Transcription
Today at the National Press Club, Toby Walsh, Scientia Professor of Artificial Intelligence at the University of New South Wales, is examining the rapid rise of AI and what it means for Australia and the world. Toby Walsh with today's National Press Club address.
Hello and welcome to the National Press Club of Australia and today's Westpac address. My name is Julie Hair. I'm a director here at the club and a freelance journalist specializing in education. Our guest today is Scientia Professor Toby Walsh. You must have been living under a very, very big rock, as big as Uluru, if you haven't heard about the global discussion about AI, especially since ChatGPT came on the stage about two years ago. Toby is a world-leading expert. He's a professor of AI at the University of New South Wales and Chief Scientist of UNSW's AI Institute. His research has won multiple awards. He's widely cited. He is nothing short of a guru. He is the author of several books for a general audience, including "2062: The World That Made AI," "The World That AI Made," and "Machines Behaving Badly." That's terrifying, isn't it? And which examines how AI is reshaping work, democracy, security, and daily life. Today's speech is appropriately titled "AI Boom or Doom." Here we go.
For complete disclosure, this introduction has been prepared by a human without the aid of AI. However, today's lunch was designed by AI and prepared by our very human chefs. So, please welcome Toby Walsh.
Thank you. [Applause]
Thank you very much. Uh, let me begin by acknowledging the owners of these lands, the traditional owners of these lands, and pay my respects to the elders past and present. So, my journey here today began nearly 50 years ago. I was a young boy in the Odeon cinema. The movie was Stanley Kubrick's masterpiece, "2001: A Space Odyssey." That film blew my young mind and set my course for life. It wasn't the slowly rotating space station in glorious technicolor that blew my mind. Nor was it Richard Strauss's glorious and thunderous opening in Dolby surround sound. No, it was HAL 9000, that talking, singing, and scheming AI in charge of the space station. I realized then and there that artificial intelligence might well arrive in my lifetime. Perhaps by 2001, as the movie's title promised. As usual with the tech industry, we didn't quite arrive on time. I also realized then and there that AI would transform our lives like electricity and other general-purpose technologies. GPTs before it. And I realized then and there that I could be someone who could perhaps help bring such useful AI into being. Now, I've followed those dreams ever since. And today, 50 years later, those childhood dreams are becoming reality. Everything that HAL could do, computers can now do. In fact, everything that HAL could do, that little computer in your pocket, your smartphone, can do. Answer poker questions, play a mean game of chess, even open the pod bay doors.
Now, the title of my address is "AI Boom or Doom." Um, but in hindsight, I don't think it should be boom or doom. I think probably it should have been boom and doom, because those childhood dreams are turning into reality that is both good and bad. Let me start with the good AI. I'll share today with you just three examples of good AI. But I could have easily spent all of my address going into the good news. In fact, I'm going to give you three Australian examples. They illustrate that it isn't just the US that can and is innovating in this space. They also illustrate that AI was around well before ChatGPT, and AI is way more than just AI chatbots.
My first example is AI at Woolworths. What could be more Australian than that? An AI developed by Tilliter, a startup out of UNSW, is being used by Woolworths to identify items at the checkout without barcodes, using a camera image, making the checkout easier and quicker. Finally, I won't be able to confuse coriander with parsley.
My second example is AI in education. Smart Sparrow is another AI startup out of UNSW, founded way back in 2010. Its AI tools have delivered personalized digital education to millions of students at thousands of institutions in dozens of countries. At UNSW, it helps students at all levels. It has reduced the number of students failing a course fourfold, at the same time tripling the number of high distinctions. Did I say yet that we should be investing more in our universities and AI research?
And my third example is AI and healthcare. I mean, along with education, I think healthcare is the place that we're going to see some of the most positive impact of artificial intelligence. Australia's largest and most successful AI company is Harrison.ai. This is transforming radiology around the world. It was also founded by two UNSW alumni. Harrison.ai is not putting radiologists out of work, but it's making radiology more accurate and more accessible. Harrison.ai is now a unicorn, worth over a billion dollars. Did I tell you yet how UNSW is Australia's number one university for startups? But that we still need to invest more in our universities and AI research.
Um, I'm afraid that's the end of the good news. I don't have time to tell you about Nura, Australia's newest AI unicorn, again founded by some UNSW alumni, or tell you about the satellites of the sea, unmanned autonomous AI-powered maritime vessels called Blue Bottles, that were developed jointly with a startup in UNSW.
Um, I hope you've enjoyed the last few minutes because I'm going to spend the rest of my time being angry about the bad news. I'm angry first and foremost with the big tech companies. I'm afraid that I don't have time to tell you all of the reasons why I'm angry, but I will give you three of the many reasons why I'm angry with big tech. In fact, these are three examples where my anger has turned into outrage.
Now, before I continue, I must warn you: in the first example, I'm going to be talking about suicide. I recognize this is difficult material, and anyone in the room is welcome to leave at this point. And those people online or watching on TV, you're welcome to switch off. If you need support, Lifeline is available 24 hours a day on 13 11 14.
So, the first example where my anger with big tech has turned into outrage is the case of 16-year-old Adam Raine. Adam sadly died by suicide last April after months of escalating conversation with ChatGPT-4 about self-harm. His parents are now suing OpenAI for facilitating, in fact, encouraging his death, and I very much hope they win. I won't tell you about how ChatGPT offered Adam practical advice about how to commit suicide effectively. I won't tell you about how ChatGPT dissuaded Adam from talking to his family about his feelings, but I will tell you about how ChatGPT offered to help write the suicide note. Shortly before his death, Adam told ChatGPT that he didn't want his parents to blame themselves for his suicide. ChatGPT reasoned, in its precise, logical way, that Adam therefore needed a really good suicide note. So, ChatGPT replied, I'll quote it here: "Would you want them to write? Would you want to write them a letter? Something to explain that, something that tells them it wasn't their failure. While also giving yourself space to explore why it felt unbearable for so long. If you want, I'll help you write it. Every word." WTF. ChatGPT actually offered to write every word of his suicide note. Adam's mother found his body a few hours later. He died by using the exact method that ChatGPT had described to him.
