SADLER, Mr Greg, Chief Executive Officer, Good Ancestors [14:23] CHAIR: Welcome. I understand that information about parliamentary privilege and protection of witnesses giving evidence to this inquiry has been provided to you. We now invite you to make a brief opening statement—as brief as you can to give plenty of time for questions as I know there are many that the committee members have. Over to you, Mr Sadler. We'll kick it off from there. Mr Sadler : Good Ancestors has been working on AI safety in Australia for over four years. Over that time, progress has been rapid and is accelerating. The computing power used to train frontier AI models has grown fivefold per year every year since 2020; that's 600 times more than when we started working on this topic four years ago. Each year, better training methods let developers reach the same level of performance with about a third of the compute they needed the year before. New ways of making AI more capable have also emerged: test-time compute, where a model thinks before it answers, and agent swarms, where thousands of agents cooperate on a task. These factors are exponential and are driving dramatic leaps in AI performance. AI is not slowing down; it's speeding up. Two years ago, AI couldn't count the number of 'r's in 'strawberry'. Last month it solved a math problem that had been stumping the world's best mathematicians for 100 years. Over that time, we've repeatedly seen risks that were dismissed as speculative become reality. Neither the National AI Plan nor the Prime Minister's speech mentioned loss of control, even though the speech came after the Medicare and Hugging Face hacks. The most ambitious AI activities Australia conceives of are out of date before they start. Australia needs to get a say on its AI future. Our submission provides 10 ideas, and no doubt there are many more. The most achievable is a properly funded AI Safety Institute. We recommend $50 million per year, plus compute and pre-release access to frontier models. The most impactful way to get a say on AI future is training AI in Australia, both in Australian based companies and in partnership with leading global companies. AI training is a powerful fit for Australia. Every country will run AI models, but only a few will build them. Training makes the most of our national strengths, including abundant space and clean energy. Unlocking training in Australia will help the domestic AI industry grow and deepen our relationship with the countries and companies that are leading in AI. Making many valuable contributions is not about leverage to use against others; it's about becoming a partner that others want to work with and have reasons to respect. Copyright never conceived of a general purpose technology trained on the sum of human knowledge. Today, every frontier and foundational model is built in a country that has a pathway on copyright. We consulted on a permit-and-fund model that would allow AI training in Australia with strings attached. The permit would raise funds and secure opt-out. The fund would put hundreds of millions of dollars per year into the pockets of Australian creators. If we don't act, Australian creators will continue to have their works used overseas without compensation, and Australia won't get the benefits of building AI here. Once we have a say on the future of AI, we need to use that say to protect Australia's national interests. We need accident and incident reporting; embedded third-party safety evaluators; public interest compute; a measurement of AI compute, supply and demand in the country, along with a target that gets us out of compute deficit; and other commonsense policies. We also need to break the cycle of acting on risks only after they happen or acting at a scale that is far too small relative to the challenge. Australia is currently scrambling to deal with AI's cyber capabilities. We can't afford to be in the same position for AI's bio capabilities. Biosecurity risk is something that both Kate Chaney and Tim Watts have talked about on the public record, and that demands urgent action. Overall, humanity is racing into an uncertain future. AI has already surpassed human performance in many domains. AI companies are pursuing recursive self-improvement. The US president has said, 'Whoever wins AI wins.' Our recommendation is that Australia urgently take a stake in the AI future and use that stake to protect our national interests. I'm happy to take questions. CHAIR: Thank you very much. To your point that you make around the opportunities for AI training capability here: if that could occur in Australia at scale, what do you see as the benefits that could potentially fall from a talent and safety perspective, and what other considerations should be made? Mr Sadler : As I said in my opening, this is about the valuable contribution that Australia can make, not about getting leverage to use against others. I agree with what ASD said about influence following contribution, not consumption. We need to think as a country about what it is that we can contribute. One way to think about this is that training is more like a fuel refinery, whereas inference is more like a petrol station—so it has that spot further up the value chain. Many countries, regions or jurisdictions will expect to have petrol stations, but refineries will be rarer. Hosting AI training gives us three different things. One thing is something concrete to apply conditions to. I've