top of page

Lowy AI Poll Finds Australians Remain Wary

27 minutes ago
11 min read
Illustrative photograph. Photo: Lukas Blazek / Pexels.

The Lowy Institute’s November 15 publication reports that 65% of Australians believe artificial intelligence’s risks outweigh its benefits. The Lowy AI poll presents a challenge for governments and businesses promoting adoption: investment ambitions face a skeptical public.

The findings come from October interviews, not November fieldwork. They measure attitudes toward AI and comfort with overseas technology, rather than actual usage or demonstrated harm. The distinction matters because concern can coexist with practical use.

Australia’s National AI Plan promises economic opportunity, wider benefits, and safeguards. Lowy’s findings put pressure on that promise without establishing that it has failed. For policymakers, employers, and technology suppliers, the central tension is between expanding AI and earning public confidence in its consequences.

What the Lowy AI poll actually found

The October findings establish widespread caution, rather than a sudden November collapse in confidence.

Lowy published its analysis on November 15, 2026. Its AI attitudes survey reports 65% placing risks above benefits, with 33% taking the opposite position.

The Social Research Centre surveyed 3,028 Australian residents aged 18 and older from October 6 through October 18. Respondents came from Life in Australia, a probability-based panel whose recruitment uses sampling rather than an open invitation.

Most interviews took place online, with a small telephone component covering residents without internet access. This provides a stronger foundation than a website poll inviting anyone to participate.

However, a representative survey still estimates opinion. It does not capture every resident’s beliefs, establish individual motivations, or explain how respondents would judge specific applications.

That last limitation is particularly important. Someone can object to AI’s overall direction while welcoming a particular medical application or using a writing assistant.

The headline question asks respondents to balance a broad collection of benefits and risks. Different people can reach the same answer through different concerns.

One respondent might prioritize employment. Another might focus on privacy, inaccurate information, creative work, or dependence on foreign suppliers. The headline result cannot separate those explanations.

There is also a discrepancy in Lowy’s published historical comparison. The November analysis describes its result as unchanged from March and 13 percentage points above 2024.

However, the institute’s earlier AI risk assessment reports 64% choosing risks and 34% choosing benefits. That page describes a 12-point increase from 2024.

The reviewed pages do not explain the difference. Readers should therefore avoid treating the exact March comparison as settled without clarification from Lowy.

The broader interpretation remains straightforward: both publications describe a clear majority judging risks more heavily. Neither supports a claim that Australian opinion reversed during November.

That distinction changes the news judgment. The issue is persistence, rather than a newly discovered public backlash triggered by a particular announcement.

For businesses, persistent concern creates a different challenge from temporary controversy. Waiting for a news cycle to pass is unlikely to answer questions about accountability.

For government, the findings raise a communication and delivery problem. Explaining an investment agenda requires showing how ordinary residents benefit and how institutions address foreseeable harms.

Those are implications of the findings, rather than additional survey results. The poll itself does not prescribe a policy response.

Australia’s AI strategy faces a public confidence test

The pressure falls on institutions asking Australians to support AI expansion before its benefits feel concrete.

The government launched its National AI Plan on December 2, 2025. Its goals include infrastructure investment, broader adoption, worker support, and protection against harms.

That agenda presents an affirmative case for AI. It treats domestic capability and access to technology as contributors to competitiveness and better services.

The poll does not disprove that case. It does show why announcing capacity or attracting investment cannot, by itself, demonstrate public confidence.

An infrastructure project and a household benefit operate at different levels. A project can advance a national strategy without answering a resident’s immediate questions.

Will a service become easier to use? Will an employee receive training? Who handles a mistake? What happens to sensitive information?

These questions illustrate the gap between a policy objective and a personal judgment. They are not questions Lowy reports asking in this mini-poll.

For employers, the same gap appears when an AI rollout arrives as an efficiency announcement. Management’s expected benefit does not automatically explain what changes for workers.

Consider a hypothetical customer support team introducing AI-generated response drafts. Faster drafting might help, but employees still need clarity about review responsibilities and performance expectations.

A worker can reasonably welcome less repetitive writing while worrying about reduced staffing or increased workload. Those positions are compatible.

For enterprise buyers, this means adoption decisions involve more than model quality. They also involve responsibility, access controls, escalation procedures, and whether users understand the system’s limits.

For public agencies, the stakes extend to people who cannot simply choose another provider. Residents need a practical way to question an outcome.

That makes public confidence more demanding than favorable advertising. Institutions have to make their commitments visible in ordinary interactions, including when technology fails.

A useful distinction is between promising benefits and making them inspectable. The former describes an ambition; the latter gives affected people evidence.

For example, an organization could explain which tasks remain subject to human review. It could also describe how users correct errors and report problems.

These are potential responses to skepticism, not findings about what Australians would accept. The survey does not test whether any particular safeguard changes opinion.

The policy challenge is therefore specific. Government and industry must connect expansion with outcomes that people can assess, rather than assume economic ambition supplies consent.

