A presentation by ATSE President Cathy Foley to the Institute of Public Administration Australia on the role of expertise and evidence in informing the public sector.
This presentation was delivered during a webinar for the Institute of Public Administration Australia ACT, bringing Dr Cathy Foley AO PSM FTSE FAA's experience at the intersection of policy, public sector science and cutting edge research to an audience of professionals working with science in government.
I want to start with a question. What is evidence-based advice?
It’s a phrase we use all the time. Governments want evidence-based policy. Public servants provide evidence-based advice. Scientists want governments to listen to the evidence.
But after four years as Australia's Chief Scientist, I came to realise that the phrase can make the job sound much easier than it actually is. Because the real world rarely presents you with a neat folder marked THE EVIDENCE, containing the answer.
Usually the evidence is incomplete. Sometimes experts disagree. Sometimes the data were collected for a completely different purpose. Sometimes we have very good evidence about what happened in the past but much weaker evidence about what will happen next. Sometimes there is a very impressive study that turns out to be not particularly relevant to the decision you have to make. And increasingly, you can now ask an AI system a question and receive, within seconds, a beautifully written, authoritative-sounding answer, with references, arguments and conclusions. And, it is quite possible that some of the answer may be wrong.
So, in this information age we are living through, our problem isn’t a shortage of information. Quite the opposite. Data has never been more available, and perhaps never harder to interrogate. The capability we need is not simply finding information. It is deciding what deserves confidence. And that is what I want to talk about.
When I was Australia's Chief Scientist, my job was to give government evidence-based advice. Not my opinion.
I used to say: "I champion; I don't advocate." And when a minister or a department asked me a question, there were essentially three possible answers:
I sometimes joked that this meant I couldn't be wrong. Of course I could be wrong. But the serious point was that "we don't know" is a legitimate piece of advice.
So is: "The balance of evidence points this way, but confidence is limited."
Or: "There are two credible bodies of evidence leading to different conclusions, and here is why."
Some may feel under pressure, particularly when briefing senior decision-makers, to make uncertainty disappear. I think that is a mistake.
One of the most important professional skills in public policy is being able to distinguish:
That isn't weakness. That is intellectual honesty.
I learnt something about this much earlier in my scientific career. I'm a physicist. For much of my research life I worked in superconducting electronics. You make a device. You measure its properties. You develop a theory. You compare the theory and experiment.
This is an enormous simplification of decades of my life's work, but that was the world I understood. Then I got involved in work on nanoparticles in sunscreen. And suddenly I was in a very different scientific world. There were biological systems. Mouse models. Ethics approvals. Experimental design. Statistics. Sample sizes. Choices about endpoints.
And I remember thinking: my goodness, there are so many ways of doing perfectly honest research and still getting the answer wrong. That distinction has stayed with me. It has also taught me not to confuse research integrity with research quality.
Integrity is about behaviour.
Quality is different.
A scientist can have impeccable integrity and still conduct a study that turns out to be poor. Questioning the quality of research is therefore not an accusation of misconduct. It is actually part of the scientific process.
In my work on trust in science, I became increasingly concerned that we blur those terms. Scientific consensus is not the weight of popular opinion; it is the weight of evidence, built through testing, challenging and refining results. And quality involves the methodology, rigour, analysis and interpretation, not simply whether a paper has been published. That matters enormously for government, the parliament, the public service and indeed anyone seeking evidence-based advice.
Because the question isn't: "Has somebody published a paper supporting this?" You can nearly always find a paper.
The question is: "What does the body of evidence tell us?"
So, if somebody puts a report, a dataset, a scientific paper, or now an AI-generated briefing, in front of you, there are six questions I would encourage you to ask.
The first is: What exactly is the claim?
It sounds obvious. But it is remarkable how often people begin arguing about evidence without agreeing on the proposition they are trying to establish.
Second: Where did the evidence come from? Not simply: "Who sent me this report?" But, what is the original source? Is it administrative data? A controlled experiment? A survey? A model? A systematic review? Expert judgement? An industry report? A consultancy? A newspaper report quoting another newspaper report that quotes a paper? The further you get from the primary evidence, the easier it is for qualification and uncertainty to disappear.
Third: Is the method fit for the question? This is where experts are enormously useful. You do not have to become an epidemiologist, economist, engineer, psychologist, climatologist and quantum physicist. That would be a fairly exhausting APS capability framework. But you do need to be sufficiently confident to ask an expert: "Is this actually a good way of answering the question?"
