BENFIELD, Mr Sean, Senior Executive Director, Financial Statements Audit Services Group, Australian National Audit Office CRASWELL, Ms Lesa, Group Executive Director, Systems Assurance and Data Analytics Group, Australian National Audit Office FECHNER, Mr Chris, Chief Executive Officer, Digital Transformation Agency FLEVARAS, Mrs Fay, Chief Digital Information Officer, Department of Health, Disability and Ageing [by video link] FOX, Ms Amy, First Assistant Secretary, Financial Analysis, Reporting and Management Division, Department of Finance HICKS, Mr David, First Assistant Secretary and Chief Financial Officer, Department of Health, Disability and Ageing [by video link] HUBBARD, Mr Paul, Assistant Secretary, Acceleration and Enablement, Artificial Intelligence Delivery and Enablement Division, Department of Finance JAGO, Ms Carla, Acting Deputy Auditor-General, Australian National Audit Office LAL, Ms Monita, First Assistant Secretary, Artificial Intelligence Delivery and Enablement Division, Department of Finance LIU, Ms Faye, Assistant Commissioner, Craft and Learning Branch, Australian Public Service Commission McLIESH, Dr Caralee, PSM, Auditor-General, Australian National Audit Office POOLE, Ms Lucy, Deputy Chief Executive Officer, Digital Strategy, Planning and Performance Division, Digital Transformation Agency SAMARANAYAKE, Mr Lasantha, Assistant Secretary, Accounting and Frameworks Branch, Financial Analysis, Reporting and Management Division, Department of Finance SAWADE, Mr Mark, Chief Information Officer and Chief Security Officer, Australian Taxation Office [by video link] TEPPER, Ms Lisa, First Assistant Secretary, Information Technology Division, Department of Health, Disability and Ageing [by video link] WATSON, Mr Andrew, Deputy Commissioner, Smarter Data, Australian Taxation Office [by video link] CHAIR: I resume the committee's hearing into the Commonwealth financial statements 2023-24 and 2024-25. Welcome. Do you have any comments to make on the capacity in which you appear? Mr Watson : I'm the ATO chief data officer and AI accountable official. CHAIR: As this hearing is a formal proceeding of the parliament, the giving of false or misleading evidence is a serious matter and may be regarded as a contempt of the parliament. The evidence given today will be recorded by Hansard and attracts parliamentary privilege. I remind all that, as chair, I will ensure proceedings are conducted in an orderly, respectful and courteous way. Witnesses are reminded that any claim that it would be contrary to the public interest to answer a question must be made by a minister and accompanied by a statement setting out the basis for the claim. Do any of the agencies have any opening statements? No? Alright. As a starting point, for the purposes of the committee's understanding, maybe each agency could just outline the work being done right now to utilise AI in whichever form—around advice, around processes, around decision-making. How is it being utilised right now? And how is it evolving? Let's start there. Mr Fechner : What I might do is start with a general position around its linkage to financial statements. I think it would be fair to say that AI is becoming pervasive across all aspects of technology and pretty much in government, so I might keep it a bit more narrow to financial statements. Most of the way that the Commonwealth delivers financial statements is through end products delivered through our enterprise resource planning, or ERP, systems. This covers areas such as SAP, TechnologyOne, Orion and Workday. They fit into that general category of ERP. In general terms, AI— CHAIR: Sorry, what does ERP stand for? Mr Fechner : Enterprise resource planners. It's a generalised term for the people that do the financial and HR systems in large part across the Commonwealth. They're a specifically broad category of use all across the economy. Within the Commonwealth, the main ERP systems that we use are SAP and TechnologyOne, and for HR there's another product called Orion that comes in. There are smaller ones that are utilised, but I think all of those products are extensively investing in how they're actually incorporating artificial intelligence and AI agents within their environments to support better processing and more effective outcomes through the use of ERPs, and the end product of those is often financial statements. I want to talk a little bit more about the history of ERPs in the Commonwealth. For a very long period in the history of the Commonwealth, between 2019 and 2024, there was a moratorium on investing in new ERPs in the Commonwealth. That was on the basis that the government was considering implementing a single ERP model called GovERP, and it was meant to address many of the issues about the complexity of producing annual financial reports. That program, through a series of changes and a series of disagreements about the way that the system would be implemented, never actually came to fruition, so it was never implemented. So, in 2024, the DTA undertook an assessment to see what we could recover from that, and the result of that work was a new government decision on a future direction for ERPs, which was to allow agencies to make their own choices and implement in a consistent way. Also out of that, there was a recommendation that the Digital Transformation Agency support the development of an enterprise resource planning panel to allow a standardised approach to go to the market to receive services in that area. We've