KING, Dr Stephen, Commissioner, Productivity Commission [by video link] NUCH, Mr Hudan, Director, Productivity Commission [by video link] Committee met at 18 : 31 CHAIR ( Ms Briskey ): I declare open this hearing of the Joint Select Committee on Artificial Intelligence. I begin by acknowledging the traditional custodians of the land on which we are meeting, and I pay my respects to elders past and present. I extend that respect to any First Nations people joining us today. These are public hearings being videostreamed live via the parliament's website, and a Hansard transcript is being made. I remind all witnesses that, in giving evidence to the committee, they are protected by parliamentary privilege. It is unlawful for anyone to threaten or disadvantage witnesses on account of evidence given to a committee, and such action may be treated as contempt. It is also a contempt to give false or misleading evidence. Witnesses also have the right to request to be heard in camera. The Senate has resolved that an officer of a department of the Commonwealth or of a state shall not be asked to give opinions on matters of policy and shall be given reasonable opportunity to refer questions asked of the officer to superior officers or to a minister. This resolution does not preclude questions asking for explanations of policies or factual questions about when and how policies are adopted. If a witness objects to answering a question, they should state the ground upon which the objection is made, and the committee will determine whether it will insist on an answer, having regard to the ground which is claimed. If the committee determines to insist on an answer, a witness may request that the answer be given in camera. Commonwealth officers appearing today are also reminded of the Senate order specifying the process by which a claim of public interest immunity should be raised. A copy of the order is available from the secretariat. The extract read as follows— Public interest immunity claims That the Senate— (a) notes that ministers and officers have continued to refuse to provide information to Senate committees without properly raising claims of public interest immunity as required by past resolutions of the Senate; (b) reaffirms the principles of past resolutions of the Senate by this order, to provide ministers and officers with guidance as to the proper process for raising public interest immunity claims and to consolidate those past resolutions of the Senate; (c) orders that the following operate as an order of continuing effect: (1) If: (a) a Senate committee, or a senator in the course of proceedings of a committee, requests information or a document from a Commonwealth department or agency; and (b) an officer of the department or agency to whom the request is directed believes that it may not be in the public interest to disclose the information or document to the committee, the officer shall state to the committee the ground on which the officer believes that it may not be in the public interest to disclose the information or document to the committee, and specify the harm to the public interest that could result from the disclosure of the information or document. (2) If, after receiving the officer's statement under paragraph (1), the committee or the senator requests the officer to refer the question of the disclosure of the information or document to a responsible minister, the officer shall refer that question to the minister. (3) If a minister, on a reference by an officer under paragraph (2), concludes that it would not be in the public interest to disclose the information or document to the committee, the minister shall provide to the committee a statement of the ground for that conclusion, specifying the harm to the public interest that could result from the disclosure of the information or document. (4) A minister, in a statement under paragraph (3), shall indicate whether the harm to the public interest that could result from the disclosure of the information or document to the committee could result only from the publication of the information or document by the committee, or could result, equally or in part, from the disclosure of the information or document to the committee as in camera evidence. (5) If, after considering a statement by a minister provided under paragraph (3), the committee concludes that the statement does not sufficiently justify the withholding of the information or document from the committee, the committee shall report the matter to the Senate. (6) A decision by a committee not to report a matter to the Senate under paragraph (5) does not prevent a senator from raising the matter in the Senate in accordance with other procedures of the Senate. (7) A statement that information or a document is not published, or is confidential, or consists of advice to, or internal deliberations of, government, in the absence of specification of the harm to the public interest that could result from the disclosure of the information or document, is not a statement that meets the requirements of paragraph (1) or (4). (8) If a minister concludes that a statement under paragraph (3) should more appropriately be made by the head of an agency, by reason of the independence of that agency from ministerial direction or control, the minister shall inform the committee of that conclusion and the reason for that conclusion, and shall refer the matter to the head of the agency, who shall then be required