How understanding AI today will help you make an impact tomorrow

In this episode of the Decoding Innovation podcast, Henry Ajder, a globally recognized artificial intelligence advisor, maps out the emerging landscape of AI and its opportunities, threats and benefactors.

Related topics

Listen now on Apple Podcasts, Spotify and Google Podcasts, and watch on YouTube.

While the concept of AI has been a subject of speculation for years — even decades — its significance, at least for the layperson, has elevated in the past couple of years. With the emergence and growing popularity of platforms, such as ChatGPT, Midjourney and Stable Diffusion, companies are eager to position themselves at the forefront of the AI renaissance.

As the relevance of AI increases, concerns and doubts associated with it grow as well. While it's undeniable that AI is capable of advancing industrial operations and human lives, it has also raised questions regarding the potential consequences of entrusting machines with power. This has given rise to debates surrounding privacy, consent validity, ethical considerations and potential for its misuse.

Henry Ajder, a globally renowned figure in the world of AI, generative AI (GenAI) and synthetic media, explores the advent of AI and the prospect of artificial superintelligence (ASI) and discusses the commercial implications. Further, he delves deeper into the ethical questions the technology brings forth.

Key takeaways:

Debates persist whether AI can be classified as true intelligence, considering it parrots to the users what they want to hear, based on the massive amount of data is has been trained on.
A large number of AI tools are readily accessible free of cost — some companies leverage this to build their reputation and user base. As there's a growing concern regarding how AI is monetized, several platforms have started embracing subscription-based models.
With tools capable of voice cloning and generating deepfakes, there's a growing need to create new legislation to curb the harm vectors posed by AI.

For your convenience, full text transcript of this podcast is also available.

Announcer

Welcome to the Decoding Innovation Podcast series, brought to you by the EY-Nottingham Spirk Innovation Hub, where we explore the innovative technologies, business models and ideas that are shaping the future of industries. During each episode, we meet with stakeholders at the cutting edge to discuss innovations in their space, challenges they need to overcome, and their outlook on the future.

Mitali Sharma

Hello and welcome. I'm your host, Mitali Sharma. And today's topic is artificial intelligence (AI), and to cut through the hype and the noise is our guest, Henry Ajder, who’s widely recognized in the field of AI, generative AI (GenAI) and synthetic media. Henry, welcome to the show.

Henry Ajder

Thank you so much. Pleasure to be with you.

Sharma

Henry, would you mind starting with giving us a little bit about your background and your journey so far?

Ajder

Certainly, yes. I really refer to what I've been doing in my work for the last six years as AI cartography. That is, I've been mapping this very fast, emerging landscape — the opportunities, the threats, the stakeholders involved, who's benefiting from this, who's facing challenges from it, how is society, business and individuals responding to this pretty disruptive technology, and the way that they live their daily lives.

Sharma

Thank you for that. If we can start with some basics around AI, because different people have different definitions of AI. How would you define artificial intelligence?

Ajder

Definitions around AI and various other aspects of AI can be a little bit controversial. People often disagree about whether AI should really even be called AI — whether we should really be referring to this as intelligence. But the broadly kind of understood definition, or at least how I understand artificial intelligence, is the ability of a machine, arguably a computer in this context, to acquire and deploy knowledge in useful ways. And that's being powered primarily by machine learning (ML) — the ability to understand and infer knowledge from experience or from data.
So, that's how I broadly understand artificial intelligence. But throughout history, there have been forms of AI. Frankenstein, technically, is a form of AI, or I should say Frankenstein's monster. The Golem in Jewish tradition, in some of the stories, is an example of what you could call AI. So, there are obviously a lot of implications around computer science in our current age, but it's a concept that, in some form or another, has existed for a long time.

Sharma

Is there a distinction still in today's world between AI, artificial general intelligence (AGI) and artificial superintelligence (ASI)? Or is it largely been obscured?

Ajder

I think there definitely is a distinction. When we talk about AGI or artificial general intelligence, what we're really talking about is the ability of AI to achieve, again, perform various different tasks across different domains — generalized domains — on par or on level with human ability. And some people often refer to AGI as commercially useful applications.
Whereas artificial superintelligence is, as the name implies, it’s beyond the human capabilities. It's demonstrating a level of intelligent activity, which is far beyond what the human is able to perform. And that's where, I think, a lot of the concerns really kick in: “Well, if we develop artificial superintelligence, how are we going to maintain control of something which is potentially vastly more capable and intelligent than we are?” So, I think there is a distinction still, but it often does get muddied, because some people refer to superintelligence as general artificial intelligence. But obviously there is that distinction.

Sharma

And if you were to think about have we reached it, since you've been doing the cartography around this area? Or we're still to get to ASI, because some people think that already there — we've passed it. Others think, “No, it's still evolving.” How do you see it?

Ajder

Yes. I think, obviously, as you mentioned in your introduction, this is a space full of a lot of hype, a lot of noise, a lot of people making some quite bold claims, and other people sort of taking things which are perhaps closer to opinion as facts. When it comes to AGI, I don't think we are there yet. I think that's quite universally understood. And the same goes for superintelligence — artificial superintelligence. We've certainly seen pretty dramatic jumps, particularly in the last two years, catalyzed by the generative boom.
But when it comes to a single unified model — which can achieve all of these different kinds of applications, all of these different sorts of knowledge categories, and then act upon them on par with human intelligence or beyond — I mean, we've certainly seen some applications where systems can operate at scale, create content at scale that humans cannot recognize patterns that humans were not able to, even if we don't understand necessarily how they've come to understand those patterns, because of the black box nature of how machine learning works. But we're, in my view, still quite a way off the AGI and superintelligence framing, but it's one of those topics where some people would argue that large language models and ChatGPT are demonstrating nascent aspects of general artificial intelligence or superintelligence.
Other people would say, large language models are just probability machines that just spitting out what it thinks we want to hear, based on the huge amount of data it's been trained on and the weighting of those models. So, it's one of those questions which will continue to dominate. It will continue to confuse and cause controversy. But in my opinion, we're certainly not there yet, although we've seen pretty incredible progress in the field of AI, in general, over the last five or six years, really.