been progressively increasing our work in defining standardised ERP patterns and working with the vendor community to bring forward those solutions. Importantly, we're working very closely with their introduction of AI within their product spaces. I also want to briefly go over some of the issues associated with the IT system controls that the ANAO have brought forward, and I want to use a specific example that demonstrates the overarching complexity and why this is such an ongoing issue for many agencies. I'm just going to use a real-world example of an agency that receives services from an internal government service provider. I'll tell you the layers that are in there. The first layer is generally getting access to the systems through the agency's own local environment and then the system control to allow access to the system. Then a second layer can be that the HR system is an independent system. The third layer can be that the finance and reporting system is an additional layer. Expense and travel management is often taken out, and in this case is taken out, as a fourth layer. Procurement goes out in a fifth layer. Financial statements consolidation—so the work that we do to aggregate all the financial statements into portfolio agencies—is done in a separate system at a sixth layer. Reporting is often done at a seventh layer. In the case where a service is provided by another entity to an entity, the eighth layer is the management assurance levels that are required to give the agency head the authority to certify the annual report and the financial statements within it. So you can see, and it's been clearly demonstrated by ANAO on a lot of occasions, that there is a high probability that one of those layers is going to have breakdowns and that things like people retaining access or having authorities that they shouldn't have—because of the cascading impact of stepping out of all of those environments—are quite likely. Within ERP systems there is a capability that's heavily developed in most of them called governance risk and compliance. It's generally a module that says: 'Have I got the right internal controls? Have I got the right authorities that have been set up for expenditure or for approvals?' They work very effectively, but because of the way most government agencies work both on-system and off-system for some components—or across systems—it's very hard to ensure that those services work as intended across the entire ecosystem, and I think that shows up in the audits. This is not a problem that's resolved very simply. What I would say is that, generally, ERPs—because of the moratorium and other things—are an underinvested digital class. We need to figure out how we work over the next three to six years on what we need to do next. Again, one thing I will tell you is that many of the systems government uses to create its financial statements fall out of support in 2027. A migration of those activities needs to be undertaken by our drop-dead date of 2030. There'll be a large amount of concurrent work going on on ERP statements, which again I think will create pressure on the controls that we need to do, because they could be doing upgrades at the same time as they're trying to operate their systems. It's a very complex environment for all government agencies right now. CHAIR: Please forgive me, Mr Fechner. I wasn't smiling at your thorough overview. I was giggling at the ridiculous but cute contribution of my daughter, who just had a bottle and has now also done a few other things in here. Thankfully, you're on a Zoom call and can't get a full appreciation of what's happening right now. Thank you for that; that all makes perfect sense and is very useful. Let's go into some of the specific examples of how it's being integrated and then also how governance is potentially a missing step. Let's go back to the department of health and Mrs Flevaras. How are you integrating AI to streamline processes around either public interaction with programs or user interaction with programs? How's AI being included in that process? Mrs Flevaras : Thank you for the opportunity to give you a bit of a background on what we're doing. The department recognises artificial intelligence and other emerging technologies are changing the way we interact, including with our fellow government agencies and the Australian public. The department has worked with colleagues across the DTA, the Digital Transformation Agency, and the Department of Industry, Science and Resources on how we set out our policies and frameworks and road maps for AI usage in our work and also in the broader health, disability and ageing sectors. We're aligning our approach with whole-of-government policy direction, including DTA's policy for responsible use of AI in government and the APS AI plan. In practical terms what this means is we've established a dedicated AI governance subcommittee that reports up into our digital committee. We are implementing an AI assurance framework, and we require everyone in the department to mandatorily document their use cases that they're leveraging AI for so that we can perform impact assessments to ensure risks are identified early and managed through the life cycle. This is a deliberately conservative approach to minimise risk. We are conscious of the risks, but we are also very conscious of the opportunities AI poses. There is the potential for great benefits, but this cannot come at the expense of the trust of the Australian public. Very specifically, at