Regulatory guide
AI risks and limits in aged care
We build AI tools for aged care providers, and we think the risks deserve as much attention as the benefits. This page sets out what can go wrong, how to reduce it, and the jobs AI should not be given.
Reviewed 15 November 2026

In short
Aged care AI risks are real and manageable if you name them early.
Generative AI can produce confident, wrong text, and the privacy regulator warns that people tend to overestimate its accuracy. It can repeat bias in its training data, encourage staff to stop checking, leak sensitive information when staff use public tools, lock a provider into one vendor, and wear down skills that staff need when the system is unavailable.
For people living with dementia, AI raises questions of consent, dignity and confusion about who they are talking to. Some older people cannot use voice or screen-based tools easily at all. Each risk has practical controls: qualified review of outputs, approved tools only, clear disclosure, an easy route to a person, fallback processes and regular audits.
Some tasks should stay off-limits entirely, including clinical diagnosis, medication decisions, care decisions and anything that replaces human contact for people who rely on it.
Wrong answers delivered with confidence
Generative AI predicts likely text. It does not know whether the text is true. The OAIC notes that generative AI tools tend to produce outputs that appear credible regardless of their accuracy, and that this may lead users to overestimate their reliability (as at November 2026). In aged care that can mean a progress note draft that says a resident "ate well" when the care worker said the opposite, a summary that leaves out a fall, or a phone agent giving an outdated visiting time.
The OAIC also points out that incorrect information an AI produces about an identifiable person is still personal information and must be handled under the privacy principles. A wrong detail in a resident's record is a privacy issue and a care issue at once.
Controls: a qualified person reviews and approves anything AI drafts before it enters a record or reaches a family; the AI answers callers only from information you have approved; and you audit a sample of AI outputs each month and report the error rate.
Bias and over-reliance
AI systems can learn and amplify bias in the data behind them. The Guidance for AI Adoption gives the example of a biased hiring tool rejecting hundreds of qualified candidates before anyone notices. In aged care the risk shows up in recruitment screening, in speech recognition that copes poorly with strong accents or speech affected by stroke, and in any tool that ranks or prioritises people.
Over-reliance is quieter. In the first weeks staff check every AI draft carefully. By month three, many skim. That is when an error gets through. Controls include keeping the review step visible (the person who approves is named in the record), spot checks by a senior nurse, and refresher training that shows real examples of errors the tool has made.
Privacy leaks and vendor lock-in
The most common privacy failure is simple: a staff member copies a resident's details into a free public chatbot to help write a letter. The OAIC recommends, as best practice, that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. The fix is practical: give staff an approved tool with proper data protections, block or discourage the rest, and explain why in induction.
Lock-in builds slowly. Prompts, configurations, call scripts and data stored only in a vendor's format can make it costly to leave, and a vendor may change its model, price or terms. Ask before signing how you get your data and configuration back, in what format, and how fast. The Guidance for AI Adoption recommends planning for retiring an AI system, including how data is extracted or deleted, and keeping alternative pathways so critical functions continue if a system fails or is taken offline.

Apply this to your service
Talk it through with a Yes AI consultant in a free 30-minute call, including whether AI fits your service at all.
Deskilling and system outages
If AI drafts every handover for a year, new graduate nurses may never learn to write a clear one. If a policy assistant answers every question, staff may stop knowing where the procedures are kept. When the system goes down, or the internet does, the home still has to run.
Controls: keep core skills in training and competency checks, run regular shifts or drills without the tool, keep paper or offline fallbacks for anything time-critical, and test the fallback for phone answering at least twice a year so the on-call path works when the AI does not.
Consent and dignity for people living with dementia
Under the Aged Care Act 2024, older people must be treated with respect and are presumed to have the ability to make their own decisions, and they can choose registered supporters to help them make and communicate their own decisions (Department of Health, Disability and Ageing, as at November 2026). That presumption applies to people living with dementia too.
AI raises particular problems here. A person with cognitive impairment may not understand that a voice on the phone or a companion device is not a person, and may share private information or become distressed.
Monitoring technology such as cameras or sensors can intrude on privacy and dignity, and the Aged Care Quality and Safety Commission has a role in reducing the use of restrictive practices, so any monitoring needs careful clinical and ethical review.
Practical rules: always disclose automation in plain words, never design AI to pretend to be a person, involve the person and their supporters in decisions about AI that affects them, and do not use AI contact as a substitute for human company.
