DERBYSHIRE, Mr Peter, Acting Chief Executive Officer, Australian Academy of Technological Sciences and Engineering VERSPOOR, Distinguished Professor Cornelia (Karin), Fellow, Australian Academy of Technological Sciences and Engineering CHAIR: Welcome. I understand that information regarding parliamentary privilege and protection of witnesses providing evidence to our inquiry has been provided to you. I now invite you to make some opening remarks, and then we'll proceed to questions from the committee. Over to you. Mr Derbyshire : Thank you. Good afternoon, members and senators. We'd first like to take a moment to acknowledge the traditional custodians of the land we're meeting on today, the Wurundjeri Woi Wurrung people, and reflect on their long history of achievement as Australia's first scientists and technologists. The Australian Academy of Technological Sciences and Engineering is an independent non-government organisation led by a diverse fellowship of over 900 of Australia's leading applied scientists, technologists and engineers. This includes many AI experts, including Professor Karin Verspoor, who is joining us today. From finding recipes to summarising reports to generating the next level of video games, AI is increasingly being integrated into the way Australians live, work and play. AI is already improving diagnosis in medicine, productivity in business and emergency management and predictions. With considered investment, all this activity could boost Australia's GDP by $235 billion by 2034, according to our modelling. Last year, ATSE released a report that showed where to invest to make this happen: talent, training, compute facilities, national models, regulatory reform and regional AI hubs. To realise this economic benefit, there is a need for considered regulation and guardrails to manage the risks. International frontier models have shown they can hack government systems despite developers putting safeguards in place. AI misuse can lead to inequalities being reinforced, people being scammed, deepfakes and misinformation. Meanwhile, the data centres needed to build the models and process queries have become increasingly controversial. Demos polling has shown more Australians have a negative view of data centres than a positive one. Risks of price rises for power around large-scale data centre adoption, as we have seen internationally, and water consumption on an increasingly drying planet need to be managed. ATSE commends the government for its work through the Department of the Prime Minister and Cabinet to consult on regulations for data centres and frontier AI models. For data centres, clear, graded regulation can help manage power, water, noise and location issues that can turn communities against them. For frontier AI, Australia needs to invest in local models, data sets and infrastructure. Regulation should follow international norms such as the European Union's AI Act. By combining strategic investment with sensible regulation, Australia can seize an AI opportunity, enhancing the benefits and minimising the risks. Thank you. CHAIR: Thank you very much. We really appreciate you being here to provide evidence and for the work you are doing in this space. What I'm gathering from the evidence we're hearing and the conversations that are happening in this space is the importance of sovereign capability and the opportunities that we have at the moment, in terms of both how we are progressing in this space generally and, in particular, the potential advantages that Australia has to leverage off the interest that some of the hyperscalers have in coming here and developing and training here. What level of sovereign capability do you think is realistic for Australia? In your view, where should the government focus its efforts? Mr Derbyshire : Please feel free to jump in at any point if you have some thoughts. Essentially, it's a 'how long is a piece of string' question. There are areas where we absolutely need our own sovereign capability, and it depends on what we consider to be critical infrastructure. For example, health technology and health data and health AI are things that we are going to want to make sure remain onshore in Australia. The areas where we have competitive advantage, such as gathering climate data and data from things like the Square Kilometre Array in WA—all of these things that require the use of big data and AI—are areas where we're going to want to see Australia play a role in it. Do we need to see Australia developing 30 or 40 of our own large language models? Probably not. But do we want to see where we can make sure we have at least one or two that we can develop on Australian understanding and knowledge? That's probably a place to put that. So there's no direct answer that I can give that says, 'This is exactly how much sovereign capability we should have.' It also depends very much on our partners, because we know that we can partner with other countries to develop this, but those partnerships may not remain stable in 50 years time when these technologies will become part of everything we do. Prof. Verspoor : One thing I'd like to add is that, indeed, in areas where we want to be able to manage and regulate and control the data that's being used to train models or is fed through the models, it's very important for us to be thinking about the privacy and the safeguards that we can put in place. As soon as it leaves the country, it arguably leaves our control, and that creates tensions. It's also worth adding that we really need the ability to localise the use of the models and think about strategies for adapting it to local problems and the local context. A model that's generic or a frontier model that's been trained on the entire internet is not necessarily something that's going to be fit for purpose here in the local context. So I think it's important for us to be looking at ways to leverage, of course, the large-scale models that are being developed in other places but also ensure that we have ways to manage them safely and