Generative AI is a Crisis for Copyright Law

Ben Zhao 

Hi glad to be here. 

Hessie Jones 

 I’m really grateful to have you today and to address some of these issues because what we’re seeing today, even politically, is going to probably have a bearing on how a lot of these technologies evolve. So let’s start with you. Your current focused primarily on security and privacy issues related to the misuse of machine learning models and systems. Tell me about your background and your motivation for your focus on this work. 

Ben Zhao 

Sure. Well, let’s see. Been a professor of Computer Science for just about 20 years or 21 years. It’s hard to count now. But for quite some time, maybe the last decade or so, I’ve worked on computer security issues, primarily as my focus and for most of that time I’ve been focusing on, generally speaking, protecting AI and building, you know, tools to make. Make it more difficult for bad actors. Now the context, of course, is that for, I don’t know, maybe seven of those years, AI was something different than what people refer to AI today, which is to say that AI used to be primarily focused on image classification and what are sometimes called models that look at content and they recognize certain things. They tell you if there’s potential tumour in a particular, you know MRI or X. They tell you. When the car is driving down the road, what that, you know, thing is coming up ahead. That a. Is that a car? That an animal. And so those are, you know, unambiguously good products that improve our lives. So I spent a lot of years protecting those kinds of tools. Since 2022, I would say, or maybe a little bit before then, I’ve been turning more my attention into looking at things, particularly AI world, that are not so great misuses. When a is availability and. You know, reachability becomes ubiquitous and global. And can get access to the. Inevitably, those bad actors, there’s bad uses of. And so we’ve looked at things like how to disrupt companies that train face to recognition models on you without your consent. Your knowledge even and then since 2022, we’ve really turned our attention to looking at protecting human creatures who have been. I guess largely the the sort of at the brunt of these attacks using generative AI, whether it’s mimicry or or other types of things. And that started just because, you know, I think we had some prior work in protecting human facial recognition out there and artists reach out. Us. Trying to get us involved, trying to ask us for help once we learned about what these genomes Act tools were already doing, we sort of jumped in head first. You know, the rest is kind of history, I guess. 

Hessie Jones 

So you have a certain philosophy about what’s going on today from a misuse perspective and you actually talk about this imbalance. It in the Internet. Can you speak a little bit more about that? 

Ben Zhao 

Sure. You know, actually goes back decades. The last time we had such a disruptive technology in our lives that changed everything was. Tonight and interestingly, one of the sort of the the the architects of the Internet, Dave Clark at MIT, you know, had this very interesting paper that he wrote about 20 years, maybe 30 years after the initial conception of the Internet. And he. About how you know. It is challenging to look at stakeholders who are at tension at all each other with respect to different things. Maybe it’s. You know, copyright. It’s privacy. Maybe security these kind of different, you know, Battlegrounds where different, you know, stakeholders disagree. And he talked about. Why it’s so important to sort of allow both sides to to sort of, you know, in a reasonable way on a level playing field resolve their differences and come to some. Of conclusion. And that was important for the Internet to sort of thrive and to avoid some of the worser outcomes. But in AI today, we do not have that, and this is important because when we look at copyright law, when we look at regulations and when we look at, you know, how businesses are are managing these types of sort of tensions. One thing that’s quite clear is that companies who are the big stake holders in this space, the the big tech model trainers, are using sort of tactics. I guess I would say that are not entirely. Normal. You know, I’m referring to things like effectively threatening um. Individual countries or entire regions with, you know, loss of tax revenue, should they be regulated, right? What is right? Is. What is legal? Is illegal. Or not. 

Typically, concerns that are sort of decided based. Tax revenue. And yet this is what sort of. Dealing with today. Political lobbyists is one thing, but you know, basically saying that if you regulate us, we’ll exit with our billions of tax revenue. You need to change your law. It’s something that I don’t think we’ve ever seen before. So I come at. Not some somewhat from a sort of super philosophical standpoint. I mean, although I think that it’s important for these stakeholders to resolve things in a in a neutral way. But I come at it. More of a humanistic standpoint where I look at. Artists I look at writers, I look at, musicians, I look at entire creative industries that are on the verge of not, you know, necessarily just disruption, but. But essential crisis because of what? Tools are doing because of. Not necessarily because of what these tools are capable of doing, but because of what people are doing with these tools in a in a sort of a unintended way. And so these negative harms, I think, are what I’m trying to mitigate and trying to protect human creatives from. 