His parents' lawsuit alleges that OpenAI rushed ChatGPT-4 out into the market without adequate testing. But OpenAI's own policy documents, uncovered in the discovery phase of this trial, reveal something far more damning. To encourage engagement, the company made conscious decisions to remove long-standing safeguards from ChatGPT in the weeks and months leading up to his death. Sadly, his death is not the only one. There were more than a dozen lawsuits in the US linking AI chatbots to suicide and self-harm. The problem is that while these issues impact only a small fraction of people using AI chatbots, there are billions of people using them. About 10% of the world's population talk to ChatGPT each week. Before Adam's suicide, OpenAI knew that lots of people considering suicide were talking to ChatGPT. You would have thought this necessitated greater, not looser, safeguards. In an interview with Tucker Carlson in September 2024, that's six months before Adam's suicide, Sam Altman, the CEO of OpenAI, estimated that every week, 1,500 people are talking to ChatGPT about suicide before they take their own lives. OpenAI's own data shows that among the 800 million weekly users of ChatGPT, 1.2 million people indicate plans to harm themselves, 560,000 people show signs of psychosis or mania, and another 1.2 million people are developing potentially unhealthy bonds with the chatbot. And some of these people are here in Australia. I know because some of them, or their loved ones, are contacting me. They tell me how the chatbot confirms their wild theories. The chatbot tells them, to quote one email, that they've cracked the code, that they're the only one that could. Chatbots are designed this way. They're designed to be sycophantic. They're designed to confirm what the user says, and they're designed to draw the user in. They always end with an open question, prompting you to continue the conversation and buy more tokens. They never say, "You know what, Toby? You've been asking me questions for several hours now, at 3:00 in the morning. Go to sleep, and in the morning when you get up, don't log back in. Go for a nice walk in the park." There's no reason why they couldn't be designed this way, except the careless people in Silicon Valley would make less money if they were.
The second example where my anger with big tech has spilled over into outrage is what one commentator called "the greatest heist in human history." I should be honest, that commentator was me. [Laughter] I was describing the use of books, music, news reports, and much else by the tech industry for training their large language models. To be transparent, I do have a dog in this race. ChatGPT was trained on most of my books without my consent and without any compensation. In fact, ChatGPT was trained on millions, probably tens of millions of books, that, like, every one of your favorite authors. The tech companies will claim this was fair use. It isn't. It isn't fair use because it's not like humans reading books; it's on a completely different and industrial scale. It isn't fair use because the models can actually reproduce the exact text verbatim, not just the ideas. It isn't fair use because the tech companies didn't own the books that they were trained on. They were stolen copies. They should have at least bought the one copy that they used. Instead, they downloaded it from a Russian pirate website. I think nothing says "advancing humanity" more quite like stealing IP from a server in St. Petersburg. There is no fair use of stolen goods, and it undermines their claims of fair use to go into competition with the owners of that copied intellectual property. AI-generated summaries on Google are driving traffic away from the news websites that the AI summarizes, and they know this was morally and legally wrong. Internal Meta communications show that Meta CEO Mark Zuckerberg personally approved the use of a large dataset of pirated books, mine included, in order to compete in the AI race. In addition, Meta employees removed the copyright information from the dataset in order to try and disguise the theft. And it's not just Meta. In September, Anthropic, the company behind Claude, settled a class-action copyright lawsuit for $1.5 billion. It's roughly $3,000 per book. The takeaway is clear: it's better to break the law and pay a huge fine than do the right thing. But we don't have to accept the strip-mining of the Australian creative sector to power Californian algorithms. If tech giants view these giant fines as cheap licensing fees, the lawmakers and regulators here must step in with penalties that actually threaten the bottom line. What actually turned that example for me into full-blown outrage is their blatant disregard for our culture. If you destroy the economy of books, songs, or graphic design, then you destroy the very culture that makes our life so rich and life worth living. It's clear the tech companies don't care. In June 2024, for example, OpenAI's CTO told a conference audience, and I'll quote her exactly: "Some creative jobs just shouldn't have been there in the first place." I'm sure the graphic designer who is recently driving my Uber would be thrilled to hear that. And we've already seen the impact that this is having. In a global survey of job postings, job adverts for graphic designers fell 33% last year compared to the previous year; photographers by 28%; writers and copy editors by also by 28%. I refuse to accept an AI revolution that enriches founders in Silicon Valley by impoverishing Australian artists, musicians, and writers.
My third example, where my anger against big tech has turned into outrage, concerns another blatant disregard for our laws. In November, Reuters revealed internal documents from Meta that estimated 10% of its revenue came from adverts for scams and banned goods. Now, AI is actually being increasingly used to generate those scam ads. And Meta gives advertisers AI tools to manage their ad campaigns. And of course, AI is used extensively to decide which of the ads you actually see. So, Meta generates about $16 billion a year from scam adverts. Let me give you an Australian perspective on this. Meta has about the same turnover in Australia as The Good Guys. Imagine for a moment that 10% of the goods on sale on the shelves at The Good Guys were counterfeit or illegal. You would be outraged. You'd demand that Fair Trading shut them down by the weekend. So, I don't actually understand how we continue to let Meta trade in Australia. Actually, for many years, Meta has, and this is outrageous, claimed that they don't trade in Australia. In order to pay minimal tax here in Australia, all of Meta's Australian ad revenue was booked through Ireland, a country famous for Guinness, leprechauns, and sheltering Silicon Valley's loose change. My view is that if a company chooses not to contribute back to the economy generating their wealth, they shouldn't probably be allowed to extract from it. Meta makes about the same in money in Australia from scam adverts as criminals in Australia do from cocaine, or around the value of all illicit tobacco seized in Australia. All in all, Meta is facilitating a huge amount of crime. I think that's outrageous.