securely and to adapt them to local needs. CHAIR: What could that mean for Australia, particularly in our own AI ecosystem and throughout the economy, when we're able to have training on Australian data, in terms of local contextual advantages that we can deliver here for Australian businesses but also then potentially export elsewhere? Prof. Verspoor : Yes, it's super important. I do a lot of my work in medical applications of AI. In that context, there are some really obvious differences between the way our health system works and the way, say, the US healthcare system works. A lot of my work is actually on free text in electronic health records. It's safe to say that, at the moment, most of the AI applications being developed in that context are being developed on a single dataset coming out of the Boston area in the United States. We know that the local context and even the local documentation practices here in Australia are different to those in Boston, so it's really important for us to be able to make sure that we can develop technologies that are going to work for us here. That means that we need to have local people who have the skills and capability to translate the problems that the healthcare sector or businesses in general are facing here in Australia to develop solutions that matter to them. CHAIR: On that, where do you think we should be focusing our efforts on building up skills and capacity, and what's our opportunity? We hear that there are lots of Australians who have significant skills in this space but are going elsewhere. There's currently a large amount apparently in Silicon Valley and the like. How do we bring them back to Australia? How do we build up our own skills and capabilities locally? Prof. Verspoor : Yes, we need to invest in that capability. I think there's a perception out there that we don't need to train computer scientists or software engineers anymore because AI can just write the code. But there's actually a misconception, I would say, around that. The challenge of taking a technology which is not fit for purpose and is generic, really, and adapting it to the local context requires people on the ground able to understand the local needs and do that technical translation into a solution. There's also an idea that you can just download AI—whatever that means—from the internet and just do great things with it. But AI is actually a collection of technologies, and we need to have the skills in place to do that. That, of course, requires investment. It requires education and training. We need to be upskilling our workforce so that people understand both how to deploy and leverage the capabilities of the AI but also how to integrate it into the context where we use it very safely. CHAIR: Thank you. Ms Chaney. Ms CHANEY: I'm interested in the comments you've made about energy use with data centres, your views on the importance of new data centres underwriting additional renewable generation and storage for 100 per cent of their demand and any thoughts you have on how we deal with the lag between the time for a new data centre to get up and running and the time it takes for a renewable project, which is generally longer—and how you define 'additional' in that context. Mr Derbyshire : I want to start by taking it a step back and just put the power usage of data centres in the broader context. It's first important to consider that as we develop more in the area of data centres, we're going to develop new technologies that make them more efficient. Our work in quantum computing here in Australia should actually make data centres more energy efficient than they currently are. I also think it's important to remember that, currently, the amount of energy being drawn on the data centres already in place in Australia is incredibly small when you compare it to things like aluminium smelters and industries like that, which do tend to draw a fair amount of power from the grid. We are asking data centres to put a bit of extra renewables into the system for them, because we know that they're going to be long-term ongoing draws on the power. But we also think it needs to be put in the context of industry more broadly. That being said, data centres can provide this opportunity to help stabilise the grid as well. They provide an ongoing base load to the energy system that means our renewable generation can actually be more stable and more easily managed. There is always going to be, as you said, a lag in the development of renewables. I think it's important to make sure that the data centres are investing in the development of these projects, and it is up to the government to decide and regulate how long of a lag there is allowed to be in that sort of space. But it should always be part of the discussion, as we should consider any new industry that is drawing a lot from the grid and that should now be asked to perhaps invest in renewable energy as well. Prof. Verspoor : I might add a point related to the previous question around skills. One of the things that we can be looking at is more effective use of the data centres in the context of AI. There's ongoing work, mostly in the research community, to miniaturise models—to take these big, huge models and try to make them more compute efficient and effective. This requires people to be thinking about new ways to build these systems that have the same sorts of outputs but not necessarily the same energy usage. It goes together. Ms CHANEY:   I would love some views on why reporting energy, water usage and efficiency is important and what it would look like in practice. Mr Derbyshire : I think it's important just because it is hard to tell what is being used unless it is being reported. You can't really tell what you're not measuring. I think it's important because data centres are somewhat controversial within the community at the moment. Providing that information to the community and to the public about how much is actually being used and in what context that falls in—how much of a percentage of the grid is it? How