this point in time we use the Microsoft 365 Copilot for generative AI purposes, which means staff within the department that have a licensed version have access to that. We do have other tools that are being considered but are not implemented or used in production. Vertex AI is approved for use in our enterprise data and analytics platform for analytics. However, it is currently only being tested, so you can see we're taking a very cautious approach. We have six approved use cases within the department, which are official information only, for the use of the Microsoft 365 Copilot, and that's to give guidance to our staff on what use cases they can freely use the Copilot for. These include documentation, summarisation and analysis; content creation and editing; meeting and task management; data insights and visualisation; information extraction and background research; and image generation. All of these use cases are there to assist our staff to do their day-to-day work, and, as such, content is the responsibility of the staff member to validate. We talk about having the human in the loop. That would probably be the extent to which we're using it at the moment. Hopefully, that answers your question. CHAIR: What's the application more specifically around aged care, as overseen by Health? How is AI being utilised in aged-care settings? Mrs Flevaras : Aside from those six approved use cases that I've just outlined, we would not be using AI in any other facet in the aged-care space. CHAIR: Wasn't there public reporting about it being used? Mrs Flevaras : Not that I'm aware of. I can tell you that we don't use it in any other fashion. There is a distinction between AI, which is an automated agentic agent, versus some business rules or policy rules that we might have automated, but that's not AI. CHAIR: What are those business rules and policy rules that you've automated? Mrs Flevaras : I can take it on notice for you. Particularly in the AI space, there was some discussion around the single assessment tool, but that is just policy that has been automated. An algorithm was calculated outside of the system, so there are a few converging topic areas there. The policy area have developed their algorithm on what a single assessment looks like, and then we have taken those rules and digitised them into a single assessment tool. CHAIR: Where is that being done? Mrs Flevaras : Do you mean the specific technology? CHAIR: Yes. Mrs Flevaras : I'll take that one on notice and come back to you, if that's okay. We have a number of different technologies in the ecosystem. I wouldn't want to misinform you. CHAIR: You don't have any rough idea of where it's being utilised? Mrs Flevaras : We have three or four different technologies in the ecosystem—coarse-grained. It would be a combination of Microsoft, Oracle, Salesforce—there are a few of them in there, and different components sit in different parts of the stack. Essentially, it is using the My Aged Care gateway and assessor platform that we've created, and then there's the government provider management system. So there are a number of technologies under those platforms that enable the transfer of data across the journey. CHAIR: So what's automated as part of the My Aged Care gateway? The purpose of these questions is not to be particularly secretive or anything like that; it's more about what's being used and how we are identifying the issues to help us manage this. Mrs Flevaras : In summary, I'd start by saying there is no AI at all in aged care. CHAIR: With the automated algorithms, what's being used? Where you take away a manual assessment and you automate it, where's that happening? Mrs Flevaras : That happens in the application from the elderly person coming in and putting their application in My Aged Care. That data is then provided to the assessors. They access that data through an assessor portal. That is where the single assessment is then performed. Once they get a result from there, the data is then passed on to Services Australia for them to do their serviceability, eligibility and entitlement. Like I said, it pretty much traverses the whole ecosystem and all technologies. CHAIR: That makes sense. You didn't misinform me. It seems like a pretty straightforward response. For the purposes of this hearing and the purpose of the committee, we don't need to have any nervousness. You have this automated process. You have older Australians who are accessing this gateway, and there are good reasons for automation because it can make sure people get access to funding. There's obviously a motivation—it makes it easier and cheaper to run. But what about the governance around that? How is the department ensuring that the decisions being made by the automated process are the right decisions and that they are not leaving people with plans that are insufficient or are overly generous? How is that being governed and assessed right now by the department of health? That is really the point of this hearing. Mrs Flevaras : We haven't taken away the manual assessment; that's still there. The assessors undertake the assessment and record their outcomes. The algorithm just applies what is in the legislation and provides that to the assessors. CHAIR: Yes, but, for the assessors, how do you make sure that the algorithm is getting it right? Do you track that? Do you audit what the algorithm is spitting out? Mrs Flevaras : I will need to take that specific question on notice and bring back a response. It's a policy