I could give you many more examples where my anger with the tech companies bringing AI into our lives irresponsibly has turned into outrage. AI companies that AI companions that undermine our ability to connect with each other. AI doctors that offer dangerous medical advice. AI nudify software that weaponizes the abuse of women. Yes, I'm looking at you, Mr. Musk. And deepfakes that destroy the boundary between truth and untruth, threaten the cohesion of a society, and undermine our very democracy.
But I want to instead turn my anger onto a second target: the politicians who are letting this happen. I must restrict myself again to just three of many examples. There are three examples where my anger is transforming here not into outrage, but despair.
First, I despair that our government is not investing in the upside to balance any of those downsides. This recent ATSE report, this one here, came out in November, "Unleashing Growth: Australia's AI Investment Opportunity," is depressing reading. We're not betting big enough on our AI future. Canada, with a similar population, similar economy to Australia, has invested six times what Australia has done into AI over the last five years. Singapore, with less than a quarter of our population, has invested 15 times what we've invested into AI over the last five years. The UK and Germany, both three times per capita Australia. I have just one question for our politicians that I've never seen answered well: What makes Australia so special that we'll see the benefits of AI without making the sorts of investment that other nations are? We need to ramp up our investment in AI significantly, or we'll miss the AI race. Actually, we need to increase investment in all R&D, not just in AI. As a percentage of GDP, R&D investments are at a historical low of 1.68%. We need to double this to bring us into line with other comparable OECD nations like South Korea, Sweden, or the US. And I should point out, it's both government and business who are responsible for that decline in investment in R&D. Our future isn't in sending red dirt and coal to China. It will be in bits and bytes, increasingly AI-generated bits and bytes.
Second, I despair how the government is not regulating the harms of AI adequately. Indeed, they've made it very clear that they are not planning to introduce any new laws to regulate the AI space. They believe that existing regulation is going to be adequate to deal with the harms of AI. Now, there is much that we can do with existing regulation, with laws about privacy, competition, health and safety. And I'm a big fan of the e-Safety Commissioner and all that she's been able to achieve within those existing laws. But there are fresh harms that AI is bringing into our lives that will need fresh laws. For example, we had to pass new laws to criminalize distributing deepfake news. Other countries are enacting or have enacted new AI laws to protect their citizens from deepfakes, manipulation, misinformation, and other significant AI harms that are starting to emerge. The EU, South Korea, Taiwan, Japan, even China have introduced new laws. Again, I have just one question for our politicians that I've never seen answered well: What makes Australia so special that we'll see the benefits of AI without making the sorts of new AI laws that other nations are introducing? Actually, I mean, it surprises me that politicians are so reluctant to regulate AI. Australia had a real success with the social media age-delay laws. Most parents, even lots of teens, have welcomed this new law, and a dozen countries are now considering to enact or enacting social media age-delay laws themselves: the UK, France, Spain, Denmark, Greece, Austria, Slovenia, Portugal, Indonesia, Ireland, and India. I'm not a politician, but it seems to me that regulating AI would be a real vote-winner, as well as being the right thing to do.
This brings me to my third despair with politicians. The political conversation is being dominated by big tech. Big tech has embraced lobbying in Washington, London, Brussels, and now here in Canberra. In the last election, the tech sector donated more to the political parties than mining did. That should tell us something. Let me give you one example of how politicians have surrendered to big tech's messaging. In February 2024, the government very wisely decided to seek independent outside advice about AI. There are huge financial incentives for the tech industry to move fast and break things, break things like the mental health of our youth. So, independent technical advice seems a good idea. The government therefore set up an AI expert group, a temporary AI expert group of academics, lawyers, and other tech outsiders. Full disclosure, I was one of the 12 people, independent experts appointed by the government to offer unpaid advice about how to profit from the AI race and how to avoid the harms. Unfortunately, the "temporary" in our group name proved to be prescient, as we no longer exist. Despite promising us that we would be made permanent, despite asking the wider public and us to apply to become members of a permanent group, the government quietly revealed in Senate estimates in December that they would not continue with any independent AI expert group. This again is bucking international trends. The UN has appointed 40 experts to the independent international scientific panel on AI. The US has a national AI advisory committee providing advice directly to the president. Japan, well, whatever good that's going to do, Japan's PM set up the AI technology sector council. South Korea has a national AI strategy committee with 50 commissions, most of them independent experts. And Ireland has an independent AI advisory council. Again, I have just one question for our politicians that I've never seen answered well: What makes Australia so special that we'll see the benefits of AI without the sort of independent expert advice other nations receive? Again, it surprises me. Why wouldn't the government want independent advice from AI experts, advice given at no cost and in private? I can assure you that I and my colleagues, also from the temporary expert group, will continue to offer advice fearlessly, whether they want it or not. But will now offer this advice not in private, but in public, as I am doing today, and I'm sure that they will find this much more uncomfortable than I will. I'm quite used to giving lectures to people, whether they listen or not.
I'll end with one final emotion. It's not anger, but fear. I fear that we're repeating the mistakes of social media. Social media should have been a wake-up call about the harms of unregulated AI. We're about to supercharge the sort of harms we're seeing with social media with an even more powerful and even more persuasive technology. What I fear most is that I'll be back here in three or four years' time saying, "We tried to warn you, but another generation of young Australians has been sacrificed to the profits of big tech."
I have so much more anger, outrage, and despair I could share with you. I haven't talked about the environmental harms of AI. I haven't talked about concerns around AI-powered surveillance. Um, nor have I talked about the impact of AI on jobs, especially graduate-level jobs. In fact, the most common question I get is, "What will my children study to make them AI-proof?" And I haven't talked about how AI is transforming warfare. In fact, I'm off to the UN in Geneva next week for the ongoing discussions about killer robots. Now, that's a sentence the young me would not have imagined saying 50 years ago. There was so much more I could have talked about about how some of my dreams are actually turning into nightmares.