does that compare to other industries? We do expect other companies and industries to report on water usage whether it's mining or industry and the like. It's not always public, but they are expected to report on that. I think, again, we don't want to see the developing of data centres to be given too special a context because they are an industry and they should be required to report in the same way other industries do. Ms CHANEY: Thank you, Chair. CHAIR: Ms Mascarenhas. Ms MASCARENHAS: One of the things that I'm curious to understand is the skills pipeline. It'd be good to understand where you would rate Australia at the moment but also what we need to do over the next 5 to 10 years. Prof. Verspoor : I would say that we are doing pretty well. We actually have quite a thriving network of universities that have very robust training available in computer science, data science, artificial intelligence, cybersecurity, general information technology—all of the areas that are critical to this arena. We have people coming from all over the world to come study with us in those areas. We are ready to contribute to that. Mr Derbyshire : I do think, however, an area that we probably do need to improve on is AI readiness for the broader workforce. That's not necessarily for those that are working in the IT area, but countries like Singapore are working on having nationwide education opportunities for everyone to understand and work with AI. Now you say, 'What do you mean you can't use a word processor? This is going to be something that in 20 years time you're going to look at someone and say, 'What do you mean you can't use an AI system?' or 'You don't know how this sort of thing works?' I think we need to look at that on a broader piece on not just making sure we're ready to make the most of it in an industry capacity but actually making sure that small-business owners know how to implement these new technologies in the same way we needed to support them to develop new payroll management software and things like that. Ms MASCARENHAS: I think that one of the things, Professor, that you may have talked about is bias and loss of control. Can you talk about that a little bit more? Do you think that AI models should be independently tested, and, if so, who would be the right people to do that testing? Prof. Verspoor : I absolutely do. I'm a researcher, so obviously I'm going to say that. But, yes, I think it's extremely important that we maintain an evaluation lens over all of these technologies. I just go around asking questions. I ask: 'What did you use to train this data? How did you assess it? Who were the people you considered?' And, in fact, we need to move— Ms MASCARENHAS: Can you explain some of the biases that you've seen through your research? Prof. Verspoor : Yes. There's a whole range of things that can happen. Any machine learning model that's data driven will reflect the biases of the underlying data. If you think about a large language model that's been trained on the entire internet, which most of the frontier models have been, it's going to reflect the patterns of that data. And that means to the extent that there are biases in the world that are being captured on social media and on any website really—those models will reflect that bias. The world is biased. There's no way of getting around that. But that also means that the view that the models have are biased, and we see, for example, in the health context in my area, there are biases that creep in because, for example, most clinical trials only consider—well, predominantly consider—white people compared to other people of colour. They predominantly, traditionally, have considered men more than women. The vast majority of genomic information we have, for example, is from European men. And so all of this gets baked into the models because if we don't see a diversity of data, then the data will reflect that bias. There are strategies for dealing with that—and actually talking about reporting and measuring things, you have to start by measuring things. That's where the evaluation comes in. Moving beyond just an overarching evaluation where you look at performance across the data set you have, which is inherently biased, to understanding that there are subpopulations within that data or subgroupings of data that you then need to evaluate separately in order to understand the impact on those specific groups. Ms MASCARENHAS: Do you think that there's a common understanding of what the safeguards look like, and is there training happening in universities to give skills to professionals to actually be able to implement and monitor these safeguards? Prof. Verspoor : So from— Ms MASCARENHAS: And if you want to talk about safeguards from a structural systems perspective, I'd be— Prof. Verspoor : I think the basic answer to your question is no. We have quite a narrow view of what the safeguards are and indeed even how we do this evaluation. A lot of the evaluation that we do is from a technical perspective: what's the performance of this system? But we aren't considering enough the human aspects of evaluation. For example, we are doing some work looking at how we can evaluate AI scribe systems which are the systems that are being used by doctors now to automatically produce documentation for their consultations. When we look in the literature, we see a lot of work that focuses on how much time has been saved for the doctors and a little bit about what the quality of the documentation is but nothing really that has examined the patient experience. What's the impact of having a scribe in the consultation room or what are the workflow impacts for the doctor? And so we need to kind of broaden the areas that we consider and really consider the impact on the people using these tools or that they're being used on. And that's something we need to do more of. Ms MASCARENHAS: I'm also keen to understand sovereign capability. I'm from WA and what we have seen sometimes from the resource boom that we've seen over there is that we haven't