question, not a technical one, in the sense that we've done a full software development life cycle. We track from policy requirements into business requirements, and then we create the software. It all gets tested, so nothing changes. It's not like AI, which— CHAIR: I'm not asking for your opinion on whether or not it's working. I'm asking: how do you make sure it is being implemented accurately in accordance with the legislation? How do you ensure that the assessments or that the information that is being put into the AI or put into the program, or the software or the automation, is being governed in accordance with the services that need to be delivered? It's not a policy question around what's your opinion; it's how it is being implemented. How is it being governed? Mrs Flevaras : The implementation is through the software development lifecycle. CHAIR: How are you assessing that? Mrs Flevaras : We have traceability from requirements all the way through to system design and system implementation. We have test strategies and test plans that test all the different scenarios, and that all gets completed and tested and validated by our business before it goes into production. That has all been done— CHAIR: And then once it's in production? Mrs Flevaras : It doesn't change. That's where it's different to AI. In the sense that the AI tool might evolve— CHAIR: With the inputs and the application of a program, I understand that whatever gets put in one end gets spat out in accordance with the program software, but how are you ensuring that that whole interface with older Australians or senior Australians is working? Mrs Flevaras : I'm not sure how to answer your question. From a technology perspective, we have built the platform and tested it against the rules that were provided, and it is a consistent output. Your question regarding how that meets elderly requirements—that's not the system. It's the technical system. CHAIR: But the department is the governing body to ensure that the software used actually meets the needs of people it's serving. Mrs Flevaras : Correct. CHAIR: You're not the department of software; you're the department of health care and ageing. How is the department assessing that interface and that usage of the software that isn't changing? I grant that it is what it is. How are you assessing and what's the governance around that automation that's ensuring that it is doing what it's designed to do, not from a technical point of view but from the point of view of the user interface? Do people understand how to use the program? Do people understand where to go and what information to use? What's the feedback being received by the department—all of that stuff? Mrs Flevaras : When we went live with the implementation of the new Aged Care Act, there was quite a lot of work that was done to uplift training procedures and manage the full change into the sector. One of the two specific things that I can speak to is the Digital Transformation Tech Talk webinars. That was a series of information webinars where we would engage every six weeks with the sector to let them know all of the changes that were coming through. A subcommunity of that was also what we called our sector partner community. They were volunteers in the sector who signed up to be a sector partner who co-designed a lot of the implementation with us, and so we got lots of feedback along the way as we were implementing. Now that we've gone live in at least those two channels plus more from our policy areas, we continue to connect with the sector, and they provide us feedback on any problems or concerns that they might have, in that sense. There was a lot of material around training and procedure changes, and so I feel there are ongoing channels for the sector to provide us information and that we can respond to. CHAIR: From the Auditor-General's point of view, the management of the automation or the assessment of the automation or the policies around assessment of the automation—I think that that answer around making sure that there was a lot of information going out before it went live and a lot of training and a lot of communication with the sector seems like a really reasonable step to take. As this becomes more and more prevalent across government, what about the building in of quality assurance and building in of governance around it? Does the Auditor-General have any views or any observations around any of that or the insufficiencies or gaps that the committee could potentially focus on? Dr McLiesh : Across our audit work, we are seeing entities increasingly adopt AI and also automation more broadly, as we have just heard. There are a number of potential benefits of this for efficiency, productivity and quality of services, but it's also clear that there are a range of risks to be managed. So the key observation that comes from our work in this area is the need for really strong governance, risk management and assurance over compliance that is commensurate with the risks of adopting emerging technologies and AI. There are risks both ways—risks of adopting these technologies and also risks of not adopting those technologies—particularly when they are increasingly being used by stakeholders, clients, providers and contractors of entities. So it's really about making sure that governance is fit for purpose to manage risk. Because there is such rapid development in this area of the technologies and of the adoption, what we're seeing is that sometimes the governance and assurance