Fourth: What does the wider body of evidence say? One paper is almost never the evidence. Science progresses because results are repeated, contested and refined. An outlier may eventually prove to be revolutionary. But until more evidence emerges, it remains an outlier.
Fifth: What don't we know? Unpacking that question means asking: What assumptions were made? Where are the uncertainties? What information is missing? And perhaps the most useful policy question of all: Would knowing more actually change the decision? Because perfect information is expensive and usually impossible.
And the sixth and final question: Is this evidence fit for the decision we actually have to make? You can have excellent science answering the wrong policy question. Even scientists don't always appreciate that.
And there is another problem. All of this sounds wonderful until a minister asks for an answer by Friday.
This is where I think Australia developed something extremely valuable during COVID. Alan Finkel was Chief Scientist when the pandemic began. The government faced questions that simply could not wait for a conventional research process. The evidence was developing almost daily. So, Alan convened what became the Rapid Research Information Forum—the RRIF.
It brought together the Learned Academies, ACOLA, government science advisers, CSIRO, universities and others. The Australian Academy of Science provided operational leadership. During 2020, it produced 13 expert reports based on the best available evidence.
And the idea was beautifully simple: A minister asks a tightly framed question. The relevant experts are convened. You rapidly review the available evidence. You expose the draft to expert scrutiny. And instead of producing a 200-page report six months later, you produce something short, often around 3,000 words, that says:
That was a very important innovation.
Because normally there is a mismatch between the speed at which science generates robust knowledge and the speed at which government sometimes has to make decisions. The RRIF bridged the two.
It was initially built for the pandemic, but during my term as Chief Scientist, we deliberately carried that capability forward through Rapid Research Information Reports and used the approach for questions well beyond COVID. The Academy of Science described the successor process as embedding the RRIF capability into routine governance and extending it to a wider range of science-policy issues.
And sometimes the subjects changed remarkably quickly.
There were questions about COVID vaccines, their safety and the reasons for vaccine hesitancy. There was the AdBlue shortage; this is the additive for diesel cars and trucks that reduces polluting gas emissions, and something that sounded highly technical could potentially affect Australia's road freight system and supply chains. And then, at the end of 2022, something called ChatGPT appeared.
Suddenly everyone wanted to know what generative AI meant. In February 2023, the Minister for Industry and Science asked the National Science and Technology Council for advice. By March, ACOLA delivered a rapid report. ACOLA brought together expertise from the Academy of Humanities, the Academy of Technological Sciences and Engineering and the Academy of Science, with Genevieve Bell, Jean Burgess, Julian Thomas and Shazia Sadiq leading the work.
Think about that timing. ChatGPT had only been publicly released a few months earlier. There wasn't a ten-year evidence base.
Government couldn't say: "Come back when the literature has settled." Policy was already having to respond.
So the task wasn't to manufacture certainty. It was to tell government:
That, to me, is what good evidence advice looks like when knowledge is moving very quickly. Sometimes you need the deeper dive. Of course, rapid advice isn't appropriate for every question. Some issues require much deeper work.
During my term, for example, we undertook a substantial examination of airborne infectious disease transmission and indoor air quality.
That work involved a systematic review by the University of Wollongong's Sustainable Buildings Research Centre, followed by advice from the National Science and Technology Council. And the answer wasn't a slogan. No single magic intervention suited every building. The evidence supported approaches including ventilation, air filtration and ultraviolet-C technologies, but their application depended on building type, use, energy implications and other factors.
That is exactly what happens with real policy problems. The answer is often: "It depends." But good advice then tells you what it depends on.
And this brings me to something I want this audience to know. If there’s one thing I want you to take away from today, it is that
you have access to an extraordinary national resource in Australia's Learned Academies. I say that now wearing a different hat, as President of the Australian Academy of Technological Sciences and Engineering.
But I used the Academies constantly when I was Chief Scientist. So did Alan Finkel. Together, Australia's five Learned Academies cover science, technological sciences and engineering, health and medicine, social sciences and the humanities. Through ACOLA, we can bring those disciplines together around a problem.
And that breadth matters. Because almost no serious public-policy question is purely technical.