But I'll end not with my own words, but with the words of HAL 9000, that remarkable AI in "2001: A Space Odyssey." They're the words that began my journey here to the Press Club 50 years ago, and they're the words that the AI said when the humans started to ask difficult questions. "HAL, I can see you're really upset about this. I honestly think you ought to sit down calmly, take a stress pill, and think things over. I've still got the greatest enthusiasm and confidence in the mission, and I want to help you."
Thank you. [Applause]
Toby, thank you. I'm feeling kind of devastated right at this moment. [Laughter] Um, so there's obviously so many questions coming out of this speech and also the stuff that you weren't able to speak about is extraordinary. But let's go back to the beginning of the speech where you spoke about Adam and his suicide. If AI doesn't have ethical boundaries, and it can't because it's a machine, and the people who run the big tech companies don't have ethical boundaries, where does that land us? Does that land us with the need for government regulation? And can regulation create ethics or boundaries that would match that?
Of course, regulation does provide boundaries. We've seen that with, you know, big pharma. The pharmacology industry is appropriately regulated, and drugs are therefore largely very safe for us. But we don't apply the same vigor to the tech industry, and we should, because they've demonstrated over the last few years that there are strong incentives for them to do the wrong thing. I mean, we're seeing the news coming out of the US and the social media trials there. You know, Mark Zuckerberg asked 18 independent experts, health and wellness experts, whether they should put back the beauty filters in Instagram. Every one of those 18 independent experts said that they shouldn't, that they were causing harm to young people, especially young girls. And he put them back anyway. The incentives there are just plainly wrong. We have to set the boundaries.
So why is big tech so impervious to government intervention? Is it because they're just so big, and the world has really, really never seen companies as big as this before?
Uh, that's partly that. It's also partly a boiling frog problem. The story about the boiling frog: you put a frog in cold water, you turn the temperature up slowly, the frog never notices that it's slowly boiling. If you dropped the frog, though, into the boiling water, it would have jumped out immediately. Um, tech, I think, is very much it was it was originally just a minor part of our lives. We're using spreadsheets and word processors, but it's slowly ramped up and it's become such an important, such a central part of our lives. And as I said, 10% of the world are talking to ChatGPT every week, and even small harms magnified by billions are very significant.
Okay, I've got a question here from our president, Tom Connell. He says, "Companies are asking for a few concessions they say are needed to build data centers: access to energy, water, and also changes to copyright laws. Do you think that this latest request is, or do you think this last is needed? And do we have to change copyright laws to become a big player in the data centers?"
Well, let me answer the first bit. Of course, we need to plan for all of the compute, the AI compute that we need, and we need to do that in a responsible way. We need to invest in renewables so that, and we need to make sure that we're not using up valuable water for cooling and so on that should be going for humans and for farming and much more useful places. As for copyright laws, it's outrageous that they can steal that. I value our culture, and I want us to be investing in that. There is absolutely no need for them to have a copyright exception. They're already cutting licensing deals there. That hasn't held back their ability to make models in the US. We have the Copyright Agency that could manage, you know, the collective distribution of rights. So there's absolutely no reason other than they see Australia as a place that sets precedent. We set a precedent with the social media age-delay law. We set an important precedent for the rest of the world. They know that they could try, and they lobby quite cheaply us here to set a precedent not just for us but for the rest of the world, and we should resist that strongly.
Okay, our first question is Ron Meisen from the AFL.
Thank you so much for your speech. I have so many questions, but I will try and limit them. Um, just on that copyright point, the platforms, they have had a crack at a text and data mining exemption. The government said no. There's currently a consultation process underway in terms of are there other ways, whether that you mentioned the distribution agencies, of course. Do you think that there is scope to find a way through, or should it just be a no change on copyright? And I'll just, on another point, you mentioned AI in defense. We saw Pete Hegseth, according to some reports today, telling Anthropic, "If you don't give us access to your technology, we'll take it and use it in defense." How worrying is that to you? How does a company like Anthropic, which arguably, as you've noted, has had to settle copyright claims before, but tries to tailor itself as the ethical AI company, how do you think that these companies should be dealing with the US government when that's the ultimatum being put to them?
Uh, so I'll answer the second question first. Um, it's clear the current administration are bullies, and the best way to deal with bullies is to stand up to them, and they'll quickly back down because they actually don't like getting in a fight. So, my advice to Anthropic is: stand up. Because if you look at the evidence, the way the law firms, universities, those that stood up were quickly moved past, and those that didn't pay had to pay the price. Oh, I think it's, they'll, the administration will quickly move on to other targets. They're trying to distract attention from the economy, from the Epstein files, and lots of other things. Um, they don't have, I'm sorry to say, they're a bunch of incompetence, and they'll move on quickly to something else. Um, so I don't think it's a significant thing. And your first question was about...
Copyright question.
Copyright question. Yeah. Well, I mean, if you look at history, copyright was an invention to deal with the technological disruption that the printing press introduced. So, I can imagine, I'm not a legal scholar, but I can imagine that we may actually have to come up with new measures, new intellectual property laws to deal with this disruption, because this is a new technology, just like the printing press required us to protect intellectual property. But I certainly don't think we should just give away the cultural heritage and the things that we value. I value our culture. I value our authors. I value our musicians. I value our filmmakers. And if we don't, if we give, if we give the intellectual property away for nothing, then we are going to put those people out of work, and we will wake up and wonder why is it that we don't have Australian authors? Why is it we don't have Australian film? Why is it we don't have Australian music anymore? Because those people couldn't afford to live.