to provide a statement in accordance with paragraph (3). (d) requires the Procedure Committee to review the operation of this order and report to the Senate by 20 August 2009. (13 May 2009 J.1941) (Extract, Senate Standing Orders) CHAIR: I remind committee members of their obligations under the Behaviour Code for Australian Parliamentarians to treat witnesses with dignity, courtesy, fairness and respect. Welcome. I understand that information about parliamentary privilege and the protection of witnesses has been provided to you. Thank you so very much for making yourselves available to give evidence to us today. Do you have any additional comments about the capacity in which you appear. Mr Nuch : I'm an inquiry manager and the director of the data and digital policy desk at the Productivity Commission. CHAIR: I'd now like to invite you to make a short opening statement, after which we will proceed to questions. Dr King : Thank you, Chair and the committee, for the opportunity to appear. As I said, I'm a commissioner at the Productivity Commission, and I was co-commissioner on the Productivity Commission's recent inquiry into how data and digital technology could increase Australia's productivity, which considered matters relating to artificial intelligence. Mr Nuch, as he has noted, is in charge of the relevant digital and data desk at the Productivity Commission. He was also recently at Jobs and Skills Australia and was one of their leads in their work on generative AI work and skills. The report of the inquiry that I led was released last year as part of the Productivity Commission's five reports on reforms to improve Australia's productivity. It's the most recent work that the Productivity Commission has done on AI. We've not done any substantive work since then around AI, and, as I'm sure all committee members are at least familiar with our work from last year, we did not think that it would be useful to give a further submission at this stage. We are very happy to appear today and hopefully be able to answer any questions that you put to us. CHAIR: I appreciate that, Dr King. I'll kick off. Further to that—and I appreciate the submission that was made to a previous inquiry last year—it would be excellent for the committee to get a reflection on any analysis done by the Productivity Commission on where AI has the greatest potential to increase our comparative advantages. Dr King : Let me start at the whole-economy level. Last year we modelled the potential impact of the current generation of AI, which is the generative AI models, on Australia's labour productivity, its multifactor productivity and its GDP over a period of about 10 years, recognising that history tends to tell us that it does take a while for business processes to be re-engineered to make the best use of new technologies such as generative AI. The headline numbers that we got from that were a 2.3 per cent increase, approximately, over the next decade in multifactor productivity, which is equivalent to about 4.3 per cent increase in labour productivity and a just-over-four per cent increase in GDP over the same period, a substantial lift, if I can put it that way, to Australia's productivity over that decade. Our approach in doing that was to break down—it's what's called a task based approach in the economics of analysing the impact of technological change. This particular one is a relatively conservative approach, and we deliberately took a conservative approach because, as the committee would be aware, there are many numbers floating around. We used an approach that is most associated with Nobel Prize winner from MIT Daron Acemoglu. That's a task based approach looking at individual tasks in jobs, looking at whether AI is able to augment those tasks, then asking if it would be economically sensible to use AI to augment those tasks and using that to say how much more productive the relevant workers would be in those jobs. That misses out on asking about what new jobs could be created from AI. It also avoids trying to speculate the degree to which generative AI would feed back into new scientific discoveries and so on. In terms of where the best opportunities and impacts are, it varies greatly across sectors. As the committee would be aware, there are a range of white-collar areas where generative AI offers great potential, such as law and such as health. One area that we looked at in some earlier work that we did was around software as a service, which has been a significant success story in Australia. The committee would be aware of companies like Canva, Atlassian and MYOB—business service providers that Australia is leading the world in. We also see that there's the potential for that same success to be mirrored in AI applications going forward. Now, that, obviously, is speculative, and we would love to see our Australian entrepreneurs and Australian companies being able to create those new opportunities for great Australian companies. CHAIR: My understanding is that there are varying degrees to which the deployment of AI can work across sectors of the economy. Is there anything that you are seeing at the moment in terms of sectors that are really embedding AI, going beyond just giving their workers Copilot or a platform and actually completely shifting the way in which they are operating via AI? Dr King : Yes. I'll pass over to Mr Nuch in a second; he might have some views on that as well. The areas in which we're obviously seeing