Take AI. It is obviously about computer science. But it is also about psychology. Education. Law. Economics. Workforce. Ethics. Humanities.
Security. Social cohesion. Energy. Even philosophy. If you frame AI as merely a technology problem, you will probably give the wrong advice.
The real strength of the Learned Academies is not simply that they contain very distinguished people. It is that they provide a mechanism for convening expertise across disciplines, rapidly and independently. So my message to you is this: Use us.
You don't always need to commission a six-month consultancy project to discover what Australia's research community already knows. Sometimes what you need is a very well-framed question and a way of convening the right people around it.
This is particularly important now because the machinery of science advice to government is changing. The National Science and Technology Council that I worked through is no longer the mechanism it was. The government is establishing a National Resilience and Science Council, with a stated purpose of better coordinating public-sector R&D and innovation investment with national priorities, economic objectives and resilience.
Government statements indicate this new council will report to the Prime Minister and the Minister for Science, with the Chief Scientist playing an important role. But under this new approach, I think we should also preserve something the previous system demonstrated very successfully: the ability of any minister, facing a difficult question, to get rapid, independent, multidisciplinary synthesis of the evidence.
Wherever that function ultimately sits institutionally, Australia's Learned Academies are ready to help provide it. We already have the networks. We already have the expertise. And we have demonstrated that it can work.
But now AI is changing the information environment. Which brings me back to AI. AI is not simply another subject on which government needs evidence. It is changing the information environment in which evidence itself is produced, found, summarised and communicated.
There are enormous benefits. I use AI.
Used well, it can make us much more productive. But it introduces a very important problem. Fluency is not evidence.
A language model is extraordinarily good at producing an answer that sounds like someone knows what they are talking about. And humans are very responsive to confidence.
One 2025 study described what its authors called a "chat-chamber" effect. In their experiment, people sometimes accepted incorrect information from a chatbot without checking it, especially when the response was convenient and consistent with what they already believed. The authors themselves caution that this was one study, using an older model and a particular set of questions, so we shouldn't turn it into a universal law.
But the underlying warning matters: a single fluent answer can discourage the comparison and verification process that researchers are traditionally required to do. That has implications for public-sector work.
We also need to distinguish misinformation and disinformation. Misinformation is false or misleading information that may be shared without an intention to deceive. Disinformation is deliberately created or disseminated to mislead. Generative AI can contribute to both.
It can generate misinformation simply by hallucinating, producing a false claim, a non-existent citation or an incorrect synthesis. But it can also make deliberate disinformation cheaper and easier to produce at scale: text, images, audio, video, fake personas and tailored messages.
And AI-generated persuasion is not hypothetical.
A large experimental study published in Nature Communications in 2025 found that LLM-generated messages could shift people's attitudes on a range of policy questions and were broadly comparable in persuasive effectiveness to messages written by people.
The important public-policy point is not which position people were persuaded towards; it is that persuasive material can now be produced rapidly, cheaply and at a very large scale.
And we should not assume AI will solve the problem it helps create. Research published in Nature Communications has also found significant limitations in the ability of language models themselves to detect AI-generated mis- and disinformation reliably.
So, we have an interesting loop. AI can create information. AI can summarise information. AI can persuade us with information. And then we may ask AI whether the information was true.
That should make us think very carefully about verification.
One of the final projects I initiated through the National Science and Technology Council asked: "What makes people susceptible to misinformation and disinformation—and what makes them resilient?"
The work has now been completed and published by our current Chief Scientist, Tony Haymet. And one of its most important findings is that this isn't simply a problem of stupid people believing bad facts. Susceptibility is influenced by psychology. Trust. How we process information. Mental and physical wellbeing. Social connection. Our communities.
And resilience can be strengthened by information and media literacy, education and trusted sources. That means you cannot solve misinformation simply by putting more facts on a website. And you certainly can't wait until misinformation has spread widely and then say:
"Here is the correct PDF."
Trust is part of the infrastructure. The dramatic AI story isn't necessarily the most useful one.
There is another interesting lesson here about AI and evidence. We hear a lot about existential AI risk, the idea that a sufficiently capable AI might become uncontrollable, seek power and cause catastrophic harm. I recently looked at a synthesis of the research literature on exactly that question. And what I liked about the answer was that it didn't pretend the literature was settled.