That is the main tradeoff exposed by the news: institutions want room to deploy AI, while public confidence depends on credible limits and benefits.

American AI has an advantage, not a public mandate

A preference between foreign sources does not amount to broad comfort with either source.

Lowy reports that 36% feel comfortable with Australian businesses using AI sourced from the United States. The corresponding figure for technology sourced from China is 23%.

These figures concern comfort with business use. They are not the percentages using American or Chinese AI products, and they do not measure market share.

The difference gives American technology a relative advantage on this question. It does not establish majority approval for American AI.

That distinction matters for suppliers and policymakers discussing international alignment. Being the more acceptable option in a comparison is different from enjoying widespread confidence.

It also matters for competitive analysis. A country label is broader than any particular model, company, deployment, or contract.

The percentages cannot establish that respondents assessed OpenAI, Anthropic, DeepSeek, or another developer individually. Nor do they rank those companies’ products.

Lowy’s earlier discussion names OpenAI’s ChatGPT and Anthropic’s Claude as examples of AI entering public debate. Its October country comparison asks a different question.

A buyer comparing those services with Chinese alternatives still needs product-specific evidence. National origin does not replace an assessment of data handling and operational responsibility.

The practical differences can include who hosts the system, who can access information, and what contractual protections apply. The poll does not evaluate those arrangements.

A hypothetical business might compare a remotely hosted assistant with a model operated inside its own environment. The country comparison cannot determine that purchasing decision.

Similarly, domestic hosting and domestic control are separate questions. The physical location of a system does not explain every organization involved in providing it.

These distinctions help prevent a public opinion finding from becoming an unsupported security verdict. Respondents’ discomfort is evidence about attitudes, not proof of a particular vulnerability.

Lowy author Charles Lyons-Jones interprets the findings through a geopolitical lens. His analysis connects public sentiment with Australia’s strategic choices around AI.

That interpretation deserves consideration, but it remains an interpretation. The survey cannot demonstrate why every respondent prefers one source over another.

Privacy, familiarity, political attitudes, product experience, and general uncertainty are possible explanations. Establishing their relative importance would require additional questions or research.

The business implication is narrower and more useful: suppliers should expect questions about provenance, meaning where technology and its supporting services originate.

They should also expect those questions to continue beyond a country label. Buyers need an understandable account of control, responsibility, and recourse.

For American suppliers, the relative advantage should encourage careful disclosure rather than complacency. For Chinese suppliers, the lower comfort figure identifies a challenge without predicting inevitable rejection.

The central public confidence problem remains shared. Neither national category removes the need to explain how a particular service protects its users.

Job anxiety is evidence about expectations

Public concern about employment deserves attention, but it cannot establish how many jobs AI has eliminated.

A separate AI employment survey published by Pew Research Center on September 17 provides international context. It examines expectations across 37 countries.

Pew finds that concern about future job losses extends beyond Australia. Its research also distinguishes employment expectations from broader excitement or concern about AI.

That distinction prevents an easy mistake. People can anticipate fewer jobs while remaining interested in the technology’s other possibilities.

The Pew findings are separate from Lowy’s October mini-poll. Combining their results without identifying the different surveys would misrepresent what the new research asked.

They also address expectations, rather than observed employment outcomes. A forecast expressed by a respondent is not a measurement of layoffs.

Australia’s government has separately examined current labor market evidence. Its July 8 AI employment assessment found no evidence at that point of broad AI-driven upheaval.

The department also reported slower growth in occupations more exposed to potential automation. It described the evidence as suggestive, rather than definitive.

That assessment concerns evidence available for the July report. It should not be presented as a complete account of developments through November.

Its caution nevertheless provides an essential distinction for interpreting the poll. Worry about future displacement and evidence of current displacement answer different questions.

An occupation’s exposure to AI means technology can affect some of its tasks. Exposure alone does not establish that the occupation will disappear.

A system might reduce time spent drafting documents while increasing demand for review. Alternatively, an employer might use the same capability to reduce hiring.

Those are possible organizational responses, not outcomes established by either survey. Determining which occurs requires employment evidence and attention to other economic conditions.

For knowledge workers, the uncertainty concerns how employers distribute the gains. Time saved can support better work, additional output, lower staffing, or some combination.

The public confidence challenge follows from that uncertainty. A statement about productivity does not explain what happens to entry-level opportunities or employee responsibilities.

Employers would therefore benefit from separating their claims. They can describe a task improvement without implying that all workers will share equally in its gains.

The same restraint applies to critics. A convincing example of substitution does not automatically establish an economy-wide employment trend.

For readers following the story, the strongest evidence will connect specific deployments with measurable changes. General predictions alone cannot settle the question.

Useful evidence could include changes in hiring, job design, training, and the work employees actually perform. Each measure captures a different part of the transition.

The disagreement between optimism and anxiety is therefore not something the poll resolves. It identifies why credible employment reporting matters to the wider adoption debate.

Until outcomes become clearer, institutions face a practical obligation to describe uncertainty honestly. Confidence built on exaggerated reassurance is vulnerable to ordinary disappointments.