the use of AI transforming the way everyday work is done, if I can put it that way, includes the health area. Obviously, they were very much an early adopter, particularly in areas like the analysis of scans—that goes back pre generative AI. We're seeing it in the way that GPs, for example, are able to interact with clients and are able to provide better advice for clients. There is an Australian startup that's working in that space which has been very successful. We're seeing it in law. We see it for good and bad sometimes. We know that the Fair Work Tribunal has had some comments to make about the number of AI generated applications that are put forward before it, and, perhaps, they're not quite as accurate as they should be. We're also seeing it in the work done in preparing briefs and in preparing advice for clients at law firms. That being said, Australia has a history, and we're seeing it here in the numbers available to us, of consumers taking up technology more rapidly than business. I think it's fair to say that it looks like it's happening in artificial intelligence, or generative AI, as well. Mr Nuch, would you like to just add anything there? Mr Nuch : Our work at Jobs and Skills Australia produced some, I suppose, similar results around which sectors and which industries were really implementing. At the time, the technology sector and technology firms were fast adopters. So too was the finance and insurance services industry. We undertook a few case studies as well, and, obviously, they are less generalisable. We heard from technology firms who had shifted the way they allocated their work—reducing their offshoring and bringing some tasks onshore—and, as a result, were relying more on mid- and senior-level workers and then finding ways to change their hiring practices for the junior level so they still had a pipeline. These are all very live questions for firms. In the tech sector, it seemed to be moving faster. Some of the evidence we saw, as well, was that sectors and industries vary in terms of the skills that they have on hand to adopt and to change their work processes. There are technology-specific skills that are obviously very available in the tech sector. The finance sector has a lot of business skills that allow more rapid or fuller changes to work processes. So what we really saw in terms of likely trajectories was that there would be multiple trajectories across the economy in terms of adoption. CHAIR: Thank you. Senator Cadell. Senator CADELL: So much talk has been around copyright and learning models versus generative models. On page 33 of your report, Harnessing data and digital technology, you say: Most of the economic benefits of AI in Australia will come from using AI models rather than training them. Can you go into that a bit? How did that come about? Was that in investment? Was that in use? Can we just expand upon that? Dr King : Yes, of course. That goes to the point that I was making before about the opportunities for Australia to be a world leader in what's sometimes called models as a service or applications built on foundation models. As the committee would be aware, there are a range of foundation models that have been developed around the world, mainly in the US, although we have Mistral in France and we have a number in China. There are, as I understand it, more bespoke models based on specific languages in Japan and South Korea. Australia obviously has an advantage in being a majority English language country and being able to access the reliable, if I can put it that way, generative AI models, including the frontier models coming out of the US. Australia could—subject to certain legal changes, which I'm sure the committee is aware of or we can discuss—develop its own foundation model here. It would be joining quite a crowded space. It would also be starting a long way behind the pack. As some American companies such as Google have found, it can be pretty hard to catch up from a standing start. We are already seeing the success of Australian companies in being able to build those applications on top of AI—I don't want to mention any brand names for obvious reasons—in health. We're certainly seeing that in law. We're seeing that with some very well-known applications. As Mr Nuch said, we're seeing that in the finance field, where our banks and finance companies are developing AI applications. Hopefully they will be able to onsell those around the world. I think we're seeing that innovation. It's early days yet, but I think Australia has shown by its history that it's got the ability to excel there. Senator CADELL: Given that you think the benefit is more likely to come from the generative rather than the frontier models and this question of copyright—you mentioned the legal questions there. I've asked this question of most people I've interviewed. Section 43B of the Copyright Act says it permits a temporary reproduction made 'as a necessary incidental part of a technical process'. Did you look at the legal impediments there? I know, in a separate part I'll go through, you identified that category 3 material would 'impact the commercial feasibility of training certain AI models in Australia'. That was also in your report. How deep did you go into that legal model? Have you formed a view on section 43B and its impact? Does it need to be changed for the generative models we currently have? Do you think it's okay, or didn't you touch it? Dr King : Fortunately, as