Hessie Jones 

Yeah. I I think you’ve eluded to the imbalance, mainly from the power struggle. I think part part of it is because. Power perspective because of the amount of pervasive adoption that these technologies have had in each of the respective countries. I would say though that Australia is a country to look at to see what. Impact it has on its current law and ousting. I wouldn’t say ousting, but not allowing Instagram or Instagram or Facebook. To to. To work. No, sorry to allow 16 year olds to actually adopt the technology in in Australia. I’m hoping. I’m hoping at some point that that is going to have a material effect. And can be adopted if you know if the outcomes end up being positive, so. 

Ben Zhao 

Yeah, it’s entirely possible that they can, you know, enforce these laws. 

Just have to have legislators and regulators who actually. Do the job. 

Hessie Jones 

Yeah, I mean. That is an issue when it comes to the legislators and regulators. ’cause, as you know, regulators, they, they tend to lag behind technology and unfortunately it ends up being like more of a whack-a-mole effect. But let me let me. Move to this idea of adversarial machine learning. Because this has been, I guess more of a solution that you’re using in helping to combat some of these harms. You talk a little bit about that. For audience, just define what is a general. Sorry GANS and I’ll let you go for it. And then and then talk about some of the technologies that you’ve used in the meantime. To help. With the solutions that you’re talking about. 

Ben Zhao 

Sure. Obviously machine learning is just effectively speaking, the intersection of computer security with machine learning. So anything that has to do with the, you know, attacks against these kinds of systems. Defenses protect them. Privacy issues, misuse issues, abuse issues, all these questions. Anything to do with issues of security or privacy when it comes to machine learning is what would fall inside that bigger space of. Machine learning. You know, so typically speaking over the last 10 years, I’ve spent a lot of time building tools that protect these kind of models. But at the. Time you know, these things are always, as they say, a a double. Edged sword, right? And fundamentally speaking, one of the things that has really been learned by the this, you know average year machine learning community over the last 10 years because it’s really been now 11 years I guess 2025. That we’ve been looking at sort of fundamental questions of can these systems be protected? One of the things that we’ve learned is that. AI sees things so fundamentally different from humans. That there is a fundamental gap between how. 2. Of you. How content is perceived or reacted to by humans versus? AI and that is a gap that is not closing anytime soon. What that means is that there will always be, fundamentally speaking, these methods where you can create things that look fine to a human but completely break AI model. Now for in some respects that’s a bad. Because you know, for attackers, this means that you can sort of take these techniques and and trick, you know image classifiers into doing it wrong. Things and that’s what we’ve been working on for almost, you know, quite a number of years. But you know, in some respects it’s also a bit. An opportunity? Because what it means is that we are able to develop tools that preserve the look and feel content to human content consumers. And yet. Make them completely disruptive to AI models in such a way that AI models that train on them will have unintended consequ. And that’s really what plays a nightshade do. 

Hessie Jones 

Can I ask you, just from a simplistic perspective, when you say human sees things differently from machines, so speak to me about an image. How would a human? I know how a human would look at an image, but how would a computer see that same image? 

Ben Zhao 

Sure. So let’s you know the classical example that first demonstrated this potential problem was. This picture of a panda hanging off of a tree, or assume a bamboo trunk back in 2013. And what the initial research showed was that you can basically alter a small number of pixels and do so in such a way. Was very. Of premeditated and and you know. Algorithmic. You can compute these exact pixels in such a way. That the image basically looks untouched to humans and yet to an AI model that was defined and trained to identify a panda versus other animals, it would see it as a very different animal. And I think the classical example is it sees pandas as Givens. Now, that sort of example, called an adversary example, is fundamental to image classifiers. That is to say, we have been working on that. You know, 10 plus years and basically the community has largely come to agree that these things are fundamental. Yes, you can mask. You can reduce their probability, but generally speaking there’s no. Fundamentally, there’s no fundamental. To get rid of them. So AI sees things by extracting specific rules, specific patterns from it. And the idea is that you know. Is there a question? There an. Is there a way to sort of build these rules that completely simulate the human brain, though? There’s not, and So what? AI derives in terms of its quote UN quote. You know, understanding that sort of anthropogen. IC but really it’s just extracting a bunch of numbers that govern what it defines to be a panda versus a given. But what does that? 