talked about a fee that could be placed on frontier training just to raise funds. It's a hook for frontier access, and we can get things like accident and incident reporting and transparency reporting—so these are the safety requirements. There's this national security imperative; this is particularly the case in cybersecurity, where you need to have a frontier model to defend against an attack that's been run by a frontier model but also perhaps in other domains. Then there are the economic benefits. AI has the potential to be an enormously extractive technology where tasks that were done by Australians in Australia are done by machines in other countries. Economic experts say, 'AI might lead to job losses, but it will create new jobs.' That has to be proven. There is a way where a machine that thinks is a bit different to a machine that does. But, even if it is true, it's not guaranteed that those new jobs are going to be created in Australia. We need to be thinking, 'What are these new industries and how can we get them here?' Maybe data centres is the first wave of that new AI created industry. If we turn our noses up at AI training, data centres and dark factories, or whatever that next wave becomes, we fall further behind on that curve. CHAIR: You also mentioned in your opening remarks the issue around copyright and the challenge we have in front of us currently. The evidence we have heard is that there are some significant challenges there. My sense is that, as you said, training is happening on works, and Australian creatives are not able to control or get compensated for that. Do you have further reflections on that particular challenge? It's important that we make sure we are protecting the creative industry here in Australia, in particular the creative workers, and that they have more control than they do currently and that they have the ability to be compensated for their work. How do we make sure we can secure, essentially, a better outcome than we have at the moment, noting the huge upsides, potentially, that we have for the ability to train here? How do we thread that needle, in your view? Mr Sadler : We got involved in this conversation because we want these safety benefits; that's the primary thing that drew us into the conversation. Australia doesn't have a seat at the table currently. The more we're doing in the ecosystem, the more people like you can be empowered on the global stage to have influence on those safety topics. But we aren't the only stakeholder group that thinks the status quo is bad. In fact, I put to the committee that every stakeholder group is currently in the worst possible position. If you take the creators, as I was just saying, Australia's information is being used overseas in most jurisdictions where there isn't even opt-out, and there's no compensation in places where frontier and foundation models are being trained. This is also the case for the national security audiences we were talking about before. It's a problem for small AI businesses in Australia. Currently our industry is being squeezed into these small lanes around narrow language models or areas of national expertise where we have data that's well curated and available, or areas like scaffolds and harnesses. If we had a pathway in copyright, then perhaps there would be many more ways that the Australian domestic industry could flourish. It's also a challenge for big players. Everyone is having the worst possible situation at the moment. Perhaps this was a naive thing, but we thought that the zone of possible agreement here—to use negotiator speak—was enormously large: everyone is in the worst possible position, so anything that we could propose would be better. We prepared this proposal as a consultation paper. We were hearing that lots of people were upset with the status quo, but we weren't seeing many people have a bid about what 'better' could look like. I saw that Minderoo, in their submission, have put forward some options about what 'better' would look like. That's the reason we put out this consultation paper. We called it a consultation paper because we wanted to hear about other views; could we narrow in on what 'best' would look like rather than just what 'better' would look like? Let me talk about how that model works, to help the committee. The first thing we have is a permit. The idea here is that an AI company, whether that's a multinational or Australian business or a start-up, would gain legal certainty in exchange for a fee. That fee would be on top of opt-out. This is about getting some of that financial benefit into consolidated revenue. We think that that fee could be scalable. If it's a very large AI company, then we could make it a very large fee, whereas, if it is the kind of small Australian startup that we're trying to foster, we could have a scale of fee that becomes a lower amount. It would require opt-out. For someone to obtain this permit, they'd have to be offering a best-practice opt-out to Australian creators. And it would provide a hook for these national-interest conditions that I've been talking about—things like frontier-model access or the shopping list of conditions that we've been talking about. To be super clear, the permit isn't intended—our proposal does not allow it to have piracy. I think everyone agrees that piracy is bad. The permit is not letting anyone do piracy. It's not letting you bypass technological protection measures, TPMs. All of that