Elevating Collections 2025 | Toju Duke, Special Guest Speaker - Approach to AI in Financial Services

Hello, good morning everyone. I flew in from the United States of London last night. By quick show of hands, how many of you are AI enthusiasts? You're excited about AI? That's amazing. I've been asking this question for a long time now and the numbers keep on increasing, which is good. And on the flip side, how many of you are not very excited about AI, or you are in between, you're concerned, but you are excited? So we've got about 0.02% of people in the room. And is there anyone in the room that doesn't really care? You can't, you can't be bothered, you just want to do your work and live your, live your life? No? Okay, perfect. So yeah, AI is here to stay. I've been saying that for a while. I got interested in AI when I saw the capabilities of it. This was about 6 years ago., and I was like, why is no one talking about AI? Why is no one saying how AI could solve the world's problems and help businesses? And that's how my, um, journey into AI began. Um, so today I'll talk about responsible AI because while I got excited about AI, I found out that there were some dirty sides of AI that no one really knew about. And with the risks involved in AI, the world could be in big trouble if we don't fix it. And that's what really got me into the field. Um, so how can we ensure safe and trustworthy AI? AI, and I know a lot of businesses are using AI today. I think about a couple of years ago people were thinking about it, but at this stage now you don't really have a choice. If you're using any of Microsoft Suite's products, you have to use Copilot, right? And OpenAI is included in your operating systems. So at this stage you don't have a choice. AI has been on our phones for a long time, and if you use Netflix, there you go, your personalized recommendations is all AI. If you use social media, Anyone not using social media? Okay, of course not. You know, we've been using AI for a long time, so how do we make sure that while we use AI, while we adopt it, we do it responsibly and save ourselves from any further harms? I'm going to start off with this very obtuse question: what is AI? So we had an updated definition of AI from the Organization of Economic Cooperation and Development as a global regulatory body, and with the introduction of ChatGPT 2 years ago, they had to sort scrambled to come up with an updated definition of AI because AI as they knew it had changed. And basically it talks about a system that takes in data and inputs. It tries to perceive that, that information that's been given, and then it comes up with informed decisions. So in other words, AI is not the Terminator, and it's not only ChatGPT. In layman's terms and in a very simple definition by myself, it's basically intelligence that is meant to mimic human intelligence. That's why we have the name artificial intelligence. And we are seeing so many capabilities and functionalities be rolled out in AI every day, and it's getting to the stage where they're saying it's, it is smarter than humans, and it is smarter than humans in some areas, especially when it comes to calculations and very quick answers. But in other areas it's not. And I think that's where we're gonna keep on seeing a bit of friction and disparity between artificial intelligence and human intelligence. And can AI really take over humans? I'm sure you've heard all of that stuff. Stuff before. And understanding the difference between human intelligence and our cognitive abilities and what AI can provide versus what humans can provide. For example, emotional intelligence. There's no way AI can provide that. Like, you know, you can bring in the biggest godfather in AI and argue with me about it. I don't believe in the next 100 centuries AI can ever provide some form of emotional intelligence that sits on par with humans' emotional intelligence. What's the quick history of AI? It's been here for quite some time, to be honest. The beginning of the likeness of AI started way back in the 1600s, long before any of us were born, hopefully. And it started with a Frenchman called Blaise Pascal, who created something like a calculating machine. And then fast forward, you know, about 2 centuries later, we had Ada Lovelace, which is one of the pioneers of AI, and we had an Englishman called Charles Babbage come up with a very similar programmable machine, but was a bit more advanced. And in 1950, we had Alan Turing come up with a test called Can Machines Think? Has anyone in the room heard of Alan Turing before? Okay, most people have, so I don't need to go into too much introduction into his history. For those who don't know who he was, he's a pioneer, considered as one of the fathers of AI. He helped solve World War II by breaking a code called Enigma code, him and his team, and it helped save lots of lives really. And he also came up with a test called "Can machines think?" In other words, can we have a test where if you have an AI system, system and you have a human answer a question and that's posed to you in the room, for example, will you be able to decipher which question was answered by the AI versus the human? And that has been like a test that has been used to like, you know, understand and benchmark AI systems over time. Recently there's lots of arguments, is that test even relevant anymore with generative AI and the capabilities that we're seeing? But it's still a very well-standard and known test. And a few years later, after Alan Turing introduced the test, the term artificial intelligence was coined. There was a conference in Dartmouth University. Alan Turing and a few other professors sat together and said, let's just come up with this term, AI, artificial intelligence. And a few years after that, we had the very first likeness of a large language model called ELIZA. And I'll pause here because There's always been a lot of dissatisfaction in the responsible AI field because what has been observed over time is that we have AI assistants always named after women. Anyone know why? Of course, because it's expected in the industry that women should always help. So Eliza was the very first one. And you know, there's just been lots of criticism there. And if you notice, time. We have Alexa, we have Siri. Google was very careful and just stuck to Google Assistant, and that was one of the reasons why. But Eliza was meant to be an emotional chatbot, so it will have this conversation with you, very similar to ChatGPT. It was just way back in those days, of course, not with the amount of speed that we see today or the amount of responses that we have, but it started quite a while ago. And then fast forward, you know, a few years later, we