arrangements are not keeping up with the use of AI. There are many good practices in place. We've just heard a few, but I think the key question is: is the governance keeping pace to manage the risks effectively? Mr Fechner : Mr Burns, if I could add a little bit about the automated decision-making, by the end of this calendar year, the Attorney-General's Department will put out a framework on automated decision-making, and that's in response to the robodebt royal commission inquiry. There is a framework coming out, and there's been some interim guidance come out on those particular things, of which the phased aged-care transformation program has actually complied with those components. When you talk more generally about the use of artificial intelligence, the DTA has got the policy authority for responsible use of AI, and we've put out standards for transparency and standards for accountable officials that are going through. We've also recently published technical standards that actually support entities managing AI through their life cycle. They mentioned the software development life cycle, or SDLC. AI has a different context in its operation, in that an AI can sometimes perform well in testing prior to go live but then display behaviours that are not anticipated while it's in operation, because it can learn from its experiences and apply those changes to it. We're currently doing quite a lot of work in conjunction with the AI plan to try and make sure that we've got all the controls in place, but, as I think the Auditor-General called out, we also need to make sure that we can adopt these capabilities because they do offer some very, very positive benefits as well. The aged-care transformation program has been very, very strong in its positioning around ensuring that the quality of its deliverables are in place before they go live. That was probably seen very much in the case of a deferment of the go live from July to November because they needed to get more confidence, that they were comfortable that everything was operating as it should. One of the DTA's roles is to sit on governance bodies right across significant delivery programs to support them with the best mechanisms and to ensure that their technology delivery works effectively, but it is also a control point about where there needs to be break points that are put in place to make sure that unintended outcomes don't occur from the delivery of new software solutions. CHAIR: I'll come back to you. Sal, I'll hand over to you. Ms SITOU: I want to approach it from a different angle. I know we've been talking a lot about risk and governance, and this is a question to all the agencies here. What have you got in place to make the most of what AI can bring? What are the mechanisms, the personnel, the resourcing or the funding that you have set up to enable staff to experiment with the use of AI in a safe and well-governed manner? Mr Fechner : If I can just jump in, I think some of our line agencies have some really advanced capabilities. I might throw to Mark and Andrew. Mr Watson : We've done AI through machine learning for nearly a decade in the ATO and currently have about 50 machine learning models that sit behind and help us, as a regulator, identify risk in the market. They also help us prompt and nudge taxpayers through predictive behaviour modelling. For example, a small business might regularly be late lodging their activity statement. We have some models that might send them a text message a few days before it's due, to nudge them along. We've had some governance around that. Since 2024, we have progressively expanded. Initially, as part of the DTA trial, we had Microsoft Copilot available to a small number of staff. We've now expanded out a licence version to nearly 3,000 and Microsoft Copilot Chat to all staff—roughly 20,000. The reason we have gone with the Microsoft product is that we have that behind what's called enterprise data protection. It all operates within the ATO data tenancy. We don't have data leaving the ATO being used in training for any models. We have a degree of risk protection there. In March we released, similar to what Fay described before, an artificial intelligence senior executive instruction. It's a lawful direction from the commissioner to staff that outlines responsible use aligned with the DTA policy, and we have a use case register which gives guidance around the sorts of things people can responsibly use the AI for. We are seeing people use it in lots of ways. We've got the 'thousand flowers blooming'; now, with our executive, we've agreed to prioritise, as an organisation, seven use cases that we believe will help us develop patterns to actually be able to use AI responsibly at a broader scale, beyond the personal productivity that it is currently being used with. Ms SITOU: Can I ask about those seven use cases that go beyond the personal productivity and agency-wide productivity. Who is driving that? Who is driving the benefits and the innovation in those seven use cases? And what are those seven use cases? Mr Watson : The decision was taken by our executive earlier this year on those seven. We're now setting up a program that sits underneath the chief data officer—myself—reporting through our data analytics assurance committee. We're working with our usual processes around the value framework of any new initiative in the ATO. AI is a little bit different because it is novel. It's a little bit of learning as we're going around how we adapt that