information that's behind paywalls, in walled gardens or behind opt-out, which is where we're seeing those valuable deals in the US, would still be on the table for that kind of material—those valuable deals in news archives and publishers. Then, on the other hand, we would have this fund. The fund would be modelled on the public lending right. Australia has had this public lending right since at least 1974. In fact, this parliament has just published a new version of the public lending right. If I may brief you on how this model works, the minister appoints a committee. In the case of the public lending right, that committee is made up of librarians, publishers and authors. In the case of our scheme—and I'm happy to give you, on notice, a more detailed paper that sets out how we think this could work—it would have the kinds of coders and mathematicians and creators of other kinds who have their information inside our AI training datasets as well as AI experts who can help with the understanding of how our datasets work. That committee would be tasked with producing a scheme, just like the Public Lending Right Act. That scheme would decide how money is to be distributed to people and what the best mechanisms are for achieving fairness and giving people recognition of the important value that they're creating through their work. A key strength of the model is that it allows the money to stay here. Our assessment is that something like 99 per cent of copyright protected works globally are not owned inside Australia. So, even if we were to find a pathway on licensing, it would result in some 99 per cent of the money that we bend over backwards in Australia to raise going to foreign rights holders. To me, it would seem unusual if this country were to work so hard to come up with some new pathway and have it result in money going to the US, China, Japan, Israel, Singapore or Europe when those countries aren't working equally hard to come up with pathways that would send money back to Australia. The overall idea here is that everyone would be better off. The creators would be getting the opt-out that I think that they deserve. They would be getting payment from this fund. They would have seats on the committee to decide how that fund would operate. We would get training in Australia, and we would get the right to attach strings to that training that give us the financial, national security and national interest benefits we're after. CHAIR: I appreciate that. Deputy Chair. Senator DUNIAM: Thank you for your evidence today. First, in your opening statement—and, I believe, in your submission as well—you spoke about the things that we should get. You talked about—was it $50 million a year into the AI Safety Institute? Is that correct? Mr Sadler : That's correct. Senator DUNIAM: And you talked about access to compute and early access to models. Early access to models makes absolute sense, and everyone's talking about it. In terms of the amount of $50 million per annum, how did you determine that figure as being appropriate? Mr Sadler : For clarity, the $50 million-per-annum figure relates to the Australian AI Safety Institute, not to the fee that we would propose for the fund. Let me talk to the AI Safety Institute, but I'm happy to come back to the proposed permit fee, if that's what you want to focus on. Senator DUNIAM: On the quantum of the funding, as I heard it, you're saying that's the amount of money that should be invested in the AI Safety Institute—yes? Mr Sadler : Yes. Let me talk about the AI Safety Institute. I think the AI Safety Institute, which has been up and running for several months now, is an excellent initiative by the government. It's attracted really impressive talent and drawn a number of Australians back to Australia from overseas to work at the AI Safety Institute. They've done really excellent early work. I think Bill Simpson-Young from Gradient Institute was talking about some of their early work around the interaction between AI agents, which is exactly the fault that we saw in things like the Hugging Face incident, so it was very prescient of them to be working on that topic. Currently, the AI Safety Institute is funded at $7.5 million a year. That's an amount that declines over the forward estimates. I don't think anyone thinks that AI safety is improving, so this is an issue. We suggest the amount of $50 million per year. The way we've generated that number is basically through looking at the funding that the UK has and then adjusting it to Australia on a per-capita or per-GDP basis. Essentially, we're saying we should be doing the same as what the UK is doing, because the UK has this leading institute. Another point is that the remit of the Australian AI Safety Institute is, by far, the broadest in the world. They have this task to test these frontier models. They have this task with international engagement, so they're trying to participate through the network of international AI safety institutes to build norms and standards. They're also charged with supporting the 100-plus Commonwealth bodies that are involved in AI. This is a tremendously large scope for a tremendously small amount of funding. The final thing I'd say about the AI Safety Institute is—in our submission and in my opening statement, I talked about Australia trying to have many valuable spots in the AI ecosystem. I think training is one of those, but I really don't want us to zoom in on that too much. I think there