Sharma

The last year has been dominated with conversation or hype around ChatGPT. How does ChatGPT or text-based AI actually work?

Ajder

Yeah, ChatGPT really has taken the world by storm. And I think a lot of people now, when they think of generative AI, they just think of ChatGPT. Obviously, there's a huge amount more going on with GenAI beyond text, but also within text. There are many different large language models, which ChatGPT is one, a large language model, out there. But ChatGPT or GPT stands for generative pre-trained transformer, and the key part of that is the transformer. That's the specific breakthrough that Google invented, which has really driven this revolution around generative text. And essentially, what we've seen from GPT-2 — that was the first really big one that they showed off back in 2019, all the way through to GPT-4 and arguably 4.5 in the works. And what really is powering that is a huge amount of computational resource and a huge amount of data.
Most of the written internet in English is being used to fuel these transformer-based models. And the transformer effectively works, again, in simplified terms, it almost works to help a machine understand almost like a universe of meaning. So, if you think about lots of different words, they might have different meanings, depending on how they're positioned in a sentence. You could have the classic example, which many people use: Queen, you know, can refer to a chess piece. It also can refer to a monarch. In slang, it can refer to a woman who's doing really well in her career or something like that. And by training on all of the different texts that's out there, these models come to understand the semantic meanings of these different words in different combinations. This is what we call a word vector. Based on this training, based on these understanding of word vectors, these models come to understand that, if you type in a prompt, what probabilistically the following sequence word by word would be.
So, really, these models almost act as a very, very sophisticated autocomplete. This is effectively probabilistic reasoning. These models are spitting out what it believes to be the most likely next word in a sequence. And the effect of that is these incredibly powerful, incredibly dynamic technologies, where whatever you type in, unless it's something which the developers have restricted because it's something they see as unsafe, you're going to get some attempt by these models to put together something, which they believe to be probabilistically the most accurate way to respond to that question. And whether that's writing poems, letters, articles, and having conversations and debates — these capabilities in a world, where text is still the most dominant form of information communication that we deal with every day, is pretty profound.

Sharma

So, from what I understand from what you just said, ChatGPT or models like that are just literally predicting the next word contextually in a conversation.

Ajder

Essentially, yes. This is where some people have referred to large language models. It was in a seminal paper. Timnit Gebru, I think, is her name, wrote about stochastic parrots. This idea that we think these systems are intelligent; we think they've got this big move toward artificial general intelligence, like true intelligence. But really, what they are doing is parroting to us what it thinks we want to hear, based on this huge amount of data that it's been trained upon.
But obviously, some people would say, “Well, does it matter if it's truly intelligent or not when it can deliver answers that sound as intelligent, if not more intelligent than someone who is a human being, which would have classed as having an actual intelligence?” So, even if there's an element of almost a facade there. Some people would argue that, for all intents and purposes, that doesn't really impact how we use it and how it's going to change our daily lives.

Sharma

And is there a difference between using it for text versus audio-visual?

Ajder

Yes. I mean, it's a complicated one to describe and go into detail on all of the different kinds of models and architectures that are powering various different kinds of media generation. But the short answer is, yes, there are lots of different approaches. So, diffusion-based models, you might have heard of from tools like Stable Diffusion, Midjourney and DALL-E 2 and 3 — those are OpenAI's models. These take a very different approach, or at least certainly is not the same as a transformer.
Then you also have things, like generative adversarial networks or GANs for image and audio generation. And then you also have techniques for audio generation, which are also different. So, the short answer is, what's powering ChatGPT or Anthropic’s Claude or Google's Gemini, that is not going to be the same as what is being used to generate images and audio. Even if there might be some similarities between some parts of the baseline architecture, it's not the same.

Sharma

What are the key differences that you see between something that is text based, predictive, like ChatGPT, and audio-visual AI? So, if I were to deconstruct ChatGPT, I would say there's a model, there's a data set and then there's a learning algorithm. How does that work on the audio-visual side, generally?

Ajder

Sure. When it comes to audio-visual content, there are loads of different models and approaches to generating different kinds of images and text and even video now. Recently, diffusion-based models have really led to the tools, such as Stable Diffusion as the name implies, Midjourney, DALL-E 2 and 3, and the likes. But then there's also models, like generative adversarial networks.
And in the audio space, there are some really interesting early voice cloning tools. All of these models rely on data sets. So, obviously, if you're trying to generate synthetic audio, you're going to need a lot of audio of voices, birdsong and music, whatever that might be. And typically, the more data you have, and the cleaner, better aligned and more diverse that data is, the better outputs you can generate. We're talking about machine learning here. So, we're talking about these algorithms being trained upon this data to understand the architecture — the way that certain elements of that media are constructed to then replicate it.
And there are different processes for different models, but one that's quite fun, always quite interesting, which I always like to reference, is generative adversarial networks (GANs). This really works as almost as a game of cat and mouse — between two neural networks pitted against each other. So, you have a generative network, and you have a discriminative network. The generator, on the data that is being fed into the model, is learning how to replicate those images. So, let's say it's a data set of monkeys, it learns how to generate a really good picture of a monkey. And then discriminative algorithm is learning what it is about those original images of a monkey that make it a monkey. And what happens is the generative algorithm, or a neural network, will generate images of monkeys, and then the discriminator will essentially judge and say, “No, that's not an original, that's not an authentic image.” And then that helps the generator learn how to refine and improve that image incrementally, until you get to a point where the discriminator cannot tell the difference between the original data set and the images that the generative algorithm is generating. But there are different approaches, as I mentioned.
Diffusion models work by understanding how images can be reconstructed from a data set — from essentially visual noise, so reducing data down to visual noise and then rebuilding it based on the problem that's being provided. Yeah, many different approaches. The fundamentals that people need to know is we're talking about large amounts of data, we're talking about specific models or neural networks, and we're talking about a lot of computational power and resources to actually power that training process.