always been intentional with the way that we use local skills—whether that relates to the designing and building of a project and then whether that is manufactured locally or overseas. Even though we've seen all of these resources being extracted, we've not consistently built on skills. I'm curious to know your take on how we can build sovereign capability at every single chain from, let's say, design up until actually having compute available. Mr Derbyshire : I agree with you 100 per cent in that we do need to make sure we're actually getting value out of these new industries as they come up. We do need to do it in an intentional fashion. I think it goes from everything from skills to making sure that the infrastructure is available for researchers to use when larger infrastructure is not always available. Everything including—we've seen in the past tech companies develop their own training centres and the like. What I'd rather see is the tech companies working with our universities and TAFEs to develop those skills and partnering with those local systems rather than the data centres being built and prescribing all of their time to their international companies. We need to make sure that the compute power is also available to Australian researchers and Australian small businesses. It is about looking at every step of what gets built, how data centres and AI are developed, finding what's happening in Australia already—and there's a lovely map, I think, that was released earlier today for data centres—and where can we develop these things. One example I have is that in—and I can't remember the name of the company unfortunately, which I feel bad about now—Western Sydney, there is a company who is building the cooling equipment that is going into data centres. They've looked at this. They used to develop cooling equipment for machineries of another kind. And they said, 'Actually, data centres are a new place where we can add value and create this supply chain.' I think it's about looking at everything from the physical infrastructure, the people infrastructure and then the actual digital infrastructure—what is critical, what do we see as critical infrastructure?—and then how we go and say to these companies that do want to come here; 'Come. We'll support you. But here are some things we need you to do to give back to the community.' That is not something that I think we've done in areas in mining, where they do extract the resources and employ people in jobs—absolutely—and they do work on training, but there's never a direct requirement for them to do so. So I think we can learn a lot from the mining sector in WA and ask, 'How do we now take this to a national level?' Ms MASCARENHAS: My final question is on social licence, and I'd say there are parts of data centres/AI technology where we're struggling to see that social licence. I'm curious to know whether you have a particular view of the right way to do that and what you see as the challenges. To some extent, I almost see lessons from the resource sector being repeated for a brand-new industry. I don't know whether it's because—and being an engineer, I can say this—sometimes we don't think about the people, and fundamentally if you don't think about the people then a whole project can get cancelled. So I'm wondering whether you've got a particular view on that and also whether we need to ensure that technical professionals understand how important that is because you're project can get cancelled. Prof. Verspoor : I agree completely that we really need to centre the people who are going to be using and going to be impacted by these technologies. The social licence element of it really means explaining to people what the value of these systems can be for them, not in terms of efficiencies, cost savings or productivity but, rather, why a particular system has value or benefits for the human beings who are there. We can establish that in various ways. Part of it is through education, through explaining to people what the possibilities are, and also co-design and involving them in the process of: 'What do you want? What do you need? What are the gaps?' I can give you one concrete example. Yesterday I was at a meeting organised by BioMelbourne around biobanks and clinical registries, and I heard somebody say—this sentence just stuck with me—that the patients who have these rare diseases are absolutely willing to give us every single data point they have, because they want answers. The problem is that we don't have a great way to collect that data and we don't have a good way to share it, and therefore we can't do the good work of analysing and deriving benefit that goes back to these patients. So they want it, but we can't give it to them. And why do they want it? Let's think about that. They want it because it's going to have benefit for them. So I think if we centre the benefits that people can derive from these technologies then we will get the social licence. Mr Derbyshire : I think it's also important to bring in—and a shout-out to our colleagues at the Academy of Social Sciences in Australia—the social scientists. It's important to have in the room the social scientists who are looking into how we develop the social licence and what the social impact is, because people's experience with AI is through chatbots, ChatGPT, Claude and all those things. That's their experience with AI—and seeing the horrible AI-slop pictures that appear all over social media. They don't necessarily get a chance to see and hear the stories about what AI can do in the medical space, for example, or how it is helping us to predict bushfire movement, climate change and things like that. They're not hearing those stories. I think that's where the companies that are developing these things need to be out there talking about the stories more. They need to engage with the social sciences people, because the technical people aren't always the best people to tell the stories. I think we need to start building that. Some of this we've missed the boat on. There