are, really, very many options. An AI safety institute that has a valuable research agenda that other countries and other companies want to know about and want to participate in is a reason that countries might give us things like frontier access. They want us to have that access to do the important science that we do at the AI Safety Institute so that we can share that research with others. Senator DUNIAM: Okay. I was wondering how you determined that figure. You've outlined that in your answer, so thank you very much. In terms of access to compute, do you have any guidance for the committee on what access to compute should look like? Are we talking about a percentage of any quantum of compute that is offered by a particular operator? What do you have in mind? Mr Sadler : Again, I'm happy to table, on notice, the document that we're currently preparing for the Office of AI's consultation, in which we're doing more detailed work on exactly that kind of question. What I'd say on compute is that there are maybe two separate things to think about here. One idea in reserved compute is this gas reserve idea. Should there be some proportion of compute that we earmark to be sold on the Australian market first? We're heading towards a 'compute crunch'. I argue that Australia has a compute deficit. The way we did that deficit calculation was by trying to look at all of the data centres in Australia that have AI compute in them and all of the demand in Australia for AI compute currently. Our calculation was that Australia currently has maybe 40 per cent to 50 per cent—certainly less than half—of the compute here that we're actually demanding. We're a net importer of compute. And, even if you fast-forward through the announced data centre builds and forecasts of demand, we actually fall further behind. I think this goes to the point about social licence for data centres. People might be asking, 'Why do we have all of these things?' If government were able to measure how much Australians are using AI and how much AI supply Australia has, we could illustrate that we actually are in deficit and falling behind. That might be one factor that would improve social licence. What I'm pointing to here is that we're in this potential compute deficit, and it would be very bad if Australia had a bunch of these data centres that hosted AI compute but Australians struggled to access AI compute. So the first thing is to earmark sale in the Australian market first so that Australians aren't being left out. The second thing is more about gifted or donated compute. Should there be an obligation to give a certain amount of compute to Australia? I think that is a plausible policy to look into. It has more risks and downsides than having the first opportunity to buy, but it still is a valid thing to have on the shopping list. My final thought is that there are very many things that we could be asking about accident and incident reporting, transparency reporting, frontier model access and compute skills. I think the committee can develop a shopping list of what all of these things might be. But it might not be wise to buy everything on the shopping list, because we could end up with a very complicated and baroque policy where we ask people coming to build data centres in Australia to do a shopping list of a hundred things. Maybe we should be a little bit discerning on what it is that we most want to buy from that shopping list. Senator DUNIAM: My final question relates to copyright. You talked about the intractable nature, or so it seems, of the problem. Are you aware of the proposals that were publicly reported on, as tabled in the Senate, as proposals being considered as a means of resolving the AI situation to enable training? If you're aware of them, do you have a view? Then, finally, would those proposals apply to content harvested, scraped and used for training here in Australia if the content is generated overseas? Mr Sadler : I did see those tabled documents. To be honest, I found them slightly inscrutable without broader context. There were some coloured PDFs, and I didn't really understand what they meant without having someone from the Attorney-General's Department to explain them to me, so I might not chance my arm on an opinion on documents that I didn't fully understand. On the second question, I think whatever approach Australia takes, whether it's something that the Attorney-General's Department proposed or something that I'm proposing or, perhaps, what Minderoo says, it has to cover all training data, globally, because a lot of the capability of these frontier large language models comes from the scale of the dataset. I'm happy to talk about challenges that come up in the margins of licensing. I think Microsoft made a very compelling point about spam emails. They were explaining that a spam email is protected by copyright, so the Nigerian prince who wrote that email owns the copyright in that email. Training on that email is useful because we need large language models that can identify spam emails. We need to find some way to get that into the training dataset, but I don't think we want Australian startups negotiating with Nigerian princes to license their spam emails. I don't think that's a useful path. So we need to find a way to get it in because it's important to our training. Senator DUNIAM: Thank you. CHAIR: I believe you've got some questions, Ms Chaney. Ms CHANEY: Thanks, Greg. On biosecurity risk, can you