Sharma

So, this is fascinating. I'm going to sort of pick apart some of the things that you just talked about. If I take a step back and start thinking about, how do you make money in this field — a couple of questions for you in that. The first one is a lot of the models are coming out as free now. The data is available on internet. So, where is the commercial aspect of these technologies?

Ajder

It's a really good question and it's a really interesting dynamic we're seeing in the industry right now. So, many of your listeners will be aware of the billions of US dollars being invested both by venture capital firms and also large technology companies — Google issuing a red alert in response to ChatGPT something's that’s extensively challenging to their business. The same with the likes of Meta, OpenAI and Amazon — and I should say EY as well. I know EY has invested a lot into this space, too. So, big companies have recognized the power of these technologies. I typically refer to them in terms of realism of the outputs they can generate, efficiency of the models in terms of how much power and data they need now to generate really impressive outputs, and also, accessibility. Anyone can now start using large language models incredibly easily and at scale. So, there's a really compelling argument that these technologies are transformative, and I believe they are transformative, perhaps in a way that previous technology hype cycles have not necessarily been.
But when it comes to commercializing them, we're in an interesting spot, which is you can access ChatGPT for free. You get a much better experience if you pay for it and indeed, OpenAI is making fairly decent money off the page subscriptions and obviously looking to trade off the API. That's the version that businesses can plug into their own applications directly. But then there's also a large number of companies out there, including Meta, Stability AI, Hugging Face and Mistral AI — some of your listeners might have heard of that company raised millions. I think it might be something in the region of — I think it's about EUR50 million for seed round — crazy money. But they're taking an approach to open source, which means that they release these models in some form freely and availably. The model and the weights used to effectively train that model, the train weights for that model, that anyone can then essentially download and deploy themselves. There are, sometimes, restrictions, sometimes restrictions on commercial licenses, sometimes they release certain aspects, but not others for safety reasons. But there's really big questions here about, “Well, okay, how do you actually make money in open source?” What is the sales pitch there? Because ChatGPT, it's got a free tier. There's also the premium tier.
Whereas with open-source companies, there certainly are murmurings of doubt or at least concern. I mean, we know particularly in the text-based models, people see adoption as key. And it's like capture the market, get everyone on your software or on your platform, and then maybe you think about monetization. User acquisition is king for some venture capitalists (VCs) at least in the early stages. But you're right, there are really significant questions about how you make an open-source first business truly profitable.

Sharma

In some ways, it seems like different arenas, same fight, right? Like we are talking about network effects again, but in a context of ChatGPT versus — we saw the same fight happening earlier in terms of operating systems and other areas.

Ajder

Yeah, absolutely. I mean, it's not a brand-new dynamic. And indeed, maybe it's not being for free. But if you look at the likes of Amazon, their strategy for growth for a long time was to undercut competition using low prices. But it grew their market; it grew their presence. And certain companies are going really big on open source. Some people are saying, it's almost philosophically driven that saying, “Out of principle, we believe this technology, which is incredibly powerful, shouldn’t be in the hands of a few select companies operating out of Silicon Valley. This should be open to everyone.”
Then some people are clearly going for an approach which is, “Well, this is a really good way to build out our reputation and our user base.” And interestingly, Stability AI, again, a developer of many different open-source models for text generation, I believe they have one come out recently for — I think it was for speech or for audio. And then they also have, of course, Stable Diffusion for images, which is their most well-known product. But they've started to release a paid tier. And for a company or for a company structure where machine learning engineers are not cheap, running large amounts of GPUs — that's graphical processing units, the hardware that powers the training — that isn't cheap either.
So, it certainly looks like there is a little bit of a rebalancing, perhaps, going on in some spaces, but other companies have huge financial resources that you can plow into this, with the intention of capturing, I guess, market and building reputation. And some people would say that's a net positive. Other people would say that's not so good.

Sharma

So, we are talking about an advantage of network effect here. How do you see that evolving?