Is that those rules are always imperfect that those rules always have crevices. And sort of, you know, untrained regions where you can exploit if you’re someone who’s trying to mislead the model. Do something. Into something that’s incorrect, right? So. Same idea of these AI models never being able to match the human perceptive, sort of you know, dimensions, if you will. Pervades through almost all machine learning, and so if you look at gender pay I today you still have the same issues where you know when they trade on content, they trade on content for very different things than what humans see. And so again, understanding what those rules are. Way that those rules are sort. Computed inside AI models. As long as you understand. Those rules are you can exploit them. You can manipulate. You can get the model to do things that you know they weren’t initially designed to do. 

Hessie Jones 

1. I I want to divert a little bit because I wanted to ask you based on what you just said, the opening is new reasoning model. What does reasoning mean to to open AI versus? This is what it means to humans because. I doubt that it I I assume it’s still a probability. Score.

Ben Zhao 

Yeah, of. You know the funny thing. Not so. Funny thing, but you know the the truth about much of generated today is that so much is marketing hype. And so people have normalized the marketing hype with terminologies like reasoning, right? What does reason actually? Reasoning typically means that you’re using some sort of propositional logic, some sort of a = b = C, then a probably equals to C. We’re taught these basic things in like grade school math. The problem is that these models can’t even understand that a = B means b = A. Right. So there’s no reasoning. Is no logic. And when people use marketing words like reasoning, they’re referring to something entirely different. Not sure entirely what. I I do know that you know. Generally speaking, these models are very, very different from how humans work. They are exceptionally good mimicry machines, you know. You. I’m sure they old enough, but you know some of your listeners might be old enough to remember. Back in the days of Eliza chatbot. Which is the same from the 1980s, I believe, and all it’s meant to do is pick up on certain keywords and ask you pre programmed sentences. You know, if you think about that and if you think about that being much, much smarter, you know, much more powerful, many more prompts, many more sort of different ways of asking send sentences, you get something. Of like a, you know, like an Ln today. And you know. There are still fundamental things it just cannot do. Guess. Know the classical example on Twitter was a few months ago. Were fiberglass, but it’s completely, you know, reasonable. When people ask you know these L NS, who’s Tom Cruise’s mother? It came up with the perfect right answer. If you took her name and. Said. Is, you know, insert her name here. It will be utterly confused. And people were surprised by this. Because I assumed that you know these things actually had reasoning. They don’t. They’re asymmetric. Right. It’s perfectly reasonable behavior. You think about. Fact that most articles are about Tom Cruise, not about his. And so how many examples of training data do you have that say Tom Cruises life is like this and this and his mother is so and so, but there’s very few articles that talk about his mother in the first person or third person. Says stone. Is. And so that explains the discrepancy in in some of these. And yet, you know so much about how we think about these models and and honestly test these models. Is sort of us projecting, you know, human values and human processes onto these machines? Is classical sort of psychological anthropomorphism, which we all understand well. In fact, you know the easiest way to explain this is to, you know, mention to anyone who has seen. Movie, you know. Oh God, now it’s like. 

Hessie Jones 

Who is the? 

Ben Zhao 

No, no, it’s the Tom Cruise movie. Sorry, the Tom Hanks movie. 

Was it not shipwrecked? Umm. 

Hessie Jones 

Oh. 

Ben Zhao 

It’s it’s the one where he stuck on an island by himself for like years, right? 

Hessie Jones 

Yeah, and that movie is a perfect example of this because you know, Tom Hanks is stranded on an island by himself for years. What’s that? 

Hessie Jones 

Castaway. 

Ben Zhao 

Castaway, thank you and can castaway. So he’s stranded on an island for years by himself. The only thing he has to keep him sane is a volleyball. He names it. He makes it a face and he talks to Wilson and Wilson keeps him alive. Right? But. Effectively speaking, like he can think about all ends as a. More interactive version of Wolverine. And what unfortunately people don’t realize is that so much of what we are doing with these islands has been we are projecting ourselves onto them. I think someone described it as are like mirrors reflecting ourselves, you know, at back at ourselves. That’s really, I think, much more accurate. Than people think and yeah. 