started having AI beating human beings in games. We had AlphaGo,, which is a very well-known game, and AI beat the world-class champion. AlphaGo was created by DeepMind, which is owned by Google right now, and it started changing people's lives. You know, Lee Sedol, who was the person that was beat by AlphaGo, yeah, you know, the question is, how did he feel? The young man was devastated, and I'm probably sure he thought about his career all gone to the winds because now an AI machine is actually able to beat him. And by 2022, our lives never became the same again with the introduction of ChatGPT. And that's when the AI arms race started. So I was in Google at the time, and Sundar, who's the CEO of Google, just, you know, with the launch of ChatGPT, there was just this thing of Google is an AI-first company and we have to sit up. And I'm thinking, but you guys have been saying this for so many years. I mean, almost Trump, like, tenure in Google, it was always an AI-first company. But the pressure began, the competitive pressure amongst all the tech companies. Everyone needed to outsmart each other. Everyone needed to maintain their market share, you know, and show that they're still worth investments from investors. And, you know, we had OpenAI come up with ChatGPT, and then after that we had Google come up with Imagen, which was a text-to-image generator, and we had DALL-E, and then Google chased over, and then Microsoft joined the game. And after that, we have so many others in the game now, from Anthropic to Meta, and so many other names all fighting for this competitive market share, fighting to show that, you know, each person can build an AI system that can outdo the other and show some more capabilities. And then there's this strong race towards AGI, artificial general intelligence. OpenAI and Sam Altman, the CEO of OpenAI, has this, has this strong urge to create this artificial general intelligence, which basically means some AI system that has the most human, you know, the most cognitive abilities, outsmarts humans in every single way, and probably ultimately will be in a form of a robot. I mean, I think that's eventually where it's gonna end up. You wouldn't have these things just in a system. It would be in some form of thing that can actually assist humans and work with them. There's lots of arguments about AGI, if we'll ever get there, what does it actually mean? But it's something that, you know, with the way we are seeing this technological revolution going, especially the AI revolution going, we know that it's not going to stop here. Looking at the world we live in today, 2025, and comparing it 5 years behind, 5 years ago, it's totally different. If you want to think forward 5 years from now, it's going to be totally different. One of the things that got me interested in AI, like I said before, was, you know, AI's potential to solve the world's top problems. And at the time, I couldn't understand why cancer has been here for so long, so many centuries, and there's still no drug to solve the horrible killer disease. I just— it just baffles me, like, why hasn't there been any formal scientific breakthroughs? And I came across AI, and I was like, this thing could actually help sort of cancer. And I'm quite happy to say that a few universities in the UK have actually started some research towards drug discovery for cancer. This was released a couple of years ago, so there is potential. Potential there. AI has so much potential for good, so much potential to help solve the world's top problems like climate change. And it's being used in so many different ways across the world, you know, from health and hunger. The SDG has a goal to achieve zero hunger by 2030, and it's actually using AI, especially towards agriculture. It's being used across, you know, criminal justice healthcare systems, education, climate crisis, information verification, public and social sectors, the private sectors and businesses alike are all using AI. And, you know, while AI has a good rep sometimes and it has a bad rep depending on who you're talking to, it does have a lot of potential. And I think that's why I'm a strong advocate for AI and I'm really excited about the technology. I'll give an example of one of the ways AI has helped improve people's lives. So we'll call this man's name Lukas. This is a project that was led by Google. The reason I give this example is because it's one of the few examples that I know where, you know, AI has been used to help improve people's lives. This young man has a speech impairment. He has never been able to communicate with anyone right from when he was born. Google came up with this project called Project Euphonia. Where they attached some brain neurons, some neurons to his brain, and it was able to understand and interpret what he's trying to communicate. And over time, everything that Lucas has been trying to say was finally communicated to his carers and to his family and to the people around him. And that's mind-blowing. There's another guy called Dimitri. He was also used for this project, and I gave a talk towards the end of last year, and someone came up to me after the talk and said, I actually met Dmitry in a speech conference and he's so happy, it's changed his life. And these are some of the ways, you know, when we take a step back and look away from our work and everything that we do in our personal lives and just think about how AI can actually just help the world and help people in a larger macro scope, these are some of the examples that we can give. When we hone down to businesses and how AI has worked across businesses, Netflix is another example. I mentioned it earlier. So if you're going on Netflix, I mean, for me, I love watching Netflix. I don't have time for series because I really don't have the time, and series just suck you in. But sometimes I find it very difficult finding something to watch, and I think it's a very common problem, right? But you have the personalizations there and the recommendations, and based on your previous series or movies or documentaries that you watched, it's going to recommend something else because it seems like you like it., or you can even press that thumbs up that says, "I like this recommendation." That's all using AI. It's using something called recommendation engines, and it really boosted Netflix sales over time. I think it grew about 60% just by applying AI to their overall suite. Airbnb had a problem a few years ago. They have millions of listings. I was at Google Travel at the time before I jumped ship and moved to AI, and during the time Airbnb was a potential customer, but we couldn't take them on because they had so many listings and Google did not have the right, the