well as having economists, we also had some lawyers on the inquiry. In the absence of legal precedent, lawyers, being fairly cautious by nature— Senator CADELL: 'Economists in wigs' I think they're called! Dr King : put forward the view that there is certainly an arguable case, which has not yet been tested before the Australian courts, that the temporary copying or reproducing of Australian copyright material in this country for the purposes of training a large language model would not fall under the current fair dealing exemptions in this country and quite clearly not do that. Certainly the view of the Productivity Commission was that it is most likely that copyright law would in fact rule out the training of those models in this country, unless the party training those models licensed the relevant material from the copyright holders. Senator CADELL: You go to category 3 and category 2 in your report, where category 3 website holders who don't hold good copyright use themselves can be still used for training. So technically the models and the use wouldn't be breaching the use terms of that website, but someone is already. It's really like receiving stolen goods. I would imagine that would be the legal precedent. Dr King : Again, these things have to be tested before the courts. Because the training hasn't occurred in Australia, we don't have those test cases occurring. In contrast, obviously the US has different laws around what they call fair use. I understand there are more than a hundred cases going on in the US currently, testing whether or not the training that has occurred of large language models in the US actually is legal under that fair use exemption that they've got in the US law. We have different law, but any law would have to be decided by the court both overseas and here in Australia. Senator CADELL: In my mind, I go back to a very long time ago when I was doing this. What was the law of nemo dat quod non habet, meaning you cannot get a better title than you receive? That a category 3 holder doesn't have the rights to pass those titles on is an interesting one. In the report you touched very briefly on energy. I think the two things you say are that a high energy demand may cause environmental concerns which AI might worsen and it may increase the cost of electricity. In your report, did you look at the opportunity cost? We've just spent $2.5 billion bailing out Tomago Aluminium, where I am, that uses 950 megawatts. The plant in Queensland has a peak use of about 2.13 gigawatts. In your calculation model, did you look at the opportunity cost of energy into other sectors, other manufacturing? Dr King : The answer is no, because our modelling is not based on AI models necessarily being set up in data centres in Australia. Of course, large language models, as they are, by far the majority now can be accessed from Australia using overseas data centres. So, no, we didn't look at that. Any negative impact or positive impact of data centres is not included in our numbers. Senator CADELL: Thank you very much. CHAIR: Ms Chaney. Ms CHANEY: I'm interested in what government needs to do to enable those opportunities—whether there's a risk that we spread ourselves too thin trying to win at all layers and what enabling infrastructure or moves the government would need to take. The BCA has recommended a national AI emissions policy: a targeted program to identify and prioritise specific areas where Australia can lead. What do you think about the need for intervention there or guidance or leadership versus trusting the market that innovation will spring up in its natural places? Dr King : There is a role for foundational legal and institutional infrastructure to make sure that Australia can seize the AI opportunity. Much of that will be around business, entrepreneurs and AI users being sure of what they can do and what they can't do. Part of our work last year was to look at that legal framework and to ask if that means we need some sort of separate AI laws or AI act, or should we be basing that certainty on our existing laws. After doing our inquiries, we came down very strongly that the best approach is not to try and predict in the abstract what an AI can or cannot, or should or should not, be allowed to do under a separate AI act but to carry out what we think is the necessary precursor, which is a gap analysis of all our regulations and laws to make sure that they are AI ready. The reason for that is really twofold. Firstly, most of the harms that were put to us that would occur due to AI were related to or were existing harms. As I said at the time, AI will allow bad actors to do bad things faster, quicker, cheaper and more convincingly than ever before, but they're the same bad things that are often covered by existing regulations. That's the first reason. We've already got excellent regulators in this country. We've got excellent rules in this country. Let's make sure they're AI fit. The second part is that, to create business certainty, it's much better to build on the existing laws where we have precedents and understanding within the legal and business community about the obligations than try to create a new set of obligations which would necessarily end up being tested in court and would create a period of significant uncertainty. Certainly the EU experience with their EU AI Act has shown the potential risk of that uncertainty and the dampening effect that it may have on the ability of a country