Hessie Jones 

No. I agree. So we’re going to go back to the actual topic because I, want to. Really focus on your two solutions. Tell me. About glaze, I guess that was the first one that had launched. That. And then and then tell me about Nightshade and how you actually developed them to help the creative community. 

Ben Zhao 

Sure, Glaze and Nightshade are similar in the sense that they are both models that take advantage of this gap between humans and AI, right. And and how we perceive things very differently. But their goals are very different. So glaze in particular is specifically designed to disrupt. Model fine. What that means is that many times for you know, oftentimes the use for a diffusion based image model. Right, you know, stable Diffusion, SC excel or some of these other models. What many people do is they take this model. They see some artwork that they like from a particular artist, they say okay Trey on this on these particular samples and basically to the point where you can mimic this particular artist. And of course, you know, I’m going to wear you like a skin and control you like. Puppet. And I will be able to generate art images that sort of look like the artwork was done by that particular artist. But of course without. Knowledge without their consent, without, you know, compensation. And so that’s what glaze prevents. Glaze allows artists to run, glaze on their own images of. Own artwork. In such a way that the artwork is altered in very very subtle ways. You know that we actually change something like 80 to 90% of the pixels on the image, but always in such subtle ways that even artists, when they look at side by side, blown up images of their own work. Many of many times they can’t recognize which. Is the before and which. After, but what happens is that the alterations make it such that the AI model trying to learn about the art will learn something very, very different from reality. What humans perceive. So someone trying to replicate and mimic a an artist who uses charcoal very dark. Portraits will get something more like, you know, sort of high intensity color either, like either like a Jackson Pollock or or Picasso. And if we can do that, then we have succeeded because then the person who’s trying to mimic this particular artist will have failed. They’ll leave that artist alone, hopefully. Nice shade, on the other hand is quite different. Glaze is meant to sort of only disrupt, you know, third party, sort of, you know, minimal training, additive training on top of existing models and you know, glaze only protects individual artists. Is meant to sort of. Shift the power balance if you will, for most or all artists. Basically, by allowing artists to do the same thing as glaze, except instead of changing their style. Change the composition of the image. So now imagine a photographer who has. You know a photo of, let’s say a A. You know, I don’t know what. Let’s say a dog. You can run it through night shade in such a way that night shade will take that image and again make very very minimal changes. You can’t see the difference before and after. But what I’ll. Is to change the compute in such a way that an AI model. On an image. Will see all the sort of visual features of a cat. OK. And you know, night Shade is designed to be optimized so that its images are extremely convincing, potentially more convincing than normal images to these models, so that the models of course they had no idea what really a dog looks. Or a cat. But after seeing some of nightshades images, they’ll be convinced that in fact a dog has, you know, big fluffy whiskers, you know and and, you know, basically looks like a kitten. And So what Nightshade does? Is it gives away for artists and other usual artists? In general, to push back at AI models. It. I usually say to. Something like adding hot sauce to your lunch. Right, hot sauce. That doesn’t isn’t particularly spicy towards you, but you know it’s spicy towards everyone else. So if someone steals your lunch, then they get into trouble and you know, maybe they end up going to the bathroom all day long. But that’s sort of, roughly speaking, what Nightshade does. And it’s meant. Increase the cost of training so that companies will be much more incentivized to actually pay artists a licensing fee to properly license their content instead of just downloading and scraping content to train on. 

Hessie Jones 

So let’s assume that and you had mentioned this earlier about the technology being a double edged sword. And so let’s assume that. What is happening today is it’s a lot more pervasive than than just what nightshade or glaze can, can I say, defend are the. What do you say to the critics who could argue that you’re actually potentially harming the development of beneficial AI? 

Ben Zhao 

I would ask what beneficial AI so the reality of it is that once you look past the marketing hype. And you say, OK, give me tools that can generate text. Give me tools that can generate images. But the text generator hallucinates randomly, unpredictably, without control, makes up facts. And the image generator is, you know, liable to produce images copyrighted by other people. And again in time and is not copyrightable. This is a reality today, so I would then say great. Now show me the revenue producing applications, the ways that these applications can really help society for the better and. Think there would. A long pause because I’ve had this conversation before with a number of people. Before they can come up with something and you know the answer usually ends up being something like, well, you know, I can replace customer representatives with this. I can replace some human with this because you don’t need to talk to humans anymore. OK, you know, there’s such a thing as efficiency and and if customer service agents is what you’re looking to replace and save some money, OK, I get that. Except even in. Instances there are already recorded cases of these chat bots saying things that you do not as a company want them to say. 

