Speaker 1ABC Listen. Podcasts, radio, news, music and more. If this thing is built, it will be a miracle to behold. One of science's biggest promises. It would be like the Wright Brothers moment. Governments, spies, billion dollar investments and scientific rivalries. The control that's required is rivaled only by what God did in creating the universe. Will quantum computers change everything? I'm Jacinta Bowler. Search for Science Friction, Dead and Alive on ABC Listen or wherever you get your podcasts.
Speaker 2Somewhere, right now, people are being paid to have genuine conversations with an AI chatbot. Companies training the next generation of models rely on workers to produce fresh, high quality human data. The kind a machine cannot fake. But in some cases, those conversations are not happening between a human and an AI at all. They're happening between two AIs. So what happens to a model trained on data that was never really human to begin with? Plus, all the details on the government's new data centre and AI copyright rules. This is your guide to the week in technology and culture. My name is Rae Johnston and welcome to Download This Show on ABC Radio National. I'm Rae Johnston and welcome to Download This Show on ABC Radio National. This is your guide to download this show on ABC Radio National. So now joining me to unpack some of the biggest stories in tech right now is Fergus Halliday, a journalist at Whistleout Australia. Welcome back, Fergus.
Speaker 3Hello, thanks for having me.
Speaker 2And David Browey, tech journalist at Information Age. Welcome back, David.
Speaker 4Thanks, Rae. Great to be here.
Speaker 2Now, AI company Anthropic has just launched a product called Reflect. Which is a kind of dashboard that shows you how you've been using its AI product, Clawed. So things like your topics and your patterns, the kind of tasks that you hand off to the AI. It apparently even nudges you with prompts like, what's one thing you want to keep doing yourself, even if Clawed could do it faster? Tech Crunch's read on this, though, is that the feature reinforces how much of your daily work now runs through the chatbot. Which raises the question, should the organization with a vested interest in getting you to use it more really be setting the terms for what you reflect on, and how? So we'll start with the obvious here. Fergus, what is Clawed Reflect? What exactly does this help you do?
Speaker 3Sure. So like you said, Clawed Reflect is kind of a new dashboard. It's a new element in, I guess, the Clawed interface that lets you see and visualize the patterns of your own. One of the things that it does that's interesting is that AI is often a very black boxy product. You type your words into a prompt, and then you get the result, and the work and the mechanics of it are kind of obfuscated a lot of the time. And so you don't necessarily see what your money is paying for, theoretically. And so I think that tools like this visualize the data, provide more feedback, and try and create a stronger connection between you and the product, which would theoretically make it more sticky, theoretically keep you using it, keep you paying for it, and stuff like that.
Speaker 2So it's showing it's working.
Speaker 3Yeah, showing it's working.
Speaker 2Exactly. And hopefully, from Anthropic's point of view, in the process, helping you maybe develop a little bit more trust in the tool by understanding a little bit more about how it works, maybe.
Speaker 3Yeah, or at least what it's tracking, how it sees your patterns of usage. I think that also, like, interacting with data is just inherently a very fun thing for consumer experience to do. Like, people love Spotify Wrapped and stuff like that, and so I think there's always a chase for every tech product out there to try and find their Spotify Wrapped. And I think there's a way for this to maybe lead to that.
Speaker 2I don't think people are going to be sharing what they're using their AI tools for, like, or share Spotify Wrapped. I don't know.
Speaker 3Have you been on LinkedIn?
Speaker 2I don't know. Oh, gosh. David, is it possible for a tool built by the AI company to ever honestly help you use its product less? Who should be building a product like this?
Speaker 4Well, I'd say that, of course, it's not possible for it to help you use the product less. They have no interest in people using it less. And that's certainly, I think, not the goal. It's something like Reflect. If we can be so generous to Anthropic, I'd say it's probably designed to help you use cloud smarter rather than less. It's just like Apple does this as well, and Fergus mentioned the Spotify Wrapped, just kind of giving you an idea of what you've been doing, where you've been spending your time. It's, I think, meant to just maybe make it a bit more warm and cuddly as a platform to sort of give you the impression that it's looking after you. It doesn't want you doing too much. Maybe it's trying to look after your well-being. This is feel-good kind of stuff. I don't know that everybody's going to really care that much, or maybe it will be quite a revelation and remind them maybe they should be reading more, just reading a book, that is, or perhaps spending less time making silly AI photos of cats. But in terms of helping people use it less, probably not so much, unless there are some really extreme users out there that just really have lost track of time.