framework for the AI use cases going forward. The other important part of it is that each of those use cases has a senior officer from the business areas of the ATO owning it. We're wanting to have the benefits driven by the business and not by the technology parts of the ATO. So we've set those up. They're different in size and scale—the various use cases. For example, one of them is about how we can use Copilot in financial reconciliation activities through our ATO finance area. That'll be leveraging some of the tools in the Copilot environment. At the other, broader end, it's about how we can use AI to do some of the heavy lifting for information assembly and sorting summarisation for some of our audit-type activities where, as a result of a request for information from a taxpayer, we might receive thousands of documents. It's about how AI can help turn a virtual shoebox into a prioritised, organised list of documents and help to summarise and assist our people to then analyse that more efficiently. Mr Sawade : I'll add to that. To summarise our approach, our seven use cases that the organisation has decided to focus on are driven out of our central data area, called our Smarter Data Program, under Mr Watson as a program manager. But, as mentioned, each of those use cases has a sponsor, an executive sponsor in the business area that the use case relates to, which is then really responsible for understanding how we are going to measure value, how we are going to understand if the activity in the use case actually delivers benefit or not and how we not just do that as a one-off but industrialise that so that we're continually measuring and understanding the benefit as tools change and as technology changes. That's managed through the central governance of the program that's looking after those seven use cases. Ms SITOU: I might have some more questions to the DTA, if I may, but to the team at the ATO: how are staff able to safely innovate using AI given the sensitivities of the data? You've got the seven use cases, but what if they've struck a gold idea? They just want to see if it will work. What are the processes that are in place to enable a bit of experimentation? Mr Watson : Those seven are the priorities that we're investing in as an organisation. As for other use cases, we have published internally a list of approved uses. If staff come up with a different way underneath an existing use case, they have permission to proceed. If it's new and novel, they need to come to my area under governance to look to register that use case. We need to then consider it under the broader DTA policy, do a risk assessment—What's the level of risk attached to it? Does it use personal information? Do we need to consider a personal information assessment?—and do a data ethics assessment. That is under our senior executive instruction. Staff need to come forward to get that use case registered and approved before they experiment with that. But the first step is checking: does it fit under an existing use case? That's particularly given the spread of Microsoft Copilot and particularly for some of those uses of it as a personal productivity, personal assistant tool that isn't using, for example, sensitive or taxpayer information. A lot of those would be more broadly covered under some existing use-case patterns. Mr Fechner : One thing I'd like to start out with is that the Australian Public Service AI plan is built on three towers. It's people, tools and trust. We've got a focus from the APSC that's looking at how we're getting our people ready to be able to use the technology, by training and guidance in that particular space, and we've got the Department of Finance that's leading on the development of tools such as the GovAI platform and the implementation of accelerators through the AIDE program. I'll let Mr Hubbard talk more about that. From a trust perspective, the DTA has really pushed to have policies in place before the rapid adoption of artificial intelligence. We've focused on transparency, we've focused on minimum capabilities and we've worked with large sections of the digital industry to put in place clauses and protections. For instance, the ATO mentioned the Copilot implementation—Microsoft's generative AI tool that has been around for a couple of years in the Commonwealth. The DTA ran a program on that that started with training and started with a security assessment that made sure that Australian government information wasn't going to leak out into the models or into the public arena. Privacy and integrity of the information were guaranteed through both an assessment through the ASD approved IRAP and contractually through our volume sourcing arrangements. The DTA has got significant guidance and implementations in contracts that protect our data against being used by models and ensures that private or personal information has got limitations on how it's used. We support agencies with an AI assessment capability, and it's a mandatory requirement for all high-risk use cases to go through this assessment criteria. We're in the process of establishing an AI review committee—where we have particularly sensitive AI cases that are going through, we're going to get a body to have recommendations to the agency about what controls would support them in safely deploying that technology. That could even go right up to a recommendation that they actually delay an implementation in those particular spaces. But I might pass to Mr Hubbard to talk more