Serious scholarly work argues that advanced AI could create catastrophic risks through misalignment, autonomous behaviour, or power-seeking. There is also scholarship challenging some of the classic arguments about a rogue superintelligence. And there is another body of work saying: don't become so preoccupied with the dramatic scenario that you overlook the nearer-term and cumulative risks, bias, hallucination, misuse, poor governance and gradual delegation of decisions to systems we don't properly understand.
That is exactly the sort of issue where a public servant should resist the demand for a binary answer. If the question is: "Will AI go rogue and destroy humanity?", the evidence does not justify simply saying yes.
Nor does it justify saying there is nothing to worry about.
A much better piece of advice would say: "There are plausible high-consequence scenarios that deserve serious risk management; the probability and mechanisms are contested; and there are other demonstrable risks already requiring attention." That is a far more useful answer. AI should be your assistant, not your authority
So what does this mean practically?
I would use AI enthusiastically. But I would give it a particular role. AI can be my research assistant. It should not become my authority.
Ask it:
Those are very powerful uses. But when a fact matters to a decision: go to the source.
If the statistic matters, check it. If the quotation matters, find it. If the paper matters, read at least the relevant part of it. If the AI says there is a consensus, find out whether there actually is one.
And I think traceability is going to become one of the most important skills in evidence-based public administration.
Can I trace this claim back to its source? Can somebody else reproduce my evidence trail? Do I know which parts of my briefing come from evidence, which are analyses and which are judgement?
And finally, perhaps the hardest situation of all. What happens when the evidence isn't what people wanted to hear? A program isn't working as expected. A favoured intervention is not supported as strongly as people believed. New evidence has emerged. Or your analysis challenges an established policy direction or ministerial priority.
I don't think the answer is to weaponise science. "The science says you are wrong" is rarely a very useful briefing strategy. But it's not our job to make inconvenient evidence disappear. A public servant's job is to make the decision-maker better informed.
That might mean saying: "The current evidence provides less support for this approach than we previously understood."
Or: "Government can proceed, but the evidence suggests these three risks need to be managed."
Or: "There remains substantial uncertainty; these are the consequences if our underlying assumption turns out to be wrong."
You preserve the right of the elected government to decide. But you preserve the integrity of the advice as well.
During my time as Chief Scientist, I found something else interesting. When reports arrived with long lists of recommendations telling government what it must do, they were often not very useful. Policy questions are multidimensional. A recommendation can collapse all those dimensions into a single answer.
I had much more success providing evidence and conclusions.
That created a very different relationship. It was advice rather than advocacy.
I have reflected on this since: in my Chief Scientist role, advice based on evidence and conclusions often created a pathway into government, whereas prescriptive recommendations could shut that conversation down.
Let me finish with this. We are entering a world in which information will become cheaper and more plentiful. Producing words is becoming almost free. Producing images and video is becoming extraordinarily easy. Producing a convincing argument can take seconds.
That means judgement becomes more valuable, not less. Your value as a public servant will not be that you can find information faster than an AI system. You probably can't. Your value will be your ability to ask: Where did this come from? Does it deserve confidence? What is missing? Who should I ask? How uncertain are we? What does this actually mean for the decision? And perhaps one more: Who can help me?
Because you don't have to know everything yourself.
Australia has extraordinary scientific, technical, health, social science and humanities expertise. Our Learned Academies exist in part to make that expertise available to the nation. So use us.
And the thought I will leave you with is this: Credible evidence is not evidence without uncertainty.
And trusted advice is not advice that pretends uncertainty has disappeared. Trusted advice is rigorous enough to distinguish what we know from what we don't know. Honest enough to say so. Curious enough to keep asking questions. And useful enough to help somebody make a better decision.
In the age of AI, I think that combination of evidence, judgement and trust will become more important than ever.
The Australian Council of Learned Academies welcomes the release of the Australian Government Discussion Paper on Safe and responsible AI in Australia today by the Minister of Science, the Hon Ed Husic MP.
Diversity and Inclusion
Reconciliation Action Plan Equity, diversity and inclusion policy Diversity & Inclusion ToolkitConnect
Australian Academy of Technological Sciences & Engineering
Level 2, 28 National Circuit Forrest ACT 2603
+61 2 6185 3240 | info@atse.org.au
ABN 58 008 520 394 — ATSE is registered with the Australian Charities and Not-for-profits Commission,
and is a Deductible Gift Recipient (DGR).