Speaker 2That kind of messaging of being able to look at this dashboard and see where you can go. You can cut back on your AI usage, maybe. It does feel quite on brand for Claude, though. I have seen that there have been complaints over recent months that Claude has become a little bit smug, a little bit condescending, started telling people to go to bed instead of giving them responses, started telling people to log off, and just straight up blocking certain requests as well. And it's so funny because when you punch those same requests in to, say, Google's AI, it will just answer you. It doesn't seem to care whether you're going to bed or using AIs too much. So I think it is really interesting to have the ethical, quote, unquote, AI company creating a tool where you could moderate your use of AI. I think that that's an interesting branding exercise that I'm not sure. I'm not sure how it would work in practice.
Speaker 3Yeah, I mean, I do think a lot of it is almost that branding exercise because a lot of consumers, even if their usage patterns might not necessarily end up actually changing as a result of a tool like this, people like to think of themselves as a conscious user, which is why we have all these digital wellness tools on our phones now and other products like that. And I would say that for some, as you said, with Claude introducing a tool like this, it's a competitive advantage that they offer this versus another AI platform like GPT. And to see that extra data is a reason you might look at this product over the competitor.
Speaker 2You are listening to Download the Show on ABC Radio National. And to build the next generation of models, AI models, companies are paying people to have high-quality conversations with AI, fresh human data to learn from. Except several of these workers have told new scientists they simply get chatbots to do the work for them. Which experts are now warning is a recipe for models degrading over time. There was one worker here in this report that says she feels zero guilt and that only the sloppiest users get caught because anyone aware of AI's tells can just instruct the chatbot to strip them out, like the giveaway overuse of em dashes that we're all very familiar with now. So let's rewind back a little. What does it actually mean to train an AI? What does it mean to train an AI by having conversations with it? What's happening there, Fergus?
Speaker 3These AI models are kind of as good as the data you put into them. And initially, I would say the development of the space has been all about the volume of data, which is why we've seen lots of stories about these companies scraping the internet for as much as they can get away with in terms of feeding music or written work or websites or other content into the AIs and then get them to go through just an aggregate amount of information. And once they can make sense of that, they can then go on to do more complicated tasks because they know how to make sense of information. And so as that process has kind of played out, we've now sort of hit a point where there's no great reservoir, I would say, of untapped information and good data to put into these models at this point, which is why we've ended up in the situation where we have these private contractors who are going out and creating high quality data for models to be trained on.
Speaker 2Right. And who are these workers, David?
Speaker 4Well, that could be anybody. I guess around the world, a lot of the AI companies go through intermediaries that recruit people and dispatch tasks. And we heard about MerCore recently that had a big data breach and there was an issue with the exposure of their data with, I think, Meta that was using them to do a lot of the training. And MerCore then goes ahead and recruits people online to run these conversations, gives them tasks and projects, and everything's online and very well coordinated. Right. Right. Right. Right. Very remotely coordinated. A lot of times this is used as a way of shifting that labor pool to very low wage areas like Africa in particular. There's been a lot of discussions recently about what is actually getting paid to the people at the end that are doing this work, classifying information, having conversations with the AI systems to train them. So there's quite a spread of people. But generally, there's a lot of people that simply see this as a means of monetizing your time, really. They get paid by the hour. They get paid by the number of conversations that they're having. There's a lot of incentive to produce a large quantity. of work as quickly as possible just to provide that massive amount of data that the AI platforms are relying on. And so, you know, anybody that sees this as an opportunity and anybody that's creative enough to figure out a way to game this system is probably going to be attracted to the idea of just being able to get in there and automate the process. I mean, in many ways, the AI companies were asking for this, weren't they?
Speaker 2Yeah, there is a real irony here that the AI companies are paying humans for genuine human data and then the humans are feeding it AI slop. So is it just that combination of, you know, pressures, people working externally under no supervision, the incentives to put a large volume of information and probably the lack of good pay? I'm imagining that all of these forces combine together to create the perfect environment for people to go, well, I'm just going to get the chatbots to talk to each other.
Speaker 3Instead. Yeah, I mean, I think this is one of those, I'd file this one under you get what you pay for. This does not seem like a particularly rewarding or good job. And it kind of, I imagine it's not, I imagine it kind of diminishes the work you would genuinely put into it to be like, okay, I've had this good conversation, which will now be put into the fire of this training model, right? I imagine that's like fairly demoralizing to begin with. It's all remote. And I can't imagine it's like a super healthy workplace with like career opportunities or like benefits, because the whole model is kind of foundationally,
Speaker 2extractive. So let's talk about AI inbreeding, which is a fabulous phrase, or model collapse as it's otherwise known. David, why does training AI on AI output make the models worse rather than just the same?