about GovAI and AIDE. Mr Hubbard : In response to the question, 'What are the areas where people can experiment and learn?' the GovAI platform that the Department of Finance runs is the whole-of-government platform to provide that shared training, common tools and secure infrastructure. It's got a range of services that are part of it. Maybe the most relevant here is an AI sandbox environment that people new to generative AI can use that allows them to see through structured tasks, structured learning, using different AI models—what you get from the model and what comes out of it. The advantage of that is that it's not experimenting with AI on your own network and your own data. It's actually in a safe sandbox environment. The GovAI platform also provides, for some agencies, separate sandbox environments where they can also experiment and build and test different AI approaches before they take them back to their agency to deploy, potentially. Finally, the other thing I'll mention is a trial of something called GovAI Chat, which has just begun, to allow everybody in the APS eventually to have the opportunity to access safe, secure models in a controlled environment as well. Mr Fechner : Maybe Ms Liu can talk a little bit about the capability uplift that the APSC is leading. Ms Liu : I work within an area in the APSC called the APS Academy, and we are supporting, under the APS AI plan, foundational capability for the whole of the APS. That's in two broad cohorts. One is things we think that all staff need to know. Under that, Mr Fechner referred to the responsible use policy. One requirement under there, for example, is that every APS staff needs to do some mandatory foundational training. We provide that training on our network, and it's available to the whole of the APS. It was actually produced and created by the DTA. So it really has, at its core, understanding what the technology is, alongside what frameworks are in place to support and provide guidance and guardrails for staff, and then a series of ways of thinking about testing and understanding if it's appropriate to use AI in this particular context, like, 'Is it the right tool for the job that you are seeking to achieve?' I would just add— CHAIR: I'm very sorry to cut across, but we're just running out of time. While that's a helpful overview, there are a couple more questions. Ali, I'll hand over to you. Ms FRANCE: Thank you all for coming here today to help us with this inquiry. Across the Public Service—this is probably for the Public Service Commission—do we have a system for constantly tracking the current use and expanding use of AI? I'm generally interested in how we track it because every day I seem to be getting, even from my staff or other people, 'Oh, I found out how to do this,' and it's changing constantly. I'm interested in how you are tracking its use across the public service. Ms Liu : There are a couple of components that I'm hearing in the question. One of them is how we transfer knowledge between people in a really dynamic environment about how to use AI. I would say there are some fantastic forums, some of which the Public Service Commission are involved in and some of which the DTA are involved in. I would point you to something like the Digital Profession, which is where lots of people—and they may or may not be in digital roles—are really interested in AI. There are, I think, more than 1,500 people in that forum who are, on a regular basis, teaching each other about how to use AI. The other whole-of-government platform we're using in the academy is a series of webinars which are open to all APS staff and really draw on some of the use cases that GovAI, for example, has established or determined that exist across the service. These are really just practitioners who are teaching and sharing their experience of how to use AI in their governance contexts and talking about how to navigate the systems that they have, both technology and broader governance and risk and assurance type systems. That's one aspect of your question, but I will pass to Mr Fechner and the DTA because I think there's a more structural question around how we understand what agencies are using when it comes to AI. Mr Fechner : Thank you. One clear area is the transparency. Under our policy, all entities within the APS that fall under the policy—which is the vast majority of entities related to here—have a requirement to publish a transparency statement on an externally facing website that identifies the type and use of AI across their entities. Under the new version of the policy, they're required to keep an internal register. You've heard the ATO talk about it, and you've heard Health, Disability and Ageing talk about their internal registers. All entities are required to have that internal register. The DTA also monitors the policy by collecting all of the details associated with the transparency statements of all entities. The complexity around this is that the use and embedding of AI into standardised products are just happening everywhere, and it makes that a very volatile situation. Another area that I would call out is a control that's been put in place. The Department of Industry, Science and Resources have the AI Safety Institute as part of their function. Three really significant recent announcements are memorandums of understanding with OpenAI, Anthropic and Microsoft that are talking about the safe and responsible use of AI and how government use cases can be safely and effectively