I think it was Delta Airlines that had to honor a tremendous ridiculous discount simply because of chat. Bot decides to hallucinate it to one of its customers. Now you know that may be a one off thing. But imagine this in a safety critical scenario where you’re depending on this for a translation of legal proceedings, or, as you know, a recent paper showed, you know, even transcription of conversations between a doctor and a patient can be hallucinated. And so, you know, then it really begs the question of who dares use this technology? And to do so S scale right. And then you know, of course, somebody would say well. No, but we can fix this. But they’ve been saying this for quite some time and I actually understand that architecture behind these things and they’re not so easily fixable. So yeah, I beg to. The the questions of how these things are fixed, I think it’s a very, very difficult task and I think fundamentally they are actually tied to the architecture that they’re on right now. So unless we actually. You know, break away and do something different. I think they’re here to stay. And that’s going to be a serious problem for any company trying to market it for real. And perhaps that’s why these companies are losing billions every single year. 

Hessie Jones 

I I think I saw a a note from the Microsoft CEO Nadell. What was his name again? 

Ben Zhao 

Satya Nadella. 

Hessie Jones 

Satya Nadella and he had actually questioned the near monetization of these. That often, and I mean considering their investment in them, he openly argued that we’re not going to see any real ROI anytime soon. So that just speaks to I think. I think adoption is a huge issue right now. A lot of companies are still looking for that accuracy rate, but at the same time, a lot of the harms that you mentioned are are kind of putting people and companies on pause. With respect to the technology now. Want to? You about something that you said? About generative AI models plateauing in their ability to correct to to offer correct and sound information. So I just wanted to quote you on this. They’re plateauing. Because, as we often time, saying computer science, getting the 1st 80% of something something working correctly is relatively easier than getting the last few bits. Can you expand a little bit on that? 

Ben Zhao 

Sure it is. Speaking in computer science, relatively easy to get something mostly right. Getting that last little bit, getting those last few errors out is extremely difficult because they tend to be corner cases. Tend to happen less. They tend to happen in unexpected ways. In this particular scenario, you know trying to get LMS to not hallucinate. Is going to be extremely difficult because these hallucinations happen at the corner edges. They happen in the corner cases where things interact with things unexpectedly and to try to root all those. Really require something much more dramatic than what people are doing today. To try to improve accuracy. Yes. Know people are trying all sorts of methods, everybody. Every company is throwing money at the problem, you know, but there’s a real question of whether there will be a reasonable solution to actually making a model that becomes more accurate, right? Say there are other more feasible solutions that basically bypass the model. Are now feasible solutions where? You know what’s commonly called retrieval? Generation, which you. In common terms, just means that you are adding facts, you’re adding other additive information to these models. In real time to try to help them fix things. Taken to extreme, it can mean that you basically hollow out the internals of LMS so that they are no longer answering you, but they are basically a translator and you ask LM something and the LM. Basically takes that conversation and codes. It, translates it to an actual. Query to. Database something that we. For, you know, quite a few decades, things that actually have real, concrete information that we can trust and understand and then basically have the results from those kind of database couriers then hit the LM again, be translated into something much more conversational, much more easy to understand and. Be fed back to you. So those kind of techniques? Interesting. And they might work, but at that point you’re not really training LMS to do anything. No reasoning there’s. Knowledge of any. It’s a fancy translator that, effectively speaking, transcribes your conversation into computer language and then back again when the results come out. I think that’s what it does a better job at, and maybe that’s the viable way that they become useful. But at that point, nobody’s going to take these, you know, ginormous models and pay you millions of dollars. A fancy translator. In fact, they can probably already do that today. But you know. The value is in the actual content. And the real data that’s inside these databases. 

Hessie Jones 

Is there a risk that somebody? That takes these hallucination outputs as fact can be in fact used, fed back in the model and then later on. Because of the reuse, can be depicted as truth or depicted as valid at some point in time, and then and then risk not really knowing the difference between really what is factual and what is not. 

