There is a lot of confident commentary about whether Australians are ready for AI. Very little of it cites a number. This does, because the number exists: the Office of the Australian Information Commissioner runs the Australian Community Attitudes to Privacy Survey every three years, and the 2026 round was published on 28 May. It is not an AI survey. It is a privacy survey, which is arguably better, because it asks about AI alongside every other institution competing for the same trust, and it asks the same questions it asked in 2023.
The trust number, in context
Asked which sectors are trustworthy with personal information, Australians put health service providers on 74 per cent and government agencies on 68 per cent. The bottom of the table is where AI sits:
- Social media companies: 3 per cent, down from 14 per cent in 2023.
- Data brokers: 4 per cent.
- AI companies: 4 per cent.
Being level with data brokers is the finding worth sitting with. A data broker's entire business is acquiring and reselling information about people who have no relationship with it. Australians rate AI companies exactly the same, and the report notes trust has fallen since 2023 across insurance, telecommunications, technology, retail and real estate as well. This is not a story about AI being singled out. It is a story about AI arriving at the bottom of an already-falling table.
What people want is specific, and it is hardening
The more useful half of the survey is not the mood, it is the list. Asked what conditions are essential before an organisation uses AI to make a decision that might affect them, 96 per cent named at least one, and only 4 per cent said no conditions are necessary. Both figures are identical to 2023, so the appetite for conditions has not grown. What has grown is how many conditions people want, and by how much:
Essential conditions before an organisation uses AI to make a decision affecting you
| Condition | 2026 | 2023 |
|---|---|---|
| A right to have a human review the decision | 81% | 73% |
| Personal information not retained by third-party AI providers beyond what is necessary | 80% | not measured |
| Being told that AI is being used | 79% | 71% |
| The organisation and any third-party AI provider must comply with strict privacy rules | 76% | 69% |
| Personal information not used to train external AI systems without consent | 76% | not measured |
| Being told if personal information will be shared with a third-party AI provider | 74% | not measured |
| Being able to challenge the decision made by the AI | 72% | 64% |
| A right to request information about how AI decisions are made | 70% | 68% |
| The accuracy of the AI is validated | 70% | 56% |
| The AI is tested for bias and discrimination | 68% | 57% |
| The way the AI makes decisions is clearly explained | 63% | 59% |
| No conditions are necessary | 4% | 4% |
ACAPS 2026, Figure 33. Base: all Australians aged 18+, n=1,504 in 2026 and n=1,642 in 2023. Three conditions were new in 2026 and have no 2023 comparison. Don't know (1%) and refused (under 0.5%) not displayed.
The two largest movements are the technical ones. Demand that the AI's accuracy be validated rose fourteen points, from 56 to 70 per cent. Demand that it be tested for bias and discrimination rose eleven, from 57 to 68. Those are not vibes about robots. They are the two things an assurance regime would actually require, and public support for both grew faster than support for anything else on the list.
Acceptance depends entirely on what the AI is doing
Treating "AI acceptance" as one number hides the real pattern. Asked how acceptable various uses of AI involving personal information are:
- Fraud detection: 64 per cent find it acceptable.
- Facial recognition or biometric identification: 37 per cent.
- Customer service chatbots: 37 per cent.
- Automated eligibility or risk decisions, such as loan approvals or benefit eligibility: 25 per cent.
The ranking is coherent. Australians accept AI most where it protects them from someone else, and least where it decides something about them. A chatbot scores the same as facial recognition, which should worry anyone who assumed the low-stakes uses were uncontroversial.
On training data the numbers are blunter still. Seventy-one per cent consider it unacceptable for an organisation to use personal information originally provided for a service to train AI systems after that service is finished, with 45 per cent calling it very unacceptable. And 93 per cent consider using personal information to train AI models and products unfair and unreasonable.
Set this against what Australia decided
Here is why these numbers are worth more than a headline. In September 2024 the federal government proposed ten mandatory guardrails for high-risk AI, and asked which of three mechanisms should make them law. Read that 2024 list next to the table above and the overlap is close to total: human oversight, testing, accuracy, transparency, contestability. Then, as we reported when we read both documents end to end, the National AI Plan of 2 December 2025 committed to none of it. The word "guardrails" does not appear once in its 37 pages. The settled position is that existing technology-neutral law plus sector regulators plus voluntary guidance will do the work, with targeted laws only where a need emerges.
The survey was in the field from 16 to 30 March 2026, about fifteen weeks after that plan. So it is a reasonable read of what the public wanted, measured after the government decided what it would do. On the evidence of its own regulator's survey, the gap is wide: 81 per cent want a right to human review, and no general Australian law gives them one.
The OAIC does not leave the implication unstated. In its own summary of what the findings mean, the report lists as a regulatory implication and opportunity: "Promote guardrails for AI/biometrics in particular purpose limitation, contestability (including human review where appropriate), strong transparency, and lifecycle controls on use, sharing and retention."
That is the privacy regulator using the exact word the government's own AI plan never uses, six months later, and attaching it to four of the conditions its survey found Australians want. It is a considered position from a body whose job is to hold one, not a political intervention, and it is on the public record.
The one thing that is actually arriving
Australians' single most-wanted transparency condition, being told that AI is being used, is on 79 per cent. It is also the one that is genuinely on its way, though by a narrower route than most people expect. Under amendments made by the Privacy and Other Legislation Amendment Act 2024, from 10 December 2026 entities covered by the Privacy Act must disclose in their privacy policy the kinds of personal information used, and the kinds of decisions made, by computer programs where those decisions significantly affect a person's rights or interests.
Note what that is and is not. It is a disclosure in a privacy policy, a document the same survey found Australians largely do not read. It is not a notice at the point of decision, not a right to human review, and not a right to challenge an outcome. It addresses the 79 per cent condition partially, and the 81 per cent condition not at all. The OAIC has said it intends to publish guidance on the obligation by around September 2026, which is the next thing worth watching on this beat.
What we take from it
Our reading, offered as a view and built on the figures above: the "Australians aren't ready for AI" framing is lazy and the survey does not support it. Sixty-four per cent accept AI for fraud detection. What Australians decline to do is extend that acceptance to decisions made about them without validation, testing, explanation or a human to appeal to, and their demand for exactly those four things has risen sharply in three years. That is not technophobia. It is a fairly precise specification, and it happens to match the guardrails that were drafted and then not adopted. Whether to legislate them is a genuine policy question with real costs on both sides. But the claim that there is no public appetite for them is now answerable with a number, and the number is not on that claim's side.