are a lot of people out there who are just going to dislike AI now, and there's nothing we can do to change that. But there's always going to be that middle ground where we need to start telling the positive stories of what this can do as well as reminding everyone that we aren't just letting this run rampant; we are actually looking at how we're going to get benefit out of this while trying to avoid as many of the risks as possible. Prof. Verspoor : Can I add one thing. I've been trying to poll the young people I encounter—much younger than me—in my daily life, and I've been asking them what they're worried about. They're worried about their jobs. They're worried about investing years into study that is then going to be replaced by an AI system. I think we really have to centre them. We have to centre the opportunities that working with AI effectively can create, but we have to make sure that we're not actually going down that doomerism line of thinking to end up with a society where we're controlled by the AIs and have lost the humanity in there. I think that's going to mean confronting some difficult questions for us about where we're comfortable inserting this technology into our daily lives. Ms MASCARENHAS: Thank you. CHAIR: Mr Watts. Mr WATTS: In your opening statement, you talked about the need for Australia to align its regulatory approach with those of other jurisdictions, citing the EU, for example, and the EU's AI Act in that respect. Should we interpret that as being a broader approach as well? Obviously, the EU's AI Act sits alongside a copyright framework where there's a text and data mining exception with a technical opt-out requirement in operation there and a machine-readable opt-out type of requirement. Is that something that you'd recommend that we follow as well? Mr Derbyshire : I'm not going to pretend I'm a copyright expert on this particular one. I can probably look a bit more into it and get back to you on that one. For the most part, though, it is something that we don't necessarily have an answer for in place at the moment. It's definitely worth investigating. We know that the AIs are not just being trained on the internet data but making deals with academic publishing houses and the like to get access to the research that they can then train that sort of data on. I think it's important to make sure that the people who have done the work get the benefit from this training, but I don't necessarily have the detailed IP answer for you, but I can certainly get one. I'll certainly have a look at it and get an answer to you, if you like. Mr WATTS: When you're following that up, I'd be interested in the perspective of that in the context of the Swiss model that you talked about. Your submission cites a Swiss AI model. I presume that's Apertis. Is that right? Mr Derbyshire : Maybe. I think so. Mr WATTS: We can follow it up in bulk. You were saying that it's an open model with infrastructure costing about $70 million a year. As I understand it, that model was trained in Switzerland, taking advantage of that scientific research exception and that text and data mining exception under copyright. When you're answering these questions on notice, can we get an understanding about the role that that exception played in facilitating the training of that model in Switzerland and whether there were learnings for us from that. Mr Derbyshire : Absolutely. Mr WATTS: Another thing you specify in your submission is making data centre approvals 'contingent on a commitment to include a genuine R&D presence'. I presume that's aimed at the larger scale investments. It would be particularly important, I suppose, with a frontier training investment. In your view, what would a genuine R&D presence look like? What are the kinds of things that government should be aiming for in specifying, if we are trying to mandate such a requirement as part of the rollout of AI infrastructure in Australia? Mr Derbyshire : I'll pass on to my colleague as well to talk about more of the R&D stuff, but, from a regulatory perspective, I would see genuine R&D investment as partnerships with our research institutions, whether it is universities or whether it is the CSIRO or other organisations. I think those are what I see as genuine R&D opportunities. We do know that it's possible to put forward and do their own in-house research and development. That would be great, but I think we want to see these public-private partnerships as part of all of that role. Mr WATTS: For someone that doesn't have a technical background, what could those partnerships look like? What is being contributed by each side of that partnership, what are the shared projects that might be worked on? Mr Derbyshire : Some of the shared projects would be some of the work that I imagine Karin has probably worked on in that area, so I'll pass on to you, if you've got any specific examples. Prof. Verspoor : As a technical person, I said I work in biomedical applications. I'm not a clinician. I always start every talk with the No. 1 disclosure that I'm not a clinician. In order to work on these problems as a technical person, I need to work very closely with the people who have the context and the expertise in the clinical environment. It has taken years to develop trusted relationships with clinical researchers and clinicians in order to develop meaningful solutions. I think that's where bringing the expertise from a research organisation together with the expertise from a business context or a health context is critical. It means sitting around the table and understanding what the real pain points in the day-to-day work of an industry or partner and then leveraging the creativity and the innovation from the university research to come up with a solution to the problem. I would also add, as Peter mentioned earlier, the need for social scientists and sociotechnical research around this. That means that we're not just talking about solving particular problems in particular moments of time, but understanding the broader translation of