Ajder

In the strategic space around developing and releasing AI, there are different approaches. There's a more, I guess you could say, traditional approach, which is a more closed way of releasing these models, which is you don't share the back-end code or the way that the models being trained or the weights. And you provide it via a graphical user interface (GUI), an app or a web app, something that you can download and use. And you never actually interface directly with the code. You just get the product as a software-as-a-service (SaaS) style product.
Then you also have the opposite end of the spectrum which is referred to as open source, which is when developers release essentially the full code base for the model as well as — I say as well as — they can choose to just release the model, but often, they also release it with the weights included, which is essentially how those models then be trained on a data set. This comes from a tradition which has gone on for a long, long time in AI, particularly in the academic and collaborative space of using open source as a way to help further the field, people to build off each other's research to what they call “fork,” which is when you take a project and then build upon it in your own way to essentially advance the field. And as I mentioned, academics really have relied on this, as have other kinds of researchers — people who are building different products and projects.
But companies — big, big companies — are investing a lot of money and hiring lots of machine learning engineers to develop open-source products. And there are some big question marks about how you effectively build a sustainable, commercially successful open-source first company. Because ultimately, you're spending a lot on compute and powering GPUs that are used to actually train these models. You're potentially spending money on data acquisition, whether that's capturing it, paying for data or perhaps scraping the internet — it still takes resources. And then you've actually got the process of marketing and building the company around it. So, when you're giving that away effectively for free, there are quite a few people that think, “Well, okay, this is great, but we're not sure how that's going to really result in a stable long-term company.” It might get you loads of users. This network effect gets you loads of momentum, you get lots of people, using and deploying your tools, but if you're not making money off it, it can be quite hard to see the future of that space.
So, there is this trade-off approach. But then there's also a philosophical argument, some people say, “Well, stuff has to be open source, because if all of this powerful technology is just in the hands of a few select companies or organizations, moving forward, that's not good for society.” So, there are lots of different arguments, but certainly, some companies that have been open-source first have now started introducing subscription models for you to get premium versions of these models or tools or capabilities. But it will be interesting to see if they can start to make some decent revenue, and if so, where that comes from.

Sharma

From your vantage point, is there an advantage of scale of the company that has the model or that owns the open-source AI?

Ajder

When it comes to open-source models and AI, if you are a very large company which is profitable, which is perhaps very multifaceted in the services you offer, then maybe you can afford, as we've seen with some of the big tech companies, you can afford to plow potentially hundreds of millions, even potentially billions, of dollars into research first projects or subsidiaries or companies. And with some of the really impressive challenges, startups that are entering this space, we've seen big tech companies investing billions of dollars into these entities. Even if they're not profitable, even if they're bleeding money, they've been able to make that investment, because they're saying, “Well, okay, they're working on incredibly powerful research. They're building out capabilities that we could then plug in to our products, our services and so on.” Or they're doing it in-house and they're offering their own open-source tools, alongside their closed products or their closed services, either potentially as a way, again, to build reputation and to build momentum around that ecosystem. Or potentially because they just think it's a really interesting space to further and to enter, and to do research and hoping that maybe something useful comes out of it.
Whereas if you're a slightly smaller company, and you're building, let's say, an early-stage startup or an open-source first foundation, you may get investors. And indeed, some of the startups we've seen in this space have garnered incredible funding rounds at massive valuations with very, very small teams and not much to actually show for it. But ultimately, I believe, and this is my opinion, that there will come a point where your investors are saying, “Okay, how are you going to make money now?” And it's all well and good, capturing users. It's all very good, capturing market. But if you aren’t able to monetize, you're in trouble. Obviously, if you're monetizing other products as a large company, perhaps you don't have to worry about that quite so much. So, yes, I do think the scale of the organization or business developing does make an impact.

Sharma

If we can shift our focus a little bit into synthetic media, where are you seeing some of the best use cases in industries, different industries? Let's pick one or two to start with.

Ajder

Sure. I guess just to start — synthetic media or AI-generated synthetic media is a term that actually predates GenAI, but also predates the term deepfake. And as the name implies, we're talking about multimodal media with images, video and text. Some people would say, “Not text,” but for me, text is included. Even XR or extended reality experiences for things like metaverse applications, gaming and so on, which are either partially generated a wholly generated using AI or generative models. And this is where the explosion of interest has really come in the last 18 months or so, starting really with DALL-E 2, which was OpenAI’s text-to-image generation model. Because we've seen the rapid progress and rapid democratization of the tools, which were, just a few years ago, the preserve of the top research labs in academia and industry, we're now playing with them on the smartphone in our pocket. It's such a dramatic jump.
And as much as the technology space is really interesting, and I do see real value with large language models — particularly large language models which are bespoke to organizations and can be trained on very specific data sets for specific applications — the space where I really came into this from was really about audio-visual content. And for me, that's where this technology's actually been adopted for quite some time and has had real uncontroversial impact. I'll give you a few examples of where. So, one is essentially the new age of movie magic. Hollywood has been using VFX and synthetic elements in film production for decades now. And what we've seen with synthetic media — AI-generated synthetic media — is a whole new set of tools and possibilities for directors and creatives to build new forms of content, whether that's making tasks much easier, such as Rotoscoping and removing elements from a scene — which was used in Everything Everywhere All at Once, the Oscar winner. They actually used a piece of generative software to remove some police from a scene all the way through to things which are much more almost bombastic or at least certainly much more in your face, such as synthetic resurrection — the ability to bring back deceased actors or actresses to live on and give performances long after they've retired or have passed. Obviously, with ethical questions abound.
But nonetheless, the ability to swap faces onto body doubles, to generate synthetic voice. There’s a really amazing application in the Mandalorian; a Ukrainian startup synthetically generated Mark Hamill's voice as a younger man for that show. So, that's one space that, I think, is really quite profound. An extension of that would be general content creation. Advertisers marked as, again, very early adopters of this technology. And part of the reason for that, which I give as a secondary, is hyper personalization. A lot of people around the world consume most of their content, particularly if it's English language content with subtitles, because they don't speak English or they get generic messaging in the services they use or the products they use. Whereas with synthetic media, we can now generate entirely new synthetic lip movements to match, let's say, a voice actor’s performance. Or we can take it one step further. We can take the actor's actual voice, train a model on that voice and of another language being spoken, and actually get them effectively to give the performance in another language, while retaining their accent and the way that they actually speak, which is amazing. And it always makes me laugh when I go to another country and say, “The Simpsons is on TV,” not the best example, because it’s a cartoon. But you always hear different accents of different performers for the different characters. This technology means you wouldn't necessarily have to do that.
And then also, we have some really interesting applications in communication — using avatars for news presentations, for educative videos and training, and things like that. There’s some really cool work being done by a UK startup in this space. We're now a really video-first generation. Most people prefer video; they don't want to read. They want to consume unless it’s engaged with video. So, these tools using avatars are really interesting. And I guess final thing, if I can put one more in, which is fascinating to me, is the rise of virtual influencers and characters. So, Instagram accounts of nonexistent, primarily women, who have millions of followers and do advertising promotions.
It's incredible to see how, again, there's this generational shift in tolerance or acceptance of synthetic content and also, synthetic K-Pop groups — nonexistent K-Pop groups that or they don't exist in real life. There're no physical, fleshy guys behind it. So, these are all applications that I find really interesting because they raise some really interesting questions about our relationship to authenticity and what that means to us in the age of synthetic content. Because if I said to you before something is authentic, you’ll probably say, “Well, okay, it's not synthetic. We can treat those as opposite.” But it's no longer the case. People have meaningful, authentic experiences in synthetic-first worlds. And I think that's a real dramatic shift that AI is powering.