What the survey cannot settle

The Lowy AI poll provides a useful estimate of opinion, with limits that should remain attached to its headline.

Lowy reports an approximate 1.2% margin of error under a simple random sampling assumption. Its actual design is more complex.

The reported design effect is 1.21, indicating additional statistical variability compared with simple random sampling. The stated margin should not become a universal uncertainty guarantee.

Lowy also reports a 74.8% completion rate and a 2.2% cumulative response rate. These measure different stages of participation.

Completion concerns participation among those invited within the panel. The cumulative measure incorporates panel recruitment and subsequent attrition, meaning losses from the panel over time.

Neither figure should be read alone as a verdict on survey quality. Together, they help readers understand how the sample came into being.

Beyond sampling, a broad attitude question has interpretive limits. Respondents can attach different meanings to the same term, especially one covering many applications.

An answer about AI overall cannot automatically be transferred to every use case. Attitudes toward document assistance need not match attitudes toward automated eligibility decisions.

The hypothetical contrast matters because the consequences differ. A draft that someone reviews presents a different decision from an outcome someone must formally appeal.

The survey also cannot identify a causal explanation for skepticism. Public concern does not prove that a particular product launch, policy, or economic event changed opinion.

That limits the historical claim as well. Even a measured shift would require additional evidence before attributing it to one development.

The discrepancy between Lowy’s March figures reinforces the need for careful comparison. Exact trend language should wait for clarification, rather than hide conflicting published numbers.

There is another temptation to avoid: treating majority concern as a mandate for a specific restriction. The headline question does not ask respondents to choose a policy.

Someone who considers risks greater than benefits might support stricter safeguards, selective adoption, slower deployment, or continued experimentation. The result does not distinguish those positions.

Conversely, the minority favoring benefits should not automatically be treated as supporting unrestricted deployment. A positive overall judgment can still include demands for accountability.

For product teams, the implication is to investigate specific concerns rather than assign users to simple camps. Broad labels can conceal actionable differences.

For journalists, the responsibility is similar. Reporting concern accurately requires preserving what was asked, who answered, and when the interviews occurred.

For readers, this makes the finding useful without making it universal. It supplies evidence about the environment facing AI expansion, rather than a complete explanation.

The Lowy AI poll is strongest when used within those boundaries. Stretching it into a prediction about adoption, security, or employment would weaken the story.

Three signals will test whether confidence improves

The next evidence should connect attitudes with institutional behavior and observable outcomes.

Over the coming months, the first signal is comparable polling. A repeat survey would help establish whether broad caution persists, strengthens, or eases.

This is a watch point, not an announced survey schedule. The reviewed material does not establish a publication date for another Lowy mini-poll.

Consistency will matter more than an isolated headline. Comparable wording, fieldwork details, and transparent historical figures would make any change easier to interpret.

Clarifying the March discrepancy would also improve the baseline. Without that clarification, readers should resist building a precise trend around conflicting published percentages.

A sustained improvement across comparable surveys would weaken the argument that public confidence remains stuck. Continued caution would strengthen the case for a deeper problem.

The second signal is how institutions turn their AI commitments into conditions people can inspect. Infrastructure provides a concrete place to look.

The government’s data center expectations, announced March 23, address national interests, energy, water, employment, and local capability.

The department says projects meeting those expectations will receive priority in regulatory approval processes. It also says the expectations operate alongside existing laws.

The relevant question is how those commitments appear in individual projects. Broad statements become more informative when they connect to specific responsibilities and public reporting.

Evidence of enforceable safeguards and visible local benefits would strengthen the government’s case. Commitments that remain difficult to assess would leave the confidence gap unresolved.

This does not mean every project should be judged through a national poll. Local decisions need their own evidence about costs, benefits, and affected communities.

The third signal is employment evidence tied to actual deployment. Future assessments should help distinguish task changes from broader hiring or staffing effects.

A slowdown in an exposed occupation would still require explanation. Demand conditions, organizational changes, and other technologies can influence the same outcome.

Evidence linking adoption with better services and improved working conditions would support the benefits case. Credible displacement evidence would intensify pressure for worker protections.

Businesses can contribute by reporting outcomes clearly. They should distinguish a completed pilot from a sustained improvement, and a task saving from an employment conclusion.

For knowledge workers, that also makes documentation valuable. A searchable AI knowledge base can help preserve policies, research, and deployment decisions for later review.

It cannot resolve the national debate. It can make an organization’s own decisions easier to trace when questions arise about promised benefits or accepted risks.

Read the next Lowy AI poll alongside those practical signals. Ask whether its questions remain comparable, whether institutions explain their commitments, and whether employment claims match observed evidence.

When an organization announces another AI deployment, look beyond the expected efficiency gain. Check who reviews mistakes, who handles complaints, and how affected workers participate.

Then compare the announcement with outcomes reported later. Public confidence becomes easier to assess when promises leave a record that people can revisit.

The October findings establish caution. The next useful question is which institutions give Australians concrete reasons to revise that judgment.

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page