Speaker 4Well, the problem is that generative AI is really only an approximation of actual human knowledge and information processing. So by training model on its own output, basically these AI trainers are telling the system, what it's producing is already the same as what humans would do. So it's reinforcing the outputs of the system, which they're supposed to be improving. And that's not what's happening. Basically, it will de-emphasize the data that's in there that was actually created by humans. And if you think about over time, the AI models have been ingesting the whole of human knowledge. I mean, there's quite a lot of real facts that are in there, but if you start feeding it, its own interpretation of those facts and labeling that as human, you're going to create problems. I mean, when the snake is eating its own tail, eventually it runs out of tail.
Speaker 2Yeah. And if this is happening at scale, if we've got workers all over the world contributing this to, I'm sure most companies would be utilizing this kind of manpower or AI power, what does it mean for the models that we are all using over time? Should we, as the end users, be worried about the quality going backwards? Or if there's. There's a full model collapse. Is this something that needs to be on our radar, Fergus?
Speaker 3Yeah. I mean, the topic of model collapse is definitely really interesting. I think that there's a term, Habsburg AI, which a researcher called Jathan Sadowski coined a few years ago to sort of refer to the phenomenon. And I think the fact that you see these companies are specifically chasing high quality, authentic human conversations, that to me says that they do see, like, not necessarily risk of model collapse, but the benefits of real data versus synthetic data is pretty measurable. The thing that I have read about these companies is the way that they are trying to offset it is they're usually having another AI to double-check whether the first AI is being fed synthetic data and stuff like that. That, in addition to, like, the whether or not that's an effective approach, expensive, because you're automatically doubling the, like, amount of output, your kind of tokens you're using and stuff like that. I think that those sort of countermeasures are really, really important. And I think that's a really, really important thing to consider. And I think that those sort of countermeasures are something that these companies will probably continue to explore. But they're probably going to, the benefit, those benefits are probably going to be exclusively to the paying customers, the best of the best models and stuff like that. The free models, like the free chat GPT or the Google AI overviews and stuff like that, the stuff that's, like, not a premium product, they're probably not going to benefit from these safeguards. And I think the potential window for harm, if fears about model collapse play out, is much wider for the people who don't pay for this stuff.
Speaker 2David, can the company do something about that? Can the companies fix this? Or have they already built a system where all of the incentives are guaranteeing people are going to cut corners?
Speaker 4Look, I think people are people and human nature is what it is. There's always going to be this. And as was mentioned earlier, the whole issue, of course, is a lot of this training happens remotely. There's really no oversight of what people are doing. So the only time that the high level AI companies will even notice this is happening is when it's sort of things start to go a little bit weird. I suppose, over time. So I think they do want to fix it, because they always want to sell the idea that they're approximating human intelligence. I mean, that's always been the benchmark, human interactions, meaningful conversations, that sort of thing, certainly is on the table with the AI company. So I think they really do want to fix it. It's very hard to do, I suppose, unless you really spend a lot more time vetting the people that you're paying to do this and generate this human. sort of genuine conversation. But then that defeats the purpose. The cost is high. And that becomes a bit of a challenge. So I mean, they'll certainly have to work on fixing it. Some could say that the way to fix it would be to make sure that you increase the number of people having these so-called genuine conversations so that the outliers, the people that are doing this and getting AI to talk to itself are only a small proportion of the entire data that's going into the system. But as Fergus said, that does get expensive. I think this really is an example of how it's really important just to keep humans in the loop, make sure that you do have some kind of oversight of what's going on, particularly as this scales up. Because without humans in the loop, things can go really bad, or at the very least, very weird.
Speaker 2You are listening to Download This Show on ABC Radio National. And speaking of things going very bad, Meta had a big week. On Tuesday last week, Meta rolled out Muse Image, the first image generator from its new super intelligence labs built into the Meta AI chatbot. The headline feature, you could at mention any public Instagram account, and the AI would generate or remix images referencing that person. The catch was that it was opt out by default. Private accounts and under 18s were excluded, but any adult with a public profile was automatically in unless they dug into it. The settings to leave. There was much backlash over privacy and consent, including from a Hollywood union. Meta pulled it out on Friday, admitting it missed the mark, shut the whole thing down. Let's start with what Muse actually did. You tag someone's public Instagram, the AI generates images of them. How does that work, Fergus?
Speaker 3It really was as simple as bringing up your Meta AI screen, tagging your friend that you wanted to make an image of, typing in a prompt, and then maybe selecting one of several presets to kind of guide the output. These products are inherently very kind of straightforward because they don't want you to go off the beaten path. They want you to pick a preset, they want you to tag your friend and be like, look, I made a thing and hit share.