Sharma

That's so fascinating. Can we double click on this aspect of authentic versus synthetic, particularly around the concept of the uncanny valley?

Ajder

Sure. This is a really interesting one. And yeah, for your listeners, I'm sure most of them are aware, but the uncanny valley is this idea that there is a curve where if something is super unrealistic, we feel comfortable and familiar with it. And then the more we try to make something seem more human-like, we reach these points where you might have seen, for example, a robotic woman who's been shipped out to a lot of conferences to give talks. And it's really creepy. It's unsettling, because it's somewhat human. We can recognize aspects of a human, but there's also some areas that seem really off, and it can be quite creepy.
But at the other end of the valley is, it's very human again. It's so realistic. It mirrors an authentic human-interaction experience. I think we've seen pretty dramatic changes or progress in this in certain applications. So, for example, in the synthetic voice space, with some of the state-of-the-art models being released now, some studies have found that synthetically generated speech, I think it's more than 50% of people can't tell the difference on at least the samples that were provided. That might change, for example, if it's their mother's voice or someone they know very intimately. But it is getting very, very good. Same goes for synthetic images. I mean, the world was fooled by the pope in a Balenciaga puffer jacket. And there are many other applications.

Sharma

I mean, we saw Tom Cruise and we saw Zielinski. We've seen political figures — other political figures — that's for sure.

Ajder

There have been very many. But there are also still spaces where the uncanny valley has some work to be done. And I think generative video is one of those, where you’re having entire videos from scratch being generated by AI. There’re some great examples. If you search AI-generated pizza commercial, that will give you nightmares. So, there's still a way to go. And it's often the really small nuances that we don't consciously recognize. We subconsciously note them; we clock them. Those are what make the real difference in putting us at ease with synthetic experiences. Which is partially why images are some of the most convincing ones. Because they're less complex forms of media. It's not across a period of time. There isn’t movement. You don't have to have spatial temporal consistency, for example. So, there's way to go, but we have also come a long way very quickly in the last few years.

Sharma

That's indeed interesting. Now, your background is in philosophy. Tell me how you started working on deception.

Ajder

Yeah, it's been a little while. I finished my postgraduate studies in philosophy back in 2017, but it's certainly a key aspect of, I guess, how I got into this space and why I work on AI. My studies focused on philosophy of perception. So, I was looking at hallucinations and illusions. In some ways — I won't bore your audience with the technical language around what that meant in philosophy. But I was really fascinated about how we experience the world and how we describe those experiences. At the same time, in Cambridge where I was based, there was a really big movement around AI and a real conversation going on about artificial intelligence and some of the breakthroughs that happened at the time, and how they were driving new applications and questions about how they would change society.
And so, those two areas really came together, when in late 2017, deepfakes first emerged. And I saw this technology, initially in its very pernicious use in nonconsensual image abuse against women, but nonetheless, I saw this technology as what I referred to as almost a non-hallucinatory hallucination, which sounds really silly. But what I mean by that is it's not a hallucination, right? What you're seeing is there. The pixels on your screen are exactly right. The experience you're having, the conscious experience you are having, is exactly as your eyes or ears are telling you, but there's nothing behind it. That's where that hallucinatory side comes in. There's nothing behind the image. There's no reference. There's nothing in the world that corresponds to what looks like an authentic recording of the world. And so, from this position, I was really starting to get interested in how these generative technologies, synthetic media, deepfakes and the likes were going to change the way that we saw ourselves, the way we saw others and how we interacted with each other. And also, how systems and procedures within society and within institutions were going to change because of this.
And because deepfakes emerged initially in a malicious context, and many people still do refer to deepfakes exclusively as malicious uses of synthetic media, I also started to think, how are these going to challenge and cause real problems to society, whether that's governments, whether that's via disinformation against people on social media platforms, or whether that's to continue to use image abuse against women. Many, many different applications then also triggered some of my trainings in ethics as well. So, that’s a big part of what drove me to do this and still very much influences my thinking. Philosophy teaches you quite well, I think, how to imagine and even work within systems which are different from the ones we're working in right now or thinking in right now. And also, to try and forecast what future, with those new systems, could look like.