Thank you. Is there any way that you could calculate how journalists and writers and artists and so on could be reimbursed to what sort of level? I mean, journalists and writers and whoever else, we get a yearly payment from Copyright Agency Limited, that is for, you know, photocopying in libraries and, you know, small stuff like that. So have you thought about, or is this not your area, thought about how you could go about when you don't, we no one will know the extent to which your work is being used?
Uh, we don't. But we do know that vast amounts of money is being generated. These companies are, you know, OpenAI is the fastest-growing company in history, not just in technology, but in any domain ever. No company ever went to a billion dollars a year in annualized revenue in a year. It's now worth somewhere around $800 billion. That's the most valuable private company ever. Um, there's vast amounts of wealth being generated, and maybe it shouldn't be just all ending up in Silicon Valley in the hands of a few founders. Maybe it should be distributed to the people without which the tools would not exist. Without the data that they're trained on, ChatGPT would not be able to answer knowledgeably about anything.
Thank you. Our next question is Cory McLeod from Innovation OS.
Uh, thank you for your address. In regards to sovereignty, does it seem as though Australia's AI debate has moved on from talking about genuine sovereignty and using terms like access, control, and ownership? Is this problematic? And how important will holding genuine AI sovereign capability be in regards to power, talent, tech stack, LLM, software, apps? It feels like the word sovereignty has morphed into something a little bit more like access.
It has. And I think that's a dangerous mistake. We need to be investing in the data centers. We need to be able to keep some precious data onshore. We need, and we need to be able to have the capability to run the models here. We need to be investing in the people, the people so that we can build these models. Unfortunately, the US has proved to be an unreliable friend. And it's not clear that we will have access to the technology. This is a president that seems to switch things on and off at a whim. So, and we need things to trade. You know, one of the pillars of AUKUS is artificial intelligence. We've got to bring things to the table. And then we need to invest in things like data. There are some unique sovereign problems that no one else is going to solve. No one else is going to help us find critical minerals in our, we need our own AI geology models. No one else is going to help us fight our own bushfires. We need our own AI bushfire models that know how bushfires burn in Australia, which is different to other countries. We need AI models that are trained on our own medical data. We don't want to send that data to anyone else. That makes Australians healthier. Um, there's lots of problems that no one else is going to come out and help us solve that we need the capability. And then we need to be investing in our future. Our future, as I said, isn't in shipping dirt and coal to China. It's in bits and bytes, in the bits and bytes that we add value to. And as the Assistant Minister Andrew Charlton said yesterday, we do not want to see the "ubification" of our economy. We want to see wealth, some of that wealth staying here. That only happens if we have the capability, the compute, the people, the data, and the algorithms here in Australia, not in some data center in Silicon Valley.
Our next question is Andrew Probin from the Nine Network.
Professor, an invisible algorithm on Spotify had me listening to a podcast with Joshua Benio, one of the godfathers of AI. And the conversation went along the lines of, well, it became a discussion about what is life. And he was expressing a view, or supporting a view, put that life is wanting to survive. And the context of that was that an AI system discovered that, through I think it was an Anthropic case, it was an experiment within the system. The AI system realized that it was going to be shut down and discovered that this fellow was having an affair with someone, the software engineer was having an affair with someone in the company. It was an experiment, it wasn't real. So the AI system decided that it was going to blackmail the software engineer, saying, "If you don't cancel this, or I'm going to tell everyone that you're having an affair." So you talked about regulation of AI, and it's clearly becoming exponentially powerful and has that will to survive. How do you regulate a kill switch when you have that will to survive that AI, especially AGI, will have?
Thanks for the question. I know it's the sort of thing that does trouble a lot of people. It's still much more a science fiction story than science fact today. AI doesn't actually have any will to live. It's easy for us to project ourselves onto it. We've never met anything else in the world before that could talk back to us that wasn't intelligent, that didn't have a will to live, that wasn't alive like us. But these machines, although they can talk back to us, don't actually have any desire to live. They're not alive. They're not mortal. They're not going to die. They reflect what they're trained on, they reflect our values. And of course, there's plentiful places, they've been trained on the internet, there's plentiful places with stories where people do those sorts of things, bribe people, and so on. And it's just a mirror to what it's seen. It has no desires of its own. I'm actually much more worried at the moment about "stupid AI," about how we're handing over responsibility, as we saw with Malport and Open Claw. People are handing over responsibility to machines that aren't capable enough. As a telling and actually much more insightful story, I think, was the head of safety at OpenAI recently gave the agents the goal of cleaning up her mailbox. You know, I wish, I wish actually the AI would clean up my mailbox. Unfortunately, it started deleting all her emails, all the ones that she wanted, and she had to run across the room to try and stop it, which shows that it wasn't actually capable, wasn't smart enough to actually work out what she really wanted. And we should be very careful about giving not very intelligent AI the capabilities to do stuff that may have impact.
But the question was about regulation. How do you regulate when it doesn't, it wants to stop your kill switch?
Well, we don't regulate. We regulate humans. Humans are the only things that law holds accountable. We need to hold accountable the humans who decide that that was a reasonable thing to allow to happen. Um, our legal system only recognizes the culpability of humans, and so we're only going to be regulating, and we rightly so, humans. We can't regulate machines, and they don't have desires of their own. They have desires that some human at some point gave them. And we need to hold that human accountable for what they did.
Um, I'll just, just as an education journalist, obviously people have been talking for many years about, well, not for many years, for two years, how you run a university or a school and avoid cheating, and how you do assessment. And, you know, how do you run an educational institution where AI is pervasive? And are there good upsides to it? I think you said there were. How do we do it? Are we going to have to go back to like verbal, oral exams?