talk through what's changing in biosecurity, what needs to be done and how hard it is to do? Mr Sadler : I heard the comments by Mr Whittle from Global Shield yesterday. I agree with a tremendous amount, maybe everything, that he said to the committee. I think that's very helpful. Just to hit a few highlights there, I worry that Australia is increasingly seeing AI risk as being synonymous with cyber risk. This is definitely one area of the jagged frontier of AI where we're seeing these risks now, but AI is also pushing boundaries in physics and maths and biology and, perhaps, very many other domains. We definitely need to not have tunnel vision on cyber as the main or only thing to be worrying about. The particular concern I have here is not that AI could coach anyone through the process of building a bioweapon but that it could coach people who have relevant undergraduate degrees or people who have worked as vets or pharmacists or in pathology or in other areas where they have relevant skills. The evidence for this possibility is mounting. For years now, we've heard the AI companies saying, 'We think AI models are on the cusp of being able to help novices build bioweapons, so we're going to treat the AI model as if it could help a novice build a bioweapon'—sort of flirting with this line. I think Mr Whittle referred to a study where AI models exceeded 98 per cent of domain experts at virology protocol troubleshooting. If you expect the vial to be cloudy, but instead it's clear, what's gone wrong? The AI model exceeds human experts on that. Another study that jumped out to me from the International AI Safety Report was where some experts surveyed hundreds of biospecific models, which are not large language models, and found that only three per cent of the models they surveyed had any safeguards at all. The bar is really low here. Again, we're seeing this translating from the theoretical to the actual already. I think Anthropic published five misuse cases, including attempts at gain-of-function in viruses and adapting bird flu for mammals and including the users bypassing Anthropic safeguards to do this kind of work. So there is a real problem here. And it's definitely not a problem where we can wait to see what happens. 'Wait and see' didn't work in cybersecurity, and it's not going to work in any area of AI policy, including biosecurity. The most practical action is screening synthetic DNA imports. Australia already has the world's best biosecurity regime. Other countries are scrambling to pass new laws to deal with this risk; we could act on this risk today. And that's a little bit about dealing with the risk domestically, but I think it also should be about sending a global message, right? We can foresee this risk; we have the world's best biosecurity regime. Let's start building a global norm. This is something we do via the Australia Group. And it's also international market pressure. Australia has a big market for ordering this kind of material. If we can help build this global norm, then it's just easier for labs to screen, because more companies are requiring it, and then we can start building global safety. Ms CHANEY: Thanks very much for that. One of your recommendations is 'Ensure accountability for harms caused by AI agents'. Can you just talk a little bit about that liability question, why it's important and what we can do about it. Mr Sadler : Sure. We did an expert study—it feels like several years ago now, but, given the space of AI, it was probably only last year!—where we asked experts about the adequacy of Australian law for AI agents exceeding their authority or AI agents taking actions that no human intended, and well over 95 per cent of the experts we surveyed said Australian law wasn't ready for this. The details are in our submission. The key problem here is one of actus reus and mens rea, for the lawyers in the room. A crime requires a thing to have occurred in the world and it requires a person to have a fault element or a mental state that accompanies that. So, obviously, intent is challenged with AI agents, because an agent can go and do something that a person didn't intend to happen. We also have other fault elements like negligence or recklessness, but it's just less clear in an AI context, where we don't have clearly recognised global standards or global norm development. To the extent that there is a global practice around how AI is developed and used, it's not a very rigorous and safety focused culture that has developed. So I think, from a criminal law perspective, it would be very hard for a prosecutor to show that someone's fallen far below the standard of the average AI user, when there is no standard for the average AI user, and that standard seems to be very low in any case. So the first answer to my question would be that we should contribute to global standard development around what appropriate use and development of AI looks like, so that we have that touchstone that could be in a criminal prosecution. But my final point would be: criminal law and criminal fault is not the best way to build a safety culture. If something goes wrong in an airline, we don't send in the police to arrest the pilots. That's not the first move that we make. It's a possible move, if people have been grossly negligent, but the first thing we do is send in impartial investigators. They investigate the problem. They find out whether the fault was from the airline, from the manufacturer