Sharma

This is a great segue in our conversation around ethics. When you had talked about the uses for AI, you talked about entertainment or actually using somebody's voice in a translation. In those scenarios, have you been in discussions around who owns the intellectual property or the original art? If the artist has already gone and not in this world anymore, what happens to the copyright? Tell us a little bit about that aspect.

Ajder

Yeah, it's such a great question, because what is at the bottom of a lot of the questions around AI is some variation of the concept of consent, whether that's consenting for your personal data, your biometric data potentially being used to train a model, maybe your artwork, may be your intellectual property being used to train a model and generate novel artwork.
But in the context of, say, synthetic resurrection, this idea of a deceased individual being brought back for new performances, there are some really interesting questions about, “Well, can you meaningfully consent to something happening after your death?” Now, there have been some performers, such as Robin Williams, who sadly took his own life. But in his will, I believe, he actually wrote, way ahead of his time in his thinking, that he didn't want anyone to do this synthetic resurrection with his likeness. But then you've got to think there are many, many more historical, I guess you could say, but not really that historical, maybe people, who died as recently as, say 2010, may have no provisions at all in their will to forbid this use case, because they didn't want it to happen. And in that case, consent for this falls to their estate. The person who legally owns their catalog of work and their likeness.
And so, we've seen many celebrities, actresses, musicians, performers and even artists being synthetically resurrected without any explicit permission from them before death that they wanted this to happen, but because their estate has either said, “Okay, yeah, this sounds like something that aligns with their values and they'd like to have it had done. They wouldn't have minded.” Or I'm not saying anyone in particular; I don't have any proof of this, but I imagine some people have probably thought, “That's a pretty big check we're getting offered to give license to have our relative to be in this commercial. ” There were some really, really key questions about this for me. And one is almost a paradox, but a really big challenge, which is, is it fair that people, let's say up until 2010, maybe even 2015, didn’t get the choice to say no to this technology?
Whereas from now on, I bet you every single celebrity will have their lawyers and their agents saying, “Make sure you have these clauses in your will, have these conversations, what you want, and what you don't want.” So, is it fair that there's almost this historical cutoff? But also, even in the cases where it does happen, should we be saying, “Well, okay, this doesn't seem like it aligns with their values?” So, there are some really big questions around consent. And what meaningful informed consent looks like, and whether that can actually ever be obtained in the context of synthetic resurrection, I don't think informed consent really can be obtained with synthetic resurrection, but I don't necessarily think that means that we shouldn't allow it. It's just perhaps something that we need to be more thoughtful when there, perhaps, needs to be a little bit more scrutiny in the legislation.

Sharma

What you said was very thoughtful and thought-provoking, because in my mind, it gets to the concept of copyright. So, copyrights didn't always exist, and the laws are always trying to catch up after the fact.

Ajder

Yeah. You're absolutely right. I mean, copyright is a really interesting one, because I think a lot of people now feel in the synthetic age — the age of generative AI — copyright is no longer really fit for purpose. It wasn't designed to protect artists or intellectual property holders from what they see, at least, as violations of their ownership of work. And that's partially because copyright is about copying.

Sharma

Exactly.

Ajder

You can’t copyright inspiration, if that makes sense. So, if I read a lot of Murakami books and then I read a lot of Hemingway books, and my writing style became a somewhat synthesis of these two authors’ styles. And I sort of made it my own through that process in some respects. Obviously, that's not the same as me copying line for line, word for word, passages from their books and plagiarizing.

Sharma

Right.

Ajder

But I'm one person and I can only do that, let's say, on a certain scale, whereas these models can ingest millions, hundreds of millions of images, billions of lines of text, and can spit out content at a scale which is beyond humans. I think the copyright question is a really, really interesting one. We're seeing it playing out in the courts at the moment with some of the AI companies being sued by artists, authors and so on. And it's one of those areas where we're seeing this case law essentially developing in front of our very eyes. And how that will then shape the future is going to be interesting, partially because copyright laws differ, depending on the country you’re in. Some countries have already said — for example, Japan has already said using copyrighted content for training data is allowable as long as you're not generating direct copies of artworks or copyrighted material. Other countries, I'm confident, will take a very different approach.

Sharma

The whole concept of what direct and indirect, how do you define it? It's gnarly. I know we talked a little bit around the ethical issues, and you advise a lot of Fortune 500 companies. You talk to a lot of governments. Is cybersecurity something that's coming up in these conversations?