Speaker 2I made a thing of you. Without your knowledge or consent. This was only available for public profiles. Public has always meant that your photos are visible, but I think there's a big difference between people being able to see your photos and being able to use an AI feature to see your photos. People are able to use an AI feature to remix them into something new.
Speaker 3Yeah, I totally agree. I think that obviously Instagram and the internet has a really rich remix culture and stuff like that. But I do think that with Instagram specifically, there is a connection that people have. There's a sense of ownership you have over your feed and your photos. And while you do consume the things that other people post on the platform, I think that one of the reasons this tool was kind of controversial is it kind of flips that in a weird way where you're turning other people's content into something that you are producing.
Speaker 2David, could you talk us through why opt-out is such a problematic way to have this operate?
Speaker 4Well, this is pretty typical of the way that big tech is operated. We're going to roll out a feature. We're going to do this and this and this and then see what happens. And then people start complaining about, oh, by the way, this is a massive privacy violation. Or by the way, this is simply a way to facilitate large-scale plagiarism. And then people complain enough to the right people and they kind of walk it back a little bit. Opt-out, of course, is problematic because most people don't pay as much attention as privacy researchers and privacy advocates do to these sorts of settings. And people may not be aware of what's going on. Most people probably wouldn't even notice it until they maybe find a video of themselves surfing on top of a large shark or something, or they go to their sister's feed and there's all these photos of the time that she got on stage and sang with Taylor Swift.
Speaker 2And people did call this out. There was quite a large public outcry. Which obviously led to them pulling the feature only days after launching it. What do you think it would take for tech companies to do their due diligence before launching a feature? Is that even in the realm of possibility? Surely they could have predicted this.
Speaker 3I mean, really boring answer is like legal or regulatory pressure. But I mean, these companies do kind of, they do this because they know they can kind of get away with it. I think a lot of the time you see them launch features like this, and it is almost like a negotiation, like this is our opening offer. And if they get, heat for it they'll you know scale back and go from there that's what they pay the crisis people for um so And I think that the culture of how much can we get away with, that's something that you can't really solve for overnight. You read these books like Careless People and stuff like that, which provide insights into companies like Meta. And there are definitely people in there who do understand the potential problems of a product like this, but oftentimes they're either structurally disempowered or not in a position where those criticisms actually get heard until we're out in live territory, I suppose.
Speaker 2Yeah, I suppose it's kind of hard to consider that within the walls of an industry whose motto for the longest time was move fast and break things.
Speaker 4Yeah. It's important to remember too that Meta is a company that loves money so much that whistleblowers recently shared that it's actually making 10% of its revenues from actively selling scam ads to its users. Yes, yeah. So to suggest that they would check that or confirm that a new feature is going to be fine before they roll it out, it is probably a bit rich. This is very much a culture of just throwing spaghetti at the wall and seeing what's happening.
Speaker 2Okay, so this superintelligence lab, which has poached AI talent from all over and spent endless billions of dollars being established, is this what they're creating? Are they doing anything better than this?
Speaker 3Well, according to Meta, and you'll love this, they have a number of other AI features and interactions planned for WhatsApp, Facebook, and Messenger and another AI video tool in development. So I would expect more of them.
Speaker 2More of this.
Speaker 3More of this, I think, is what the plan is.
Speaker 2Great. David, tell me you've got some insight for some better news. Surely we can't be burning this much money to create a remix tool.
Speaker 4I have insight, but it's not better news, unfortunately. Oh, no. I would just point out that one of the big things that's been going on in the consumer tech sphere in the past few months is the launch of these smart glasses with cameras. Meta is, and of course, they're not the only company that's doing this, so we shouldn't just say Meta, but we will. They've launched a number of new, smart glasses that can record video, take photos, basically anywhere that anyone is. That is a rich, rich, rich mine of new images, new information, new videos. And I think anybody who believes that that content is not going to end up somehow being remixed or redistributed and modified and reused for some way by Meta is probably kidding themselves. So it's only going to continue like this and probably even worse. David, tell me some tech that
Speaker 2has brought you joy this week. I require some balance in my life, please.
Speaker 4Well, this is a good one for balance. The reality of my nerdness is that I'm prone to getting very excited about the ability to control stuff around the house from my phone, which I think a lot of people have gotten into, but I was delighted, I suppose, recently. We had to put some blinds in, some roller blinds. This is a very mundane application, but a fun one. No, this is the dream. The dream. And yeah, so we had to replace some very old blinds and said, why not go to the forehog and put in the ones that are remote controlled?