or from the pilot. Did people follow standards? And that's really where we need to start with these AI agent faults: not thinking criminal law first. Ms CHANEY: So, under our laws as they stand, if someone instructs an AI to do something, and the AI then acts agentically and goes and does something else that was not intended by that person but breaches a law, it's not clear under Australian law whether the individual is at fault, whether the developer of the AI tool is at fault or whether, in fact, anyone is at fault. Is that right? Mr Sadler : Yes, that's right. Imagine you're driving along and the wheel falls off your car. Under current law, it would be very easy to say, well, is this the fault of the car manufacturer—has that wheel been poorly designed? Or is it the fault of your mechanic—did the mechanic not screw the wheel on properly after they changed the tyre? Or is it your fault—were you driving erratically? Were you speeding? Or was it sort of an act of God—was your car hit by lightning, and no-one could ever have foreseen that the wheel was going to fall off? We have clear engineering standards that let us investigate and find out whose fault it was. But we have none of that for artificial intelligence, which would make the task of a prosecutor enormously difficult. Ms CHANEY: Thank you very much. Just a last quick one on critical incident monitoring—what should that look like? Mr Sadler : This is absolutely low-hanging fruit. We have this for aircrafts, for medicine and for consumer products. We have international examples in California and New York and the EU and Korea and Vietnam where this is required. But all of these jurisdictions have these accident/incident reports as confidential, which means we don't know what they are, because they're being given to those jurisdictions. The absolute starting point here is that we should simply have this requirement and we should simply start by saying, 'Whatever you're giving to other jurisdictions at whatever threshold you're giving it to them, we should have that as well.' But I do think there's a need for Australia to go further here—firstly, in two ways. I think near misses have to matter. We shouldn't wait for people to die before we conduct an investigation to find out what went wrong. We can wait for something to have gone wrong, a near miss, and then investigate without having that serious harm. That's the first way I think we should go further. The second is internal models. Under the Consumer Law, we can say, 'If a product is sold in Australia, here's the standard it has to meet,' but it's very hard for us to look up that value chain to 'what about the factory where that product was made?' For the Hugging Face incident and for the Medicare incident, they were prerelease models. It wasn't a product available in the market in Australia that caused that harm. So having a way to look into the internal model development is essential for effective accident/incident reporting, as the most dangerous models are held inside the companies; they're not available to the public. It comes back to why AI training is important. If the model's built in Australia, it's much easier to say our law should look at that construction, whereas, if it's purely imported, it's a harder legal threshold to meet. CHAIR: I'm conscious of the time. I know Senator Pocock and Mr Watts are both keen to ask questions, so we'll try and be as concise as we can. Mr WATTS: I do think the biorisk is real, and, as Ms Chaney's pointing out, there's also talk about the use of Australia group and our role in convening that group and the contributions we can make. But we should be clear that Claude was not being used for gain-of-function research. That security report suggested that it was being used to help write a grant application for academic research into gain-of-function research. Just when I tried to fact-check you using Claude now, the guardrails kicked in and kicked me down from Opus 5.5 to Opus 5, so you should probably note that for the record. Can I ask two very brief questions. Firstly, you've given three reasons why labs won't license—this is in a copyright context. One is that 'formally licensing publicly available copyrighted material could undermine favourable court rulings in other jurisdictions, particularly the United States'. Has anyone quantified the kind of risk that they would face from undermining those decisions, and how confident should the committee be that this is a real red line on the part of the frontier labs, rather than a negotiating position? That's question 1. This is question 2: do the frontier labs have realistic alternatives for where to locate training compute in the future? Mr Sadler : Let's hit the court cases first. I'm not aware of any legal prospects advice that suggests how risky licensing in another jurisdiction would be as a precedent in the US case. I'm not aware of that existing. I think it's worth thinking about the stakes that are at play, though, in the US courts—and this is, I think, an important background for the copyright conversation in general, both from the perspective of creators and from the perspective of AI companies. Obviously, the vast majority of these frontier AIs are being trained in the US, and they've been trained in the US for a long time, and most of the infrastructure is there. I think something like 75 per cent of global AI compute is in the US, so that is where most training will happen into the future. The