to use AI applications. Ms CHANEY: I'm wondering, given that this was done last year and how quickly things are moving, how much these conclusions took into account agentic AI and the perhaps novel risks presented by that specifically in relation to the liability issue with determining who is liable if someone gives AI a general instruction and it goes off and does something specific that you hadn't contemplated. Dr King : That's, again, an excellent question. The answer is it was considered as part of the analysis that we did. Agentic AI was certainly being developed, on the cards and known about before we finished our inquiry. The issues of liability and the standard of liability are exactly the sorts of issues we think should be carried out through this gap analysis. As a simple example, there have been both internationally and in Australia some very unfortunate and undesirable AI-hacking incidents. There are issues certainly within the United States, as I understand, of an inability to take action against a particular AI firm. My understanding of the US law is that it requires intent and they did not intend their AI to escape whilst being tested and to go and hack into another company. We already have laws in Australia around hacking, misuse of data, the incorrect use of data and the incorrect seizure of data. We need to make sure that those rules are fit for agentic AI, including the liability. So it should not be the case that—I was about to give my opinion rather than the PC's. Apologies. The question that needs to be asked is: should there be a liability, such as strict liability, so that if you're using an agentic AI tool and it goes out and does something bad, you can't say, 'Oh, that was an accident, sorry'? Should it be strict liability? Should it be negligence? For example, did you not take reasonable care? Or is intent the right level? That is going to differ on the specific laws that we've currently got in Australia, but that needs to be checked. Ms CHANEY: I'm interested in understanding this. You didn't do an analysis of the economic benefits of training AI models, but, looking at the potential economic upside for Australia, how dependent do you think that benefit is on accessing frontier models as opposed to not-quite-frontier models? In order to achieve the economic benefits, is it important that we have access to the bleeding edge or will most of the economic benefits come from less novel models? Dr King : The Productivity Commission did not explicitly look at that question. If you're happy to take my personal opinion rather than the PCs work— Ms CHANEY: Sure. Dr King : My personal opinion is that, at the moment, that is an open question. There are certainly experts out there who are saying that, particularly for those applications built on top of AI engines or frontier models, there is a large range of models that aren't the absolute frontier models like the Anthropic model or the OpenAI model that are more than adequate for the vast majority of uses of AI. At the same time, I know there are experts out there who say, 'No, the frontier models are critically what we need.' I suspect that it depends very much on the application. If you are solving cutting-edge, unsolved mathematical problems, as we've recently seen, then frontier models are what you need. If you're building applications that are likely to be used in everyday business, I suspect that there are a large range of models that aren't quite as good as the frontier models but are more than adequate for the task. Ms CHANEY: Thank you. CHAIR: Senator Pocock? Senator DAVID POCOCK: Thank you for your time. I was really interested in recommendation 2.1 in the Productivity Commission's December 2025 final report, which recommended that the government monitor licensing markets, creative incomes and overseas court decisions for three years before making any changes to copyright law. I'm interested in this approach and why you think that's the best course of action for Australia. Dr King : As the committee is probably aware—because there was a fair bit of publicity around it at the time—we floated in our interim report around July last year potential changes to copyright law, particularly a text and data mining exemption that may make clearer the legal situation around training large language models in Australia. When we followed up on that and did further analysis—particularly talking to the tech companies, the publishers and the creatives themselves—we found a number of things. Firstly, whilst it might sound like a silver bullet to enable training of large language models in Australia, any change to copyright law—such as a text and data mining exemption—would be challenged in court straightaway. We would very likely be in a situation, hopefully not as bad in terms of number of cases as in the US, that would require court cases to determine exactly where the boundaries are. That would involve appeals, and as the committee is probably aware that can take a number of years. It became very clear that a broader change to the copyright law like a text and data mining exemption was unlikely to be an immediate fix. It also became fairly clear that, whilst that is one of the barriers to training large language models in Australia, it is not the only barrier. Obviously, having the data centre capacity, having the skills—and there is a skill shortage in Australia. Other barriers would be relevant. The second area was the impact on the creatives