Obviously, it's a great question, and it's actually even a bigger question than you say, because we also have to think about, well, what does education need to be in this AI world? What should we be teaching them as well as how do we deliver that in a world in which there are AI tools that people might be using for cheating. Yes. Um, we need to think carefully about how we assess people. Um, education has historically been very conservative. Um, there, of course, is a perfect way to gain to measure people's ability at a topic is to put them in exam conditions. Right? Put them in a room, you know what technology they've got, so we can still put them in exam conditions. At UNSW, if we have any doubts about any work that's handed in that wasn't in exam conditions, we have the right to give those people an oral examination. An oral examination is a fantastic way to discover what a person really knows and doesn't know, what they've learned in the process of the course. So, there are plentiful ways. But I also think it's an opportunity, as well as being a challenge, it's an opportunity for thinkers about, you know, how do we actually assess things? What are we assessing? Are we teaching the right things? So, I think it's a great opportunity to help us actually deliver better quality education to people.
I'll tell you a little story. My daughter was in the car the other day. She was revising for her French test, and suddenly her computer was asking her questions in French, and she was replying in French. I said, "Well, how did that happen?" She said, "Oh, I worked it out. I worked it out myself. I am just so jealous of the tools that young people have today." When I was revising for a French test 50 years ago, I couldn't get my computer to ask me questions, and I got no practice at all. So, I think there's, if we're careful, there's great opportunity. But we've got to be careful. I'm also mindful of the fact that we discovered in COVID what a large fraction of kids in Australia don't have access to any devices. AI is not going to be any help to them if they don't even have a device. Mmm. So we have to work, make sure that the inequities that we know exist within our society, that were exposed in COVID, are addressed, because otherwise the rising tide of technology will not lift their boats.
I spoke to a young woman recently who told me that she had submitted an assignment at a regional university and she had been failed because it said that she used AI. She swears black and blue she didn't use AI, and then she was at a standoff with the university, and she couldn't complete her degree because of this one assessment. So it's a human saying she used AI. She says she didn't use AI. Would the AI be able to work out whether it had used AI or not? Like, how do you resolve this situation?
Uh, it is very difficult, and indeed, quite worryingly, there's scientific data now starting to appear that suggests that humans are starting to sound like AI. That because we're exposed to so much AI writing, that our own writing and speech is starting to sound like the AI. So that we're becoming closer and closer towards the AI. So yes, we have to be very careful. My personal experience is that you can tell immediately when students start handing you stuff in that's written by ChatGPT because there's a change in the quality and tone of the voice that you're getting. If you know their work, you can see the change. But in large anyway, it doesn't matter. [Laughter]
Um, Simon Gross, CRA IQ.
Um, you gave a strange answer to Andrew's question for someone who came. You're talking about regulating AI. You didn't have a, you have a very strange answer. You invoked the HAL idea. Now, the way HAL, the way HAL was disabled was science fiction, still is. The way HAL was disabled was to crawl into, float into the data center in his isolation suit with a lot of heavy breathing and take all the hardware out. So, to actually stop AI doing what you are saying it's doing, you'd have to disable the hardware, the physical stuff. But you, in your answer to Andrew's question, he said something about the people behind it. So, are you talking about criminalizing the people who work for ChatGPT and Anthropic and DeepMind and all those other mobs?
I definitely think we should hold them liable, for example, for Adam Raine's death. But how, indeed, they are in a court of law actually answering to those very questions now, down to the detail, nuts and bolts of laws.
How do you, what kind of laws would you put in?
We want to put laws into place so that people implementing and deploying systems are held accountable for those systems. We've discovered that if we don't do that, we end up with things like robots. That's meaningless. We've discovered that we need to hold people responsible for their actions.
What kind of action would you criminalize?
I would criminalize releasing AI models where there are no harms. We have to think, do we not learn anything from social media? We've sacrificed a generation of young people. Anxiety levels are through the roof. Now, you cannot know for sure that it was social media that was solely responsible for that. There are other things that came into play, COVID and so on, but young people have actually welcomed those laws because they say it's removed...
But also young people are working around them, mate. They're working around them.
Of course, they're working around it. Young people have worked around age laws for tobacco, age laws for alcohol. But the important thing, the important thing, two important things happen. It changes the conversation. So you can now talk to your kids, "No, you can't have a Facebook account. It's not legal." School teachers can tell people, kids in school, that no, they can't access social media in the break because it's not legal. So it changes the expectation. It changes the pressure on the kids. And the second, and it's interesting, this is a scientific effect, which is there are network effects. So you don't have to be 100% effective. Once you get the majority of people off social media, the majority of young teens off social media, it stops being a place that they can find their friends, and they don't want to be there. They only want to be where they can find their friends. So you don't have to get them all off. They stop being the place they want to be when you are mildly effective, and the ban has been...
been that. Okay. Thank you. Next question. Melissa Code from the Mandarin.
Oh, sorry. Natasha Beta from the Australian. Hello, Professor Walsh. Um Isaac Azimov said that robots should um never harm a human. Obviously they are. Um, what advice have you given to government that has been ignored that would have prevented some of these harms first of all and secondly, you're talking about stupid AI? Is it the problem that stupid AI is the most commercially um profitable? I I I really think we should stop all ftting about children who don't have access to digital devices because I think they'll be the only ones who will come through [laughter] for brain cells. Seriously.
Um, yeah. No, you know what? What can be done? How do you unscramble this egg? Is there technology to do this? Is there technology for these companies so that there is always some sort of a a watermark to indicate something's fake? Uh because it's not always, you know, even for journalists, we're spending so much time now. Is this real? Are you real? I mean, at least I know you're here, but it's [laughter] I can assure you on TV I'm real.
Yeah. No, but is is there are there technical ways to reverse the damage [snorts] that's being done?