Ajder

Very much so. Very much so. And with good reason. So, as early as 2019, we saw reports coming out that people were using, particularly synthetic voice cloning, to conduct what has been referred to now as “vishing” or voice phishing attacks. That is, calling up a CEO, a C-suite executive, a financial controller, or someone of this kind, or indeed private individuals — mothers, fathers and so on — and essentially impersonating someone else, directly using a clone of their voice, whether that's asking them to transfer money or impersonating the CEO, saying, you need to take this action or share this information. In the case of private individuals, there was a case earlier this year of a mother who got a message from her daughter, saying that she'd been kidnaped, and a ransom needs to be paid, which is horrific. So, these reports have been coming in for some time.
We're talking, in some cases, about tens of millions of US dollars allegedly being stolen. And again, it's only getting more challenging as the technology becomes more accessible. This is a real trend with many technologies. The more people can get their hands on these tools, the more people that are able to get to learn how to use them and modify them, the more harm we see — the more volume of attacks we see. And if you're a bad actor targeting, let's say, a C-suite executive, the payoff is likely very high if it's successful. So, even if it took a lot of resources to put together these kinds of cyber attacks, the payoff was still worth it. Whereas for private individuals, it may not have been worth it. That's fundamentally changed as it's become so easy to clone voices and now targeting everyday people, even if they're only transferring, let's say, a US$1,000, US$2,000, something like that. When it's as easy as it is now to do, it could be pretty compelling. And it's not just voice.
We're also seeing, for example, in the biometric space, people using liveness detection as spoofing tools to fool systems into believing that an actual person is in front of the camera when they're not. Also, for general face identification, and indeed, voice identification and tools, like online banking, have also been spoofed. So, there's a lot of concern. There's a lot of concern, because I think a lot of people still don't quite fully grasp how good the technology is. And a lot of organizations are trying to grapple with, “Okay, how do we respond? What new protocol do we need to get in place to make sure that we don't fall victim to this?” Because email phishing attacks, many, many people fooled them who are very smart, who are very capable, because they're so clever, they're so fine-tuned. And I think the concern is that, when you hear someone who sounds so much like that loved one in your life or that person that you work with every day, that's going to be part of your brain, that’s going to be part of you that reflexively trust it. And I think there will be, again, a societal or organizational shift in how we need to think about hearing a voice reflexively and maybe even in the future, Zoom calls like this or the video calls like we’re on.

Sharma

Henry, how easy is it to deepfake a person? How much data do you need about which is linked to a person's voice or video or pictures?

Ajder

Yeah. That's a really good question. And I guess the first question in response is like, “Well, what kind of deepfake are we talking here?” Are we talking about still images of them? Like some of the ones, as I mentioned earlier. Are we talking about voice cloning for text-to-speech applications? Are we talking about face swapping, just the facial region? Or are we talking about entire synthetic avatars? It really does depend, and it also depends on the quality you want to achieve. In certain areas, we've seen pretty rapid progress in this regard, and I'd say the one that has seen the most progress is voice — synthetic voice cloning. So, previously, the models that were used to do this were pretty janky. As we say, they're pretty low quality. They weren't easy to use and so on. And you'd probably need pretty good data — pretty clean voice data. And we're talking a good maybe 30 minutes to an hour of someone speaking to get results, which sounded certainly like them it was recognizably their voice, but it was also recognizably not real.
Fast-forward to today and with — I think even less than five minutes in some contexts, you can get a really good text-to-speech model of someone's voice using one of the most popular tools available in the market now, which I think costs something in the region of US$5 a month to use — so super cheap. In the image space, I mean, if you're a public figure and there're lots of your images online already, then chances are you're going to be very susceptible. So, if you type in a celebrity or a politician's name, chances are you're going to be able to generate images of them as we've seen pretty extensively. And people have used these for satirical and parody uses. They've also used them as part of the campaigning process even to aggrandize candidates or to actually target them with ridicule. So, if you're a public figure, you're probably already in the data sets. If you're not in the data sets, there are tools now, where with just one image of you, you can do face swapping, but you can also do style transfer of your image.
So, you might have seen tools — some of the really popular ones that you got on smartphones where you just need to upload a few images of yourself, and it would generate brand new portraits of you that looked really professional or made you a space man or an astronaut or a space woman or something. So, it really has dramatically reduced in terms of the amount of data you need. But I would say to reach that other side of the uncanny valley really convincingly, you still do need a lot of data and you need expertise to do that. So, we’re moving very quickly toward a world where I think that might change. But for now, hyper realism still requires a little bit of work and a little bit more data, but baseline, I guess, realism or something looks realistic, scarily easy now, compared with what it was even a year ago.

Sharma

I guess the keyword is “not yet,” which brings me to regulations. I know you are probably involved in some of those conversations. Some people think regulations are important and others think they're a waste of time. They would stop innovation. Give us a flavor of what you were interacting with and how you're thinking about it, how you're advising different actors in this ecosystem.

Ajder

Sure. I think, to start, anyone who says we don't need any kind of regulation is, to put it politely, a little bit silly. Whether it's cars, electricity, the internet or cell signal, all of these breakthrough transformative technologies have needed some kind of regulation to make sure that they're deployed fairly, safely and responsibly. But there is absolutely a meaningful discussion to be had about what that regulation looks like, how burdensome that is on developers, how enforceable it is to protect people, and what the actual value is. Obviously, in the last few months, we've seen real movement in this space. We've had the trialogue ending around the EU AI Act, really significant, very wide-reaching regulation, but also regulation that had some pretty large changes or rather some scoping changes in those final stages, particularly around things Like, frontier models or what people are often referring to in the context of large language models.
But then we've also seen Biden sign an executive order around this putting down some baselines. The AI Safety Summit here in the UK is bringing together, I think, around 25 countries, signing a memorandum on recognizing the dangers of AI. China passing some pretty big AI regulation at the beginning of last year — the Synthesis Act — and looking to do more as well. So, it is really around the world that countries are grappling with this, but as you mentioned, there is this almost waiting game on one level, where, of course, countries want to protect their citizens. They want to make sure that the laws address harms and mitigate harms. But also, everyone wants a slice of the action. Everyone wants to be on the frontier of the AI revolution, and they want companies to build there. They want to be attractive for investments.
And a lot of countries see this as a way to get out of some of the economic stagnation that certain nations around the world are facing. And so, balancing that pro-innovation approach with a real commitment to addressing the harms meaningfully is going to be difficult for a lot of countries, I think, to achieve. And I think that's why we're seeing some countries waiting, trying to see where some of the first move is and see how the reaction is, perhaps, from some of the big companies that are investing in this space, and to maybe be more reactive than proactive. I can certainly sympathize with that approach. But I also know, having done a lot of work on the harms that AI can cause, the longer we wait without potentially putting in legislation, the more people are going to be impacted. It's not to say that existing legislation doesn't capture many of the harms AI can cause.