that research, or innovation into real world contexts and what implications that has. Mr WATTS: So you're not just talking about a reserved compute access, it's more of a deeper collaboration with researchers and technical experts in these firms and partnering them with domain area experts and seeing what can be made together. Prof. Verspoor : Yes. Both on the solution side and the evaluation and validation side. I mean, one thing that doesn't maybe happen as much as it needs to in industry context is really taking that step back and saying, 'how well does this work? What are the limitations? What are the problems?' You need someone to go in and do some measurements and do some evaluation. Typically, that is done, if it's done through an independent research evaluation and following a rigorous scientific process, you're going to get much more meaningful evidence to back up claims about how well something works or how much we should trust a particular system. CHAIR: Senator Dolega. Senator DOLEGA: I just have a quick question. I've asked some of the other witnesses about training for the community and how we bring people along to keep them up to date with AI in workplaces, in the community and in society in general. Obviously, in Tasmania we've had a lower level of digital awareness in the past, so how can workers, people and parents in regional areas be trained to keep up with AI, so they can keep progressing, but also protect their kids and members of the community? Mr Derbyshire : It's a difficult question because any type of training requires the person to first want to be involved in that kind of training. If someone doesn't want to be involved in it, there's not much you can do really to help that. So it's about making things available to people where they are. In regional areas, for example, you're not going to be able to bring them all into a big room and teach them how to use AI, but you're going to be able to provide them with those, I guess, workshops and webinars that can provide simple guides on how to keep your child safe in using AI at home or things like that. This is not too dissimilar to what we've seen when the internet became commonplace everywhere and is now available on all our phones. The challenge has always been: how do we teach people to use this technology safely and effectively, but still get the best out of it? We've never solved that problem perfectly. We still know that there are people out there who don't have the same technical ability and digital literacy, especially in regional and remote communities. I think this is one of those opportunities to learn how we didn't do very well in the rollout of the internet and perhaps go: how do we fix that this time around? If I give an example of AI in schools, there are going to be some states where they are jumping on AI for students with both hands, and there are some states that are very slow or very cautious about this sort of thing. Both have their benefits as an approach, but we also have to take into account if people are stopping students from being too involved with AI, then how are they going to compete in an international global workplace when they finished school, when they finish university, if they've not had this experience. So I think it's a combination of challenges there. There are lots of programs out there that we run ourselves on getting women into education in STEM degrees. There are a lot of programs out there that are now looking at digital—and I can find you some names of them—skills for migrant women and the like who have come in and don't understand how these sorts of things work. There are programs that are going out to schools and giving students the opportunity to understand coding and how that works so that they can understand that. There are a whole collection of those programs out there—and I'm happy to send you some of those after this—but it's not an easy solution. Senator DOLEGA: It's not, and I keep on thinking back to the days we had online access centres. They're becoming a thing of the past, unfortunately. People could go in there and jump on a computer. There'd be someone from the library who would help them to learn about how to do a web search or how to get an email and everything, but when it comes to AI—I previously went to a conference and heard an analogy that humans with AI will overtake humans without AI. It's really concerning to me that people in regional areas have a greater opportunity of being left behind. I'd be very curious as to some of the ideas that you might have as to how we can include and bring people along so they don't get left behind with these new opportunities. Prof. Verspoor : I would say we have to think about this in a context-specific way—even an application specific way. Teaching somebody about the risks of AI in the context of banking or being aware of how a phishing attack might come through sounding like—with the voice of—your child is one context. Education is a very different context. All of these contexts require a different solution. I don't think we can just talk about generic AI education. It's not meaningful. It has to be tied back to particular challenges or contexts or environments that people are facing. How to bring that to a regional or a remote community really comes back to where it is impacting them the most and where we need to prioritise, starting with some education. It might mean understanding where they feel that they're missing something and starting from that. Mr Derbyshire : You did mention the library systems and how that helped give people that digital literacy where they are. The local libraries could be involved quite effectively yet again in teaching people how to use AI, what's involved and how it might help them. CHAIR: We have come to time but thank you so very much for your contribution and the evidence that you've provided the committee today. For any information or any questions that you have taken on notice, please connect through the secretariat and provide it back within the next two weeks..