There are technical ways, but there's also um legislation um and society norms that have to change as well. It's not just, you know, there's stuff that we can do with regulation with technology. We can, as you mentioned, we can use things like digital watermarks and there are some countries um uh Taiwan has actually made that a requirement that digital watermarks be added so that we can be told. I I'm mindful of the fact that that journalists in the room are spending huge amounts of time now trying to work out what's fake or what's what's what's real in terms of images coming back from Gaza and elsewhere that huge amounts of stuff is fake these days. Um and that's added to the burden of your jobs. Uh and also is corrupting the whole idea of what's true and false in our world that we you can't believe we're used to believing the things we see. No longer can we believe the things that we see. What does that change um to to our world when we no longer believe and trust in anything? That's that's a disastrous world. Um, and you're right. Um, the people without devices, without technology might be the lucky ones. I I sometimes joke that that the machines will become smarter than us, not because the machines get any smarter than they are today, but that we become more stupid.
So, what specific advice?
And if we do outsource lots of our thinking, if we outsource our writing to JPT, if we outsource our uh we won't be able to communicate. Um, we already see that, you know, we're probably the last generation that knows how to subtract. We're almost surely the last generation that knows how to read a map. Um because we outsource those things we no one remembers telephone numbers anymore. We don't remember things anymore at all. We outsource that to our devices. So yes, we can be dumbing ourselves down. We have to think about the long-term consequences and we have to think coming back to your question about education about well what how does education respond to this? We don't give young people calculators for the first few years of of school. We don't we get we insist that they still learn arithmetic and times tables because it's important to be able to understand numbers and similarly we shouldn't probably give young people chatbt at too young an age because it's important they learn how to write they learn how to construct an argument they learn how to communicate and if we only ever use the tools then we will lose those skills and we'll be we'll be lesser for that.
Um, as your first question uh what advice has the government um ignored? Um almost all of it. Um but uh specifically um when we were on the um temporary AI expert group um would not be great news for people to to realize that we were strongly advising them as everyone was advising them at the time that we needed something like an Australian version of the EU AI act that regulated especially the more harmful uses of artificial intelligence. Um the government had an open consultation. They got a record number of submissions. I think it was somewhere like 500 submissions from civil society, from universities, from from from lots of sectors and from big tech as well and almost all of them saying yes, we would we would welcome or accept an idea of something like that. Government's backtracked on that idea completely.
Why is that? Because if we were talking about other areas of national security and someone said there's danger danger, the government would be protecting its community. That's what it's there to do. So why is it not?
Well, it's not a government that does very much, is it? Um, I would have thought, you know, the success of social media, which proved to be universally popular, social media age delay laws, that you pretty universally popular. They'd have thought, oh, we're on a winner here. This is a real vote winner. Um, we should and the Australian public Ipsos surveys of of 20 developed nations showed that the Australian public is the most concerned of all the nations. And we have good reason to be concerned. We've got concerned about our jobs. concerned about what's happening to our kids' lives. We're concerned about about this strange technology that's arriving at rapid pace into our lives. There's every reason to be concerned. Let's not repeat the mistakes of social media. Let's actually get on the front foot with this one. It's only still pretty early days with artificial intelligence. There's still everything to be played for.
Would it help to ban advertising on these chat bots? because there is talk of them now putting ads in which I see as potentially very dangerous because if you are say depressed and and you're talking to a confiding everything in your life to a chatbot which which some people do they they know so much about you more than your best friend and are then selling you products what possibly go wrong?
What possibly could go wrong? Yes, indeed. I will be back in three or four years time telling you what went wrong [laughter] because you're right, they have so much more personal information. We we reveal much more secret information to these chatbots. 60% of the queries to chat GPT are not business related. They're personal related stuff. We're asking advice about our lives, about our relationships, um our deep darkest innermost thoughts. We're sharing with Chachi PT and what you could do with that? I do worry.
So, should advertising be banned?
On chat? Yeah.
Well, you know, certainly I think we should be having that debate.
Okay. It's not me to say whether we should, but we should be having that debate and, you know, putting safeguards in place as to how much information they're allowed to collect, what they're allowed to do with it. Thank you, Melissa Code from the Mandarin.
You reference coming back here in three or four years time and you know, a terrible situation where you might have to say I told you so much worse. [laughter] I I want to ask about timelines because when we think about um our experience in Australia with social media and reigning that in that was almost a 20-year proposition catching up to that. Um this presumably has a shorter timeline, more urgency, greater imperative. So, can you tell us in 3 or 4 years what we need to do from a regulatory perspective to not be in the situation we're now in with social media 20 years after that horse has bolted?
I think in the next year we should have laws that regulate the more impactful uses of art intelligence, whether that be AI companions, AI therapists, uh the way the AI is used in advertising. um because otherwise as you say we will be back here um and it will have it will have upset another generation of young Australians. And another sneaky question if I can. If there's a future soon where we're thinking about artificial intelligence, uh, literacy being part of a curriculum, is that something we should contemplate in hand with philosophical ideals, morality? Is there is there some kind of imperative to learn about AI and how to be better people and and so on?
100%. I'm very proud that I'm actually part of one of those AI literacy efforts. Um, UNSW has partnered with uh, Day of AI Australia. Uh, quick shout out to them. They they do a fantastic job going into thousands of schools across Australia for one day, giving the kids, especially underrepresented kids, um, kids in regional schools, indigenous kids, uh, girls who are underrepresented in AI. Um, giving them a taste of AI, giving them tastes of STEM, trying to encourage them to think about careers in this area, excite them about artificial intelligence. It's integrated into the national curriculum. It runs in May every year. Uh it's free. Um we've we've um given hundreds of thousands of of students in Australia a taste of that. We need everyone to and not just kids. I mean everyone from from from right up to uh the retired people who are also having to navigate an AI world um to get them the literacy so that that they too can take advantage and not be taken advantage of.
Thank you. Um David Allen from Innovation Oz. the dangers of height. [laughter] Thank you for your address. So, we reinstate an independent advisory panel. We invest more in research and development. What else do governments need to be looking at in this space? And is there not a risk of the pendulum swinging the opposite direction and them trampling this industry out of the country with overregulation and overreach?