Sharma

I think that's the key, right? Because a lot of — at least what I have heard is people are saying those harms are already being captured. So, why do we need more laws? And the second piece is, like you alluded to, enforceability, how do you enforce these things? Because just like you said, it may be right now it's maybe traceable, but very soon may not even be traceable as it becomes smarter.

Ajder

Sure. So, the new laws are going to be focusing on both the production and development of AI, and then also the impact. But the area which is most critical is going to be development, because if we don't set down frameworks and legislative guidance on how to develop AI, then that means that some of those harms, which may well be captured by existing laws, are going to be significantly amplified.
And so, looking at new harm or threat vectors, as we would say if you're talking the cyber space, AI is arguably the biggest source of new threats that exists in the world right now. And if something is a source of threats, you may want to consider how you can regulate that source to minimize the amount of threats, which could be covered by existing laws, but ultimately, being caused by something which is not very regulated at the moment. That's why you see, for example, in the EU AI Act, the classification of high-risk AI or high-risk technology. Which may have amazing creative, powerful applications, but also, could be weaponized and use maliciously, quite easily.
And that's why they're then having some auditing requirements in there and essentially some steps to make sure that you're showing how it's safe, how you've got in place sufficient, potentially, restrictions to stop it generating certain kinds of outputs, maybe have moderation processes in place to make sure that if it does get captured and then don’t get published, making sure that, for example, bias and alignment of the model — alignment means if the model has the values that we want it to in society, it isn’t spitting out hate speech or telling people how to create bombs and things like this.
And these are all things which are very much to do with the development of AI less than, I guess, the deployment of it by individuals.

Sharma

These are fascinating and actually, very convoluted topics. I guess I would ask you one more question around the whole concept of democratization versus elitism, because it's a powerful tool, as you explained to us. It's also a tool that's difficult to control. So, where should the control lie? That is an interesting concept. I don't know how governments are thinking about it, how you're thinking about it. Can you educate our audience?

Ajder

So, it's a huge question. It's a really interesting question. And it comes back to what we were talking about earlier in the context of open versus closed approaches to releasing AI. It's what I referred to as the openness dilemma, which is this idea that can you really make a truly, thoroughly safe, reliable and robust model in the open-source release format? And the answer is: no, because it's open, because anyone can go in, change code line by line. It makes it really, really difficult to stop people from potentially misappropriating it, weaponizing it and so on.
At the same time, open source is the bedrock of a lot of collaborative work in the AI space, which has powered incredible innovations, which has helped people who are perhaps less well-resourced than others to get into the space of developing or deploying. It's really integral to the story of AI. But some people, I mean, such as OpenAI — their name is OpenAI — and some of their early models, they were releasing in a much more open fashion. But they said of reason, “We recognize that our tools are so powerful that where we don't believe they're safe to release in this way anymore.” Now, skeptics say, “Well, okay, you would say that when you're trying to monetize your tools.” But other people would say, “Yeah, there are real harms that come with releasing these powerful systems in this way.” We're no longer dealing with the much less powerful tools that were being open source ten years ago, say. But then other people would say, “This is just a way of concentrating power.” If everything is closed, and we're just talking about these already huge titans of technology becoming even more concentrated in their power and control over them, that's not acceptable. And yes, okay, there's inevitably going to be some misuse.
It's an incredibly difficult question. It's one where I have real sympathies with both sides. A lot of my work, the harms that have been caused, in the research I've done, have been caused by open-source systems. I think that would be much harder to cause if they were closed. But at the same time, I've worked with some amazing organizations who have leveraged open source to build fantastic pro-social tools to help people as something which is kind of really powered. Again, the advances that we've seen have really been driven by that open approach. So, there isn't really an easy answer.
I think, in my mind, it's the trillion-dollar question. I think this is going to be the big area of debate is as these tools get even more powerful, can we change, or should we change that release model? And if we do, what are the implications going to be for society and essentially control of some of the most powerful systems humans have ever created?

Sharma

I think this is a perfect point at which to end this podcast. Henry, it's been a pleasure talking to you. You have made an incredibly difficult topic, which is fast evolving into something that is understandable, at least by me and hopefully by our audience. Thank you so much.

Ajder

Thank you so much, Mitali. It's a pleasure. As I often say, doing AI cartography is like doing cartography in a sandstorm. So, I hope this is useful to your audience. But no doubt, in a few weeks’ time, the sand dunes will have shifted again. So, I hope we can do this again sometime.

Sharma

Thank you so much.

Announcer

The Decoding Innovation podcast series is a limited production of EY-Nottingham Spirk Innovation Hub, based in Cleveland, Ohio. For more information, visit our website at ey.com/decodinginnovation. If you enjoyed this podcast, please subscribe, leave a review wherever you get your podcasts and be sure to spread the word.

The views of third parties set out in this podcast are not necessarily the views of the global EY organization or its member firms. Moreover, they should be seen in the context of the time they were made.

Presenters

Mitali Sharma

Mitali Sharma
Former host, Decoding Innovation podcast

Henry Ajder

Henry Ajder
Globally recognized advisor in the field of AI and synthetic media

Podcast

Episode 26

Duration

1h 5m 15s

In this series

series overview

EY refers to the global organization, and may refer to one or more, of the member firms of Ernst & Young Global Limited, each of which is a separate legal entity. Ernst & Young Global Limited, a UK company limited by guarantee, does not provide services to clients.