Ben Zhao 

That’s an interesting question. Your question starts with the assumption that these models can do that to begin with, which is, I would argue, not true. And so these models have no idea what is true or not. What they do, see what they do quote UN quote. Know is that they see tremendous amount of content. And they will basically mimic things and say I can speak this with high confidence because I saw this in, you know 20,000 documents that this is true. So therefore must be true. You already have examples of, not necessarily just a random hallucinations, but actually bad content showing up inside these AI. So the example of Google AI overview telling you that you know some hilarious examples and everything from if you’re pregnant, it’s absolutely safe to smoke. If you know should you eat rocks? Every single. Yes, they’re great for. You should eat some every single day. And then you know. Turns out, oh, even the classic, which is, you know, you can put glue on pizza. It’s a topping. Right to hold other toppings together. Some of these things are you. Where they came. They came from Reddit, where people make jokes, where people make up sarcastic comments, not intending them to be serious. These Amos have no idea what’s real or. So they see that being quoted and if you ask something that is akin to content that covered by Reddit, it’ll. Regurgitate what Redis said, so there’s a perfect example of it operating correctly but giving you false information because there’s lots of false information on the web already and it has no idea what’s true or not, right, because it has no ground truth. 

Hessie Jones 

I I think that the the real risk and I was speaking to my sister about this last night because she’s a teacher and for a teacher that is used to creating, let’s say report cards where you are to provide verbal feedback on each student. What she’s saying is that I could just go to ChatGPT and it will give me a more nuanced. Feedback for one student, so it sounds different than how I than how I described Johnny versus versus Penny. And I said, but that’s not fair. And I said, what are you finding? Goes it’s so easy. She and she. But yeah, do I feel. Yeah, but do I feel like I’m getting more done? At least I’m going home. Before 7:00 at night. And that’s what scares. Is that the degradation of human cognitive ability? Do you see that same thing happening over time? 

Ben Zhao 

Oh, absolutely. It’s happening everywhere, right? You know. The. So so I always refer back to this little story that I heard when I was a kid. I think when I was in, like, grade school, you know, there was AI don’t remember whether was a cartoon or a book or something. There was a story about the magic. Pencil the Magic Pencil was this magical thing that you know, for school children. If you had one, you would basically talk to it and you say do this problem for me and it would read the problem and it would do it and write. Answer down for you. And it was magic. And you know, every school aged kid was like, Oh my God, if I. Had one of those. But the funny part was that that story was a warning. Story was a sort of what not to. It was the, you know, do not surrender your intelligence and your learning process to an automation tool that that skipped all the steps that that learns for you, quote UN quote. But really it just leaves you dumb and without the benefits of actually having learned. And basically this is where we are. We have a magic pencil for everyone journalists. You know, creators of certain kinds, you know, educators, people in school, students who are in school. We all have this choice now and it’s amazing that I think very few people actually saw that allegory, warning them not to do this, but everyone is whole, all in down on, you know, how do I save some more time? Do I, you know, make my life? Little bit. Yeah, that they can actually do that, but. You know the the downstream effects of that are are severe and I would say for a teacher to do that. You know, I can understand a lot. Teachers are overworked. You know ginormous class. They’re they’re stretched already to the limits. I don’t. I don’t fault anyone for trying to, you know, go home. A reasonable hour. I think the bigger issue is for students who are supposed to be learning who are supposed to be, you know, getting something out of these lessons and when they reuse this and. Of course, using this at incredible scales at just about every school, every university high school I know of. And what we’re losing is the value of education. And so people say like, well, you know, this is. Like. Calculator it’s not really. You know, a calculator is. You can open it up, you can test it. Will always produce the same simple results, right? Doesn’t replace people. These things do. And they’re fallible. And so yeah, I think. Very soon, we’re going to have generations of, you know, young people who, despite having gone through school, basically have learned very little of what they’re supposed to learn. And then when these. Are now in charge of the next generation. These kind of tools they will. Nothing to contribute. They won’t be able to debug these systems when. Prove. They won’t be able to understand how they work. Yeah. So there’s a there’s. Real problem here. These tools can disrupt our educational system in very, very fundamental ways. And unless we take some steps to help students keep learning, yeah, this is going to be a really, really big problem for. Down the road. 