evaluated. It is a rapidly emerging environment, but there is a high degree of obligation across all Commonwealth entities to officially and transparently make their use of AI in particular areas. When we talk about the areas, that could be in service delivery or it could be in forecasting or it could be in policy development. Then they'll talk about the particular tools, whether it might be generative AI, facial recognition technologies, video transcribing or those particular things that are broadly used. But the DTA, in a final piece, also has specific arrangements with most of the large software providers that require them to act in particular ways under terms of our contracts. Ms FRANCE: In terms of AI generated material coming into departments, whether it be applications for things or people impersonating somebody else—even into our electorate offices we get AI generated stuff now and, quite often, it is very hard to detect—I'm really interested in how the departments are training staff. I know AI can help with productivity, but, if there's a huge influx of AI generated materials incoming, how are you dealing with identifying that? Do you rule it out? How do you deal with that generally across the service? Mr Fechner : Certainly we're not promoting people not being able to use generative AI to engage with government. It is very pervasive right now. What we're actually doing with some of the fundamental training is to get public servants to understand and realise that content that's coming through that's being used for generative AI is not always to be trusted. We do have within our contractual agreements for people that are engaging on a commercial basis with government that they disclose they've used AI in the production of materials that are used and given to government. We also require them to check their own sources before presenting them to government. You may have heard of a couple of cases where that process needed to be refined. But it is really saying that public servants should not just trust it. The great risk associated with generative AI is the quality and the convincingness of how it's put together, where it could be completely and totally hallucinated. It's actually about that fundamental awareness training that says: 'This technology is pervasive. It's here.' There is the secondary or the stopgap position. You often hear scenarios associated with people creating resumes that are AI produced for jobs within government. That's why we have multiple evaluation techniques. The first one might be a filtering one, but you would expect in an interview or an assessment process that they'd be discovered as being faked through AI. There is also a large number of standards. The ISO 42001 standard is about how you manage the software lifecycle. We have new developing standards about trying to find fakes in images that have been generated, and the government's already making some moves in that particular space as well, with some legislation especially about sexually explicit images. The fact is the models are changing so quickly. Some of the earlier AI technologies that were used to discover where AI content had been created are becoming less and less effective because the technologies are getting so much better. I personally have a little thing: most models are North American; whenever I see too many Zs in a document, I know that AI has been used in it. There are things that we can go through, but it is past the point of no return where people in all aspects of their life won't be using generative AI to draft documents, to submit grant applications. so it really becomes about how we educate and support our people to make good decisions afterwards? CHAIR: I'm conscious of time, and there's a lot of ground that we have not covered yet, especially around data security, AI infiltration of government data, processes and governance around the protection of data by AI, from AI and the safety controls of AI. Have there been any incidents where AI has been adopted in a particular area of government and there have been incidents or issues of concern? We haven't covered those sorts of things. I think there's still more to go into around governance of the adoption of AI programs and processes, beyond the webinar style or level of engagement, but actually around the oversight and the implementation of policy. I still feel like there's a lot more to cover. We will submit questions on notice, but I think there's a high degree of probability that we might recall this group at a later date to have a second hearing, because I think there is a lot more to be done. I would definitely encourage agencies to give as much information as can be provided on notice. That will then potentially alleviate the need to gather everyone again. I know everyone's got a lot to do and is very busy. But, for the purposes of making sure this committee is able to provide recommendations on such a substantial matter, an hour is a difficult task. I'll finish with saying that. I do thank everyone for being here today. If you've got questions on notice where you've been asked to provide additional information, could you please forward the responses to the secretariat by Wednesday 13 May 2026. If any other committee members have further questions, they'll be sent to you in writing through the secretariat. Thank you, Broadcasting. Thank you, Secretariat. Thanks, members, for being a part of this today. Thank you to all of the agencies and the witnesses. Committee adjourned at 12 : 40