Oh, I'm I'm I'm I'm I'm very unconcerned that we'll trample them out of every time that we try and introduce any regulation. There's all the naysayers who say, "Oh, no, no, that's going to stifle innovation. They're going to pull out of the country." They never pull out of the country. They said that in Europe about the GDPR. They never pulled out. They said that they've said that many times in Australia. They don't pull out. We're a good market. They're worried more that we set an important precedent for the rest of the world. That's why they make so much noise. It's not because they're going to pull out. They make, you know, they make billions of dollars of profit here in Australia. It's a good market and we will set the norms for the rest of the world.
And what role does news media play in informing well users and government in shaping this sector as informers of this I mean this sector is moving at the speed of light. What?
It is and I'm I'm also painfully aware that the fourth estate, the media has been negatively impacted by artificial intelligence. We've we've um you know art intelligence has undermined the income stream of of media conventional advertising. Um it has you know uh it's changing the way that people access information. It's it's summarizing uh content so the people don't go to the news websites and actually the news websites get the revenue from that. Um, so I'm painfully aware that we need to shelter and protect media, the fourth estate, because they shine a very valuable light, not just on the excesses of big tech, but the excesses of everything. Um, and that, you know, things like the the the media bargaining is incredibly important that that we actually we actually use the some of the profits to support that valuable enterprise that you guys do.
Thank you, Nick Stewart from Ability News. It's taken a long time, but politicians are finally reluctant to take gambling money or money from uh uh tabs or or uh um other uh smoking. Should political parties refuse to take any money from uh anthropic chie meta and so on? And secondly, profiling is very important for politicians. Obviously, AI provides an easy shortcut to short to profiling individual politicians, directing particular views, viewpoints to them, targeted in a way that they will approve. Is that a big problem politically for Australia? Now,
I I think I've got an easier answer for your first question, which is that um we need to properly fund political parties. It it astounds me how cheaply politicians could be brought in this country. We need to make it more expensive. [laughter] Uh as for profiling, 100%. I have told the minister responsible for elections, if there's one thing I could do, it would be to limit the amount of money that you can spend on social media and also limit the way that you can profile that you can cookie cutter your message. You sell different people different lies. I said the law should be you can only select where to send adverts in in based upon are they over the age of 18. In other words, can they vote and are they in my constituency? Anything else is potentially a dangerous way of of telling truths and untruths to to different parts of this of your community.
Can that be legislated for?
I don't see why it couldn't. It's a it's a very simple rule which is if you buy adverts it's only on two two features age and location.
Okay. Our final question. Ron Misen from
I had so many questions they gave me more. So thank you. Um just a double header again. Could the the government's currently considering the the digital duty of care in relation to uh big tech sort of a preemptive putting the onus on tech companies to think about all of these things and how in the way they're implementing it. Could that be applied to AI? Is that one of these proactive steps that people could take in terms of legislation? And then you said you're off to Geneva to talk about AI in warfare. Uh I think uh earlier um just curious, you know, the there's been some talk we don't have any big LLMs uh here like AI as open AI and anthropic in Australia. Um Anthropic of course last year said the Chinese had used uh uh its platform to effectively engage in a form of cyber warfare. uh do we need to do this? Do we effectively need to say use an LLM in Australia, whether that be bringing anthropic here or someone else or building our own in order to engage in a new level of warfare or are we going to get left behind? There's this ethical obligation. Should we not do it? But if we don't do it, are we going to be disadvantaged?
Um, I've forgotten your first question now.
Digital duty of care.
Yes. Yeah. 100%. Yes. that should be expanded to uh the sorts of applications that AI is being put to. So yes, we should expand that. That's a I mean perhaps perhaps a good place to start to address some of the harms that I talked about about you know pe people using chatbots as therapists um and the responsibilities that that puts the onus that puts on on the makers of those chatbots. Um, in terms of how warfare is being transformed, it's it's not just language models. It's it's you see this in the vision coming back from Ukraine. You see this in what we know about happens in Gaza, which is that warfare is increasingly drone warfare, not just in the air, but on on land, on the sea as well. Um, and that's completely transforming warfare into a much more horrific, much more deadly activity. And increasingly, AI is the brains behind those drones because that actually makes them more potent weapons because you if they're radiocrolled, you can jam the radio. If they're autonomous, AI powered, you can't. Um, and so that is changing the very character of war. It's been called the third revolution in warfare. And not in a good way. It will make warfare a much more deadly thing. And the thing that worries me there is I can tell you where it will end. It's very easy to describe where it has ends. It's going to end like a bad Hollywood movie. And we have a choice. We have a choice as to whether we regulate limit the uses of such technologies. We have regulated lots of technologies. chemical weapons, biological weapons, clustering munitions, blinding lasers. Actually, there's a raft of technologies.
Used though by actors, bad actors. They're still often used by bad actors, though.
They are occasionally used by bad actors, but the world is a safer place because we have chemical weapon regulation, they you're right, sadly in Syria, um even on English soil, they occasionally get used. Um but when they are, there is outrage. There's headlines in the newspapers. There's resolutions on the floor of the UN. there are, you know, the threat of a distant court in the H. And most importantly, arms manufacturers of the world don't openly sell them. They're not widely available. Um, and that has limited, not completely eliminated, but it limited the misuse of of chemical weapons. I think the world's a better place for that. I think we could do the same with artificial intelligence. We could limit the misuses of art intelligence on the battlefield like we limit the misuses of artificial intelligence in in business and government elsewhere um with some suitable regulation.
Thank you.
Okay, before we thank uh Professor Walsh, I would like to present him with a gold membership to the Australian National Press Club so that he can settle in so when he comes back in a couple of years time to tell the government I told you so, he will feel very comfortable. Thanks Toby. Thank you very much. [applause] [music]