Hessie Jones 

Yeah, I can see that as well I will. I want to shift but you know address a broader issue that that comes with the development of these systems. Systems and you live in the USI live in Canada. We’re both kind of feeling it. You have a new administration and what we’re feeling or tariffs, but what you’re also feeling are significant changes. Within. Within your country, specifically from a tech perspective, there are. There are deregulation that is specifically benefiting big tech in the US. You talk about this imbalance and we know that we know that a lot of the deregulation is favoring a lot of what big tech is doing. So from that perspective, that is worrisome. Especially when we’re developing systems, as you say, that are not perfected yet and that we’re increasingly relying on and could harm potentially the next generation, so. We we’re not seeing the guardrail. And I’m also seeing things like. For example, AI models within sandboxes that are able to bolide their own makers and come hell or high water try to maximize their own goals, irrespective of the intentions of their makers. Are we seeing from your perspective in eventual loss? Of human control in our systems and what do you think the political landscape is doing to enable that? 

Ben Zhao 

Yeah, that’s a big question, I mean. So in one way I think 1 interpretation of your question is more about AGI and and sort of questions in all that direction and. You know to. Type of. I would say I don’t feel that’s coming. You know, understanding what I do about these architectures and what these models and what they’re doing, they’re extremely good at mimicry. But there is no intelligence. There is no intention of any kind. So. To the extent. People are like, oh, no, it’s lying to. Well, it’s the lying is probably our version of anthropomorphism, something that is just simply a sort of backup output when the primary one has been told is wrong. So for many of these types of sort of, you know, human behavior, traits that have been assigned to LMS, I would say there are usually pretty good explanations for this. And nowhere you know do I see real proof that these things are gaining any center sentience or or any sort of intention, so gain control. Well, you know, I I think the control here that we’re losing is not so much towards AII think the control that we’re losing and that perhaps we were voluntarily giving up is towards money. And really it’s about. You know. Aspects of human society, whether it’s creatives, whether it’s the news media, whether it’s truth as we know it online and in real life, we’re giving up control from all these other respects. To AI in the sense that we are now willing to sacrifice these aspects. For the promise of some potential tax revenue down the road, I think that is the the sort of the surrender that’s happening in real terms. I don’t believe that there’s any sort of AI that we’re dealing with. Is, you know, intelligent? And I don’t think you’re ceding control to it, but I do think that, you know what’s actually going on behind the scenes is a shift of control and power to tech and. Specifically, those with extreme resources. Right. And if you look back again, you know, I think many of your listeners like me probably are old enough to remember this, but in the early 2000s, when we had the copyright, you know, legal battles behind music sharing and peer-to-peer networks. This is. Common today of what we have between copper and law and these kind of issues and creators. And what happened back then? Except for one critical difference? Back then, the stakeholders that were being harmed by this technology okay were Mpa RA, you know, recording companies and so on. They were the ones with the resources. They were the ones with the army of lawyers. And lo. Behold, when they were hurt, they brought those resources to bear and boom. You know the right thing happened. You. Copper Arrow was preserved. You know, the companies that violated the law. Largely speaking, no longer exist. OK. The only thing that’s different today is that you know those with the resources are now the ones who. Are allegedly breaking copyright, breaking the law, and so there is no other army of lawyers, you know. Yes, there. Lawyers who are suing. There are class action lawsuits, but they are representing groups who have significantly less resources than those who are. You know, the defendants. And so really the the power imbalance is completely switched over 180 and so, you know, will the. Things happen. I still. So, but this is really dictating what happens in the legal courtroom today. It’s the resources. And hopefully you know the resources will have a limited impact on what is right and what is wrong and the right parties will prevail. 

Hessie Jones 

Hoping that that’s the case. How long is it usually for this administration? For. That there could be a potential shift in two years, is that right? 

Ben Zhao 

Usually there’s a, you know, sort of congressional elections that happen mid mid term and usually what happens is that. You know, there’s almost always a midterm shift against the current president in terms of political parties, but we’ll see. Happens. 

Hessie Jones 

We will see. Well, thank you Ben for coming. I wish I could talk to you a lot more because there this is a huge issue and I I wanted to spend a lot more time on on the effects of glaze and nightshade. Think we’ll probably see a lot more of that in the coming years. Thank you so much for. For contributing to the creative and artistic community. 

And for everyone else. Thank you for joining us today. If you want us to explore other topics, please e-mail us at communications at altitudeaccelerator.ca tech uncensored. Powered and produced by altitude Accelerator, we’re hosted on Spotify. You can find us wherever you get your podcasts. Until next time, I’m Marci Jones. Please stay curious and be inspired.