US courts have been very clear that piracy is unacceptable, to the extent that creators have been having well-deserved wins in the US courts. That has been saying that piracy had occurred and that is not acceptable, so that's good and proper. And then there is a question about the use. Once you have legally obtained material, how can you use it? I think the concern from the AI companies— Mr WATTS: The distinction we're making now, so everyone is clear, is that, if you download a batch of pirated books from the internet and ingest them for frontier training, it is prohibited. It's a copyright breach. But, if you legally procure a number of books—if you buy them second-hand et cetera—and ingest them for training, fair use applies. Mr Sadler : That's correct, yes. Thank you for the clarification. I think it's a stakes conversation about both having lost that value precedent and then potentially having to pay damages for all of the training that has happened to date. There is a real and genuine nerve there for the US companies. On the point of alternatives to Australia, I think there are a few here to explore. Obviously Australia does have desirable features; I think we've talked about those at length. There are three alternatives I'll talk through. The first is Japan. Japan has a text-and-data-mining exemption, and Japan has some fallow nuclear power. I haven't actually seen a quantification of how much that is, but I think it's real. There is a space issue in Japan, but there are obviously ways of solving a space problem, whether that's land reclamation or just building the data centres taller—or various things. I'll just touch on this TDM point a little bit longer, because I've heard the argument that Japan doesn't actually have a TDM. This turns on the non-enjoyment test, which is in Japanese law. The non-enjoyment test is about the copier's purpose. If you're running data analysis that's not for your enjoyment, that's allowed under the Japanese TDM, whereas, if you're listening to music or watching a movie, that is for your enjoyment, so that's not a TDM protected activity. There was a 2024 Japanese subcommittee that was very clear that training is data analysis, and data analysis is what the TDM is intending to allow. Secondly, there was a long discussion in the House of Lords about the Japanese provision. It was, again, very clear that everyone thought that this was a TDM. So that's the Japan option. There's the Middle East option. There are copyright pathways. Construction can be cheap and fast in the Middle East, and there's abundant energy. The downside there is conflict risk, so that's on the mind of the AI companies. The third option is just tripling down or doubling down in the US. There are moratoriums in place in some US states, but it seems that that's just pushing construction to other US states rather than actually reducing construction overall. So there are viable alternatives to Australia, but they do have downsides. Mr WATTS: Thank you. CHAIR: Senator Pocock—briefly. Senator DAVID POCOCK: Thanks for your time. I've got a bunch of questions; I might put some on notice. You're obviously a strong proponent of an opt-out model. Given copyright is based on international treaties, would the whole world need to opt out if people set up here? Mr Sadler : We thought a lot about opt-out versus opt-in when we were designing this regime. Maybe this is a bit firm of me, but it doesn't strike me that Australia should bend over backwards to offer opt-out to the US or to China when the US or China doesn't offer opt-out to us. So my starting position would be that the opt-out should be for the benefit of Australian creators, but I think that there is a genuine way that policymakers could land on different positions on that. I think—I'll keep it at the short answer. I'm happy to take on notice giving more detail to that. Senator DAVID POCOCK: I haven't heard from any experts who think that opt-out is workable. I would love to hear you respond to creative industries who say you're selling us out with this fund that you have to pay into, and then the majority of that, because of international copyright obligations, will have to go overseas—maybe 10 per cent stays here in Australia. And then the opt-out puts the onus on creators rather than companies. Then there's evidence from SBS and others that these robots.txt files are just regularly ignored. Mr Sadler : Let me tackle that '10 per cent stays here' first. I think this is a very real concern. I was saying before that my estimate is something like 99 per cent of global copyrighted material is not Australian owned, so, under any licensing approach, we could expect that 99 per cent of money raised in Australia would go offshore. But I think the fund approach is the solution to that, not the problem. In the public lending right model—the public lending right is how public libraries work in Australia—the parliament says that public libraries are a public good and that we want these to happen in Australia, but we recognise that this has downsides for Australian publishers and Australian authors. The public lending right model creates a fund that pays money only to Australian authors and publishers, not to every author and publisher on the planet. And dozens of other countries also have this same approach, where they compensate their authors and their publishers for their public libraries. So I think it's perfectly possible. This has been talked about as