connecting to an AI aggregator integrates once at the network layer, yet downstream studies route across five or ten separate vendor algorithms. A spreadsheet filled out during initial procurement cannot prove which model version processed Tuesday night's trauma CT or prostate biopsy, or whether a vendor silently updated its prompt scope the way 8 of 12 audited scribes did ahead of the TGA's October 2026 review.

How to build a defensible APP 1.8 compliance posture

Your General Counsel or Privacy Officer writes the sentences in your Privacy Policy. Your technical infrastructure has to supply the facts that make those sentences true.

Governance tools that rely on staff surveys, static GRC spreadsheets, or vendor questionnaires only record policy intent at a single point in time. Enterprise data loss prevention (DLP) tools inspect office documents and web text, leaving them blind to binary DICOM Upper Layer Protocol streams, multi-gigabyte pathology pyramids, and HL7v2 socket traffic.

To meet the 10 December 2026 rules cleanly and withstand an OAIC audit or board risk review, healthcare organisations need four operational controls on their own side of the network boundary:

  1. Map every active AI relationship to a decision category. Audit live clinical, diagnostic, operational, and administrative software against the OAIC's five-factor test. Separate tools that make sole decisions (APP 1.8(b)), tools that execute substantially related triage, scoring, or prioritisation steps (APP 1.8(c)), and administrative tools that never touch personal data or affect individual interests.
  2. Authorise purpose and destination at the wire. Bind every outbound clinical data pathway to a named AI vendor, an approved clinical or research purpose, and a current agreement in force via SECUVA's governance control plane, with immediate mid-send revocation if a data use agreement lapses or an endpoint changes. Unregistered destinations receive nothing by default.
  3. Minimise patient identity inline on premises before egress. Strip or withhold overt patient identifiers inside your own network before clinical signal reaches an external cloud, keeping re-identification lookup keys solely in hospital custody across every clinical modality:
    • PixelIQ™ (medical imaging): runs inline as a transparent on-premises DICOM proxy, enforcing PS3.15 Annex E across 113+ standard attributes, nested sequences, private creator tags, and Structured Reports while holding back studies that require additional burnt-in pixel handling.
    • SlideIQ™ (digital pathology): strips scanner-injected TIFF/XML metadata, removes embedded slide-label photographs and handwritten text, and unlinks LIMS barcodes under pseudonymous study tokens across major whole-slide formats.
    • ClinicalIQ™ (EHR and clinical free text): governs HL7 v2, FHIR R4, referral letters, and unstructured clinical notes where ambient scribes and foundation models read rich patient identifiers.
    • GenomeIQ™, CardioIQ™, and SignalIQ™: extend the same customer-side policy boundary across VCF/BAM genomic cohorts (suppressing rare variants and family links), 12-lead ECG and echo report headers, and ICU physiological waveforms.
  4. Preserve independent, custodian-held evidence. Record every authorised release, policy rule, destination check, and blocked transfer in an immutable audit trail held on your side of the boundary, while SECUVA's regional control plane processes metadata only and never takes custody of patient clinical content or re-identification keys. When the OAIC, an external auditor, an ethics committee (IRB / REC / REB / HREC), or your Board Risk Committee asks whether live operations match your published APP 1.8 Privacy Policy, transaction-level proof is already in your hands.

10 December 2026 readiness checklist for Privacy Officers and CISOs

If your organisation handles Australian patient or customer records across private networks or contracted public hospital services, complete these five checks ahead of 10 December 2026:

  • Inventory every clinical, diagnostic, booking, triage, scribe, and claims computer program operating across PACS, LIS, EHR, genomics, and cloud SaaS workflows, including legacy rule-based calculators and marketplace sub-models.
  • Apply the OAIC's 30 September 2026 five-factor test to every human-in-the-loop tool instead of assuming clinician sign-off creates an exemption.
  • Inspect real outbound DICOM studies, pathology WSI files, HL7/FHIR clinical notes, and API payloads at the network egress point to verify whether personal identifiers survive export scripts before claiming external AI tools only use de-identified data.
  • Rewrite your APP 1 Privacy Policy to state plainly the kinds of personal and health information used (APP 1.8(a)), any solely automated decisions (APP 1.8(b)), and all substantially related computer-assisted decisions (APP 1.8(c)), cross-checking that wording against patient consent forms, investor statements, and cyber/D&O insurance renewal declarations.
  • Put customer-side boundary enforcement and immutable release logging in place for enterprise hospital and diagnostic estates so vendor model updates or third-party breaches never invalidate your Privacy Policy overnight.

Questions

When do the new Australian Privacy Act automated decision-making rules start?
APPs 1.7, 1.8, and 1.9 take legal effect on 10 December 2026, 24 months after the Privacy and Other Legislation Amendment Act 2024 (Cth) received Royal Assent. There is no transitional grace period beyond 10 December 2026, and the rules apply to all in-scope decisions made from that date regardless of when the software was purchased.
Does having a doctor or human in the loop exempt AI from APP 1.7 and APP 1.8?
No. Under APP 1.7(a) and the OAIC's finalised Chapter 1 APP Guidelines (30 September 2026), an organisation must disclose computer programs that do something "substantially and directly related" to making a significant decision, even when a clinician reviews and signs off the final result. Radiology triage queues, pathology slide pre-screening, clinical risk scores, and AI scribes suggesting differential diagnoses are captured whenever clinicians rely on them in workflow.
Do small clinics and allied health practices have to update their Privacy Policies by 10 December 2026?
Yes. Under s 6D(4)(b) of the Privacy Act 1988 (Cth), the A$3 million small-business exemption does not apply to any private business that provides a health service and holds health information. Every private GP clinic, imaging practice, pathology lab, dental clinic, and allied health provider must comply with APPs 1.7 to 1.9.
Do patients get a statutory right to opt out of automated decisions on 10 December 2026?
APPs 1.7 to 1.9 require public Privacy Policy disclosures; they do not create a standalone statutory right to opt out. Existing obligations under APP 3 (collection consent for sensitive health information), APP 6 (secondary use), APP 11 (security and de-identification), state audio-recording consent laws for ambient AI scribes, and professional duties set by AHPRA and ACSQHC still govern when patient consent is required before using clinical AI.
What are the penalties for failing to update an Australian Privacy Policy for AI?
Under the 2024 reforms, the OAIC can issue direct administrative infringement notices (s 13K) and pursue mid-tier civil penalties (s 13H) for non-compliant APP 1 Privacy Policies without proving a serious interference. Where undisclosed or ungoverned AI data flows cause serious or repeated privacy interferences, civil penalties under s 13G reach the greater of A$50 million, three times the benefit obtained, or 30% of adjusted turnover, alongside direct private lawsuits under the Schedule 2 statutory privacy tort.