themselves and the fact that they have currently got a property right set under copyright law. There is a reasonable case that that law has been put in place and has existed for many years for good reason: to encourage creative work in Australia, just as it is designed to encourage work overseas in similar jurisdictions. There is a cost of changing copyright law, which may disincentivize creatives. At the moment, the degree to which AI is either reducing or improving creatives' opportunities and potentially their income is unclear. Many creatives use AI as a tool now in their creative endeavours, but obviously they're concerned when they compete directly against AI produced products. The music industry is one of the areas where this has been most prominent. Musicians are using AI to help compose and produce content, but AI based content is also flooding streaming services. That is all uncertain. The third point was that we are already seeing licensing solutions develop, and that makes sense. If you think about the applications space as opposed to the training space, you want your application to be extremely reliable. You do not want a health application, for example, that occasionally hallucinates or is trained on anything except the best possible quality input. We have already seen, for example, at an international level, licensing agreements with the New England Journal of Medicine so that the AI applications can access high-quality academic research to be able to train for that application. Again, seeing those licensing markets start to develop in Australia, we felt there was no need to move too rapidly but that it was better to let those licensing markets develop. That will reward the creatives. That being said, we felt that there needed to be a period of time before review, so we said three years—let the markets develop. If they haven't developed, if there's still a significant problem, then we should start considering alternative solutions. Senator DAVID POCOCK: Just looking at your report, you said that AI could add more than $116 billion to Australian economic activity over the following decade. I'm interested if you've looked at or you have thoughts on who benefits from that. We've heard from unions who say, when there are productivity gains, it's the company—the multinational in that case—that picks up those gains in profitability. We know, in this space, that a lot of these AI providers are multinationals and very unlikely to pay tax here, so who benefits from that, and do you have a view on how we can ensure that we're looking at retained economic benefit, not just, yes, someone's making money but they're not Australians? Dr King : I'll let Mr Nuch talk from the employment and the jobs perspective. Again, from the Productivity Commission's perspective, looking at the potential developments of AI and the uses and companies based on AI in Australia, then we would expect to see those application companies, which would be, often, new Australian companies, being able to create value and create new jobs. We would also see, obviously, companies who adopt AI change their business processes to make the best use of AI. They will be able to gain benefits from that subject too. Of course, the AI companies, abiding by our competition laws—if our companies, our businesses, are forced to pay anti-competitive prices to US AI companies, then, clearly, that's going to limit the benefit. I can say, as a former Australian competition commissioner, I think we have [inaudible] on the job side and the employment side. Mr Nuch : As I mentioned before, I speak mainly about my previous work at Jobs and Skills Australia where we looked at the exposure at a task level of gen AI on the Australian labour market. You may have heard previously from others that the research found AI and its use weighs more towards the augmentation of tasks and jobs, but it also weighs more towards higher skilled jobs, to towards knowledge work and to repetitive and routine tasks. Does this necessarily mean that workers in those occupations would win or lose, as such? It's unclear because, if the technology is used—this is going very micro into the task level—is it the case that they have an hour or some other unit of time freed in their day that's applied to other work or that's applied to leisure? There are dynamics within the workplace, within the labour market, that will be relevant and not necessarily uniform either. In some cases, firms redeploy either time or workers into other uses and in other cases not. I think it's an important question that we tried to look at with fresh evidence in that study. I will also point out that there are occupations that are less exposed to gen AI, including many occupations that are undertaken in regional Australia especially. In some ways you might think that there's less disruption to think about in the future, but potentially there's also less scope for productivity gains in those lines of work. I think there's still a lot to consider on the employment side in terms of who might win or lose. Senator DAVID POCOCK: Thank you. CHAIR: I believe there are no more questions from committee members, so we'll conclude this session here. Thank you both so very much for making the time this evening to come and give evidence to our inquiry. We appreciate that. If you've been asked to provide any additional information or anything on notice, please work through the secretariat to do so in the next two weeks. .