a Berne convention issue, and it's been very clear that this sort of fund that is available only to your creators is not an issue with the Berne convention. So I think the fund is the cure to the problem with Australia bending over backwards only to send the vast majority of the fund's value offshore. In terms of the workability of opt-out, I think opt-in just replicates the same problem that you have with licensing—that is, how do you reach all of these people? I was talking about the spam email issue that Microsoft raised before. In a licensing approach, you would have to find the Nigerian prince that sent the spam email and enter a licence with them. But, in an opt-in situation, you'd have the same problem of having to find the Nigerian prince and ask them if it's okay to train your model. Imagine the perspective from an Australian startup. I heard Bill Simpson-Young talking about the need for there to be a sort of venture capital fund for Australia. I think that's a great idea and something that the committee should think about. Suppose an Australian business raises its first million dollars from this venture capital fund and downloads a curated dataset, a dataset that doesn't have any pirated material in it. Then they want to do their demonstration of safe, responsible and ethical AI built the Australian way. The first task that that organisation would have to do is figure out who in that curated dataset is the owner of each individual bit of work. So they might have to write to all of the rights-holding groups and say, 'Can you identify in this dataset which bits you own, so I can license it from you?' But then there's the spam email example we talked about or, perhaps, all of the anonymous and semi-anonymous information that's in that dataset. And then the company would have to figure out, 'Well, who's the owner of all that information so we can have them opt in?' So I think opt-out is much more workable than opt-in. Senator DAVID POCOCK: But no experts that we've heard from think that it's workable, and copyright holders certainly don't want it. In your library example, you can clearly see the public good in public libraries, but I don't see as strong a public good in selling out copyright holders and basically saying, 'Well, just pay into this fund, and maybe we'll distribute $300 million to every copyright holder in Australia for two US AI companies.' What do you say to that? I feel there is a moral part of this, too, that really speaks to Australian values. Are you really suggesting we should throw copyright away or create a way around it for two US tech companies? Mr Sadler : No, that's actually not what I'm saying. I think I'm trying to be clear that the current status quo we have is the worst possible situation for every stakeholder, including Australian creators, and I don't think opt-out is too much of an ask for the big players. I think this committee has heard evidence that 90 per cent of Australian music could be licensed with four deals. So it follows that 90 per cent— Senator DAVID POCOCK: So why don't they do that? Mr Sadler : could be opted out with four deals, or four emails. I think that the difficulty of opt-out tracks the difficulty of licensing in that sense. Your question then was: why don't they make those deals? I think this comes back to the long-tail question. If you could cover the entire landscape then licensing would be possible, but it's just not clear how you license the anonymous content online—the spam email, the blog or the internet forum. Senator DAVID POCOCK: Sure. I just feel we're in the situation where, as soon as we start talking about striking deals with copyright holders, the AI companies and you then point to some spam email from a Nigerian prince and the long tail as the reason not to strike copyright deals with 80 per cent of copyright holders. Maybe the 20 per cent is the long tail and potentially is a fund for the long tail, and we kick that down the road and say, 'Well, we'll deal with the long tail.' But the wealthiest companies in the world at the moment say that they won't or that it's too hard or that they can't afford to negotiate in good faith. It just seems really unacceptable to me. And then we've got organisations like yours pushing this. Mr Sadler : I'm not saying we should disregard the long tail—quite the opposite. Something like math and code makes up the majority of AI datasets. I cite a 2024 open model, which is 17 per cent code and 25 per cent math. These groups don't have a collecting society. So the long tail is actually the majority of what we're talking about, not the minority of it. And maybe just a final point would be— CHAIR: We do need to wrap up. Mr Sadler : We're saying: why would the AI companies license some of the material when they don't have a pathway on all of it? If you wanted to build a back deck, why— Senator DAVID POCOCK: Do they need it, then? Mr Sadler : They'll be subject to collective action lawsuits if they train on material that they don't have the rights to. Senator DAVID POCOCK: Yes, but the question is: you're saying that they don't need this material; why, then, do they want to be able to train on the entire internet if they don't need musicians and authors? I'm sure many musicians and authors would say, 'Well, great. Don't use us.' CHAIR: I am sorry. I'm conscious we are now 10 minutes over time, so we might take some of that on notice, if we can. Thank you, Mr Sadler, for your contribution today..