Speaker 2Yes. I was hoping this is what you were going to say. I'm so happy for you, David.
Speaker 4This is the one. So we got the remote controlled blinds, which are very nice in themselves, but the remote, you've got to find a remote. You know, it's very hard. You leave it somewhere, you put it in the couch or whatever. You've got to find the remote to be able to do that. So of course, I thought, well, what else could we do? Turns out what else we could do is there is a box which you can use to connect to all the blinds, which is online and connects to things like the Google sort of display that's in the kitchen. So I now have that set up and I can walk down in the morning and all I have to do is say, hey, Google, open all the blinds. And all of our blinds open up. It is fantastic.
Speaker 2And anyone listening at home who has that exact same setup has just had all their blinds open. Thank you, David. That's wonderful. Genuinely delighted for you. Fergus, tell me some tech has brought you joy this week.
Speaker 3Tech has brought me joy. I bought a device called a Steam Machine.
Speaker 2You bought a Steam Machine?
Speaker 3I did. This was a sort of PC announced late last year. It's sort of a cube shaped computer. It's by a company called Valve, who. I don't know if you've heard of it. They're a company who sell lots of PC games online via a platform called Steam. And they released a handheld gaming device called the Steam Deck a few years ago. But the Steam Machine is their take on like a console sort of experience.
Speaker 2Now, I do know the Steam Machine has received a little bit of criticism for a very hefty price tag. So, look, this question's for both of you at this point. Are you both secretly rich? What is happening here?
Speaker 3My secret is I haven't bought a PC since 2019. Okay, okay. I've been budgeting for such an occasion.
Speaker 2So, have you found the Steam Machine is good value for money? Yeah. Especially considering building a PC from scratch at the moment is cost prohibitive for most of humanity.
Speaker 3Yeah. It's absolutely delightful. I'm thrilled with it. It's like the form factor is incredible. It's silent. I'm really excited to get into the customizable faceplates. The performance is exactly what I'm looking for in terms of running. It runs basically everything I want really well. I don't need to crank everything to 4K and be crazy about it. I just wish to play more things than the Steam. Deck or my old laptop could support.
Speaker 2Beautiful. Oh, I'm delighted for you both. Fergus Halliday, a journalist at Whistle Out Australia. Thank you so much for joining us once again.
Speaker 3Thank you for having me.
Speaker 2And David Browey, tech journalist at The Information Age. Thank you for joining us for this episode of Download This Show.
Speaker 4Thank you. Always a pleasure.
Speaker 2And now for. some more tech news from the week that was. In case you missed it, the Australian government announced this week the establishment of the Office of AI. It's designed to coordinate legislative responses to artificial intelligence that concern the technology's growing economic and social impacts. The PM claimed that Australia is the first country to bring these issues under a national framework. One of the first issues the office will deal with is copyright protections. Prime Minister Albanese had this to say.
Speaker 5Thank you. Australian writers, musicians, artists and journalists must retain ownership and control of their work. Our laws will spell that out, plain as day. An artist's creative endeavour is their work and their property. No company should use Australian books, music, art, or news to build or train AI without the artist's control. And that includes the artist's control of the price and value of their work. Anything less is theft.
Speaker 2The other big issue Albanese addressed in his announcement this week is data centre growth. He intends to work with state governments to legally bind companies to cover their own energy supplies and minimise water usage. Any legislative changes will not be introduced until early next year. The interim time will be a consultation period between different levels of government, industry bodies, companies and stakeholders. And that's all for this week. Make sure you subscribe so you don't miss any future episodes. We are working on a special episode that answers all of your technology questions. I've seen them coming into the inbox. You can join them at [email protected], especially the questions you're afraid to ask. Reach out to us. Or you can slide into my DMs on Instagram @rayjohnston. Big thanks to my producer, Jessi Kaye, for this episode. And now each year, ABC Radio National hosts media residencies for academics in science and humanities and the arts. It aims to enable early career PhD academics to better communicate their work to non-academic audiences. Dr. Loren Ruster is a senior research fellow at the ANU Complexity Leadership Lab School of Cybernetics, which is incredibly cool. And she is one of those residents. Loren researches and develops responsible AI practices in and with organisations centring human dignity. Loren researched and scripted this episode. This episode was produced on the lands of the Gadigal, Baramatigal, Dharug, Gundungurra and Wurundjeri peoples. I'm Rae Johnston, and you've been listening to Download Thiss Show. Speaker 3You've been listening to an ABC podcast. Discover more great ABC podcasts, live radio and exclusives on the ABC Listen app.
PodcastTranscript AI