Accessibility for older people
Many older people live with hearing loss, low vision, tremor or limited experience with technology. A voice agent that talks fast, a chat window with small text or a menu with no option for a person can shut them out. For web-based tools, the Web Content Accessibility Guidelines (WCAG) 2.2 from the W3C are the reference standard for accessible design.
Test any AI tool with the people who will use it: residents, clients and families, including people from culturally and linguistically diverse backgrounds. Offer an obvious and fast way to reach a person, and keep phone, paper and face-to-face options for those who need them.
Tasks AI should not do in aged care
Some limits are about current technology. Others are about who should be responsible. These tasks should stay with qualified people, whatever a vendor claims:
- Clinical assessment, diagnosis or deciding whether a change in condition needs escalation
- Medication decisions, including dose changes, administration decisions and PRN judgements
- Falls prediction or risk scores that drive care decisions without clinical review
- Care planning, admission decisions or decisions about services a person receives
- Deciding whether an incident is reportable under SIRS, or submitting a report unchecked
- Quoting fees or giving financial advice to older people and families
- Replacing human contact for people who depend on it for company or reassurance
- Any decision that significantly affects a person without a human who can review and change it
How it works in practice
- 1
Name the risks per tool
For each AI tool, list what could go wrong using the risks on this page, and who would be affected.
- 2
Set controls
Assign a review step, a fallback and an owner for each risk, and record them on the risk register.
- 3
Tell people
Disclose automation to older people, families and staff, and give a quick route to a person.
- 4
Audit outputs
Check a sample of AI outputs every month and report the error rate to management.
- 5
Review limits yearly
Revisit the off-limits list each year, with clinical, privacy and consumer input.
Common questions
What are the main risks of AI in aged care?
Inaccurate outputs, bias, staff over-reliance, privacy leaks through public tools, vendor lock-in, loss of skills, outages, and specific consent and dignity issues for people living with dementia. Each can be reduced with review, approved tools, fallbacks and audits.
Can AI make mistakes in care records?
Yes. Generative AI can produce confident but wrong text. A qualified person should review every AI-drafted note before it is saved, and the provider should audit a sample regularly.
Is it ethical to use AI with people living with dementia?
It can be, with care. Always disclose that a tool is automated, never let it pose as a person, involve the person and their supporters, review any monitoring technology clinically and ethically, and never use AI as a substitute for human company.
What should AI never do in aged care?
Diagnose, make medication or care decisions, decide admissions, decide whether an incident is reportable, quote fees, or make significant decisions about a person without a human who can review and change them.
How do we avoid being locked in to an AI vendor?
Before signing, get written terms on how your data, scripts and configuration are returned, in what format and how fast, and keep fallback processes so services continue if you leave or the system fails.
Will AI replace aged care staff?
No. AI can take on drafting, sorting, reminders and routine calls. Care, clinical judgement and human contact need people. The realistic aim is to give staff back time from paperwork so they can spend it with residents and clients.
Sources and further reading
- Guidance on privacy and the use of commercially available AI products (Office of the Australian Information Commissioner)
- Guidance for AI adoption: implementation guidance (National Artificial Intelligence Centre)
- About the new rights-based Aged Care Act (Department of Health, Disability and Ageing)
- About aged care regulation and oversight (Department of Health, Disability and Ageing)
- Web Content Accessibility Guidelines (WCAG) 2.2 (W3C)
Related pages
AI Governance for Aged Care Boards
Directors of aged care providers already carry legal governance duties for quality and safety. AI is now part of the operations those duties cover, and boards need a simple, regular way to oversee it.
AI Privacy Guide for Aged Care
Aged care records hold some of the most sensitive information a person has. Before any AI tool touches it, you need to know where the information goes, who can see it and what the law expects.
Choosing an AI Vendor
Aged care providers are being offered AI tools by care software companies, start-ups, consultants and large technology firms. These are the questions to put to every one of them, ourselves included, and a simple way to score the answers.
AI Progress Notes for Aged Care
Progress notes are often written at the end of a long shift, from memory, by a tired care worker. AI drafting lets the worker speak or type what happened in their own words and get back a clear, objective draft to check, correct and approve.
Talk it through with us
A free 30-minute video call about where admin time goes in your service and which task is worth automating first. No obligation, and no resident information needed.