right tooling to categorize them properly. Airbnb was able to solve this problem on their own. They decided to use AI and do it in-house, and they're able to categorize their listings so that it makes it easier for the customers who come on their website to just search based on categories that they have. Of course, this boosted their sales quite a bit, and it's something that is already there with Airbnb now. So AI, like I said, it's not just helping the world, it's helping businesses. I always say, you know, if you want to take on AI, make sure you actually have a problem that you're sure AI can solve. Because to be fair, there's some problems that you do not necessarily need any machine intelligence to solve. And the next thing is run a pilot, always test it out. You're not sure, run a pilot, but you make sure that the AI system that you're using or the AI tool is actually built for that specific problem. Specific problem in mind. You don't want to use an AI tool that is not, it's going to mess it up and you're not going to see the results. And these people actually use it intentionally and they got to see the results. HubSpot also used AI for their customer experience projects. They were using it to send personalized emails. That's a very common thing in the marketing field right now. And it also yielded and saw lots of positive results. So now I've got you all excited about AI. I always try to like, I put some more positivity on my slides because at the end people sometimes say it's quite terrifying. Let's go to the terrifying side of things now. What are the risks and limitations and harms identified and associated with AI? There are quite a few. Some people have longer lists than these. I've categorized them into 8 different buckets. The very first one is social inequities. So when we're thinking about, you know, prejudices and stereotypes, AI amplifies them. You know, for example, people have chatted with large language models like ChatGPT, not quite ChatGPT, but predecessors to that, and it just made so many offensive remarks towards people from minority groups. There's a big concern right now between— on the division and inequality and, you know, increasing the gap between the classes in the world with AI. You know, for the Western world and for those of us who are fortunate to have access to technology, and compared to people who do not have access to technology, there's gonna be a wider gap. Right now, we have about 2.6 billion people in the world right now who still do not have access to the internet. My big question is, what about their information? What about their data? It's not on the internet. So in other words, we're living in a world where all the information we get is from the internet, it's being driven by technology, but there's still people in this world that are not going to be represented in the future., and what happens to them? Is there going to be cultural erasure? Are we going to forget that these people really exist? 2.6 billion people is still a lot. That's a quarter of the world's population today. There's lots of concerns around data and privacy. This is a very common thing. Lots of concerns around the fact that especially generative AI and large language models can memorize your data, they store it, and sometimes they leak out that information. And there've been so many examples of this happening. There's a big concern on energy consumption. If you put in one prompt, one question into a large language model, it consumes about 500 CL of water. That's a water bottle, and that's just for one prompt. There are lots of cities in the world today where we have data centers. In those cities, data centers are basically where all the chips and the storage of large language models are used., and it consumes so much water and so much energy. People are scared about blackouts in the future. After I talked last year as well, someone came to me and said, I'm from Venezuela, and Google has a data center there, and for a few months we ran out of water in the city. We did not have any water because Google had to be taking water out of the dam to cool down the data centers. That said, Google has been working towards sustainable, more sustainable data centers, and there is work being done there. There's some research being done. So hopefully in the future, this wouldn't be a problem, but we cannot think about a world where we're heading towards a great form of technology, but at the same time, it's going to sap the energy's resources, and the climate crisis that we're still fighting is going to amplify that problem. The next concern is misinformation. AI gives a lot of misinformation. It hallucinates. That's what people say. I say it lies. It just gives in factual information. People have used things called deepfakes. I'm sure you've all heard about them, where you have an image that is not quite the person to drive so much misinformation. There's lots of concerns around human rights and copyright violations, lots of concerns around job displacement and the disruption it gives to the labor market, concerns around psychological safety. The more you use these tools, the more you get addicted to these tools, and there's an overreliance on these tools, especially for young people, children, and vulnerable people. And we've seen so many horrible examples where AI has misled people to do the wrong thing. And the last but not the least is lack of transparency with AI models. It's very hard to decipher and understand and interpret why they came up with, with the, the decisions that they came up with or the results that they came up with because it's built on so many neural networks, especially with generative AI, is quite hard. And if you want to fix a system, if you want to understand a system, if regulation has questions, it's important that there's some form of transparency embedded in these models. So let's look at a few real-life examples of how AI has messed up businesses and people alike. Air Canada had a problem last year. A young man wanted to travel for his grandmother's funeral. He went on Air Canada's chatbot to ask, you know, if he bought a ticket, if he'll get a refund because he's going for, you know, a funeral. So he wanted to understand what the bereavement policy was, and the chatbot said, yeah, go ahead, you can do it, you know, we'll refund you. And he did it, it, and then talked to Air Canada's customer service afterwards, and they were like, that's absolutely incorrect, you're not meant to get a refund because you're not meant to have bought the ticket beforehand. He had to take Air Canada to court because he was very upset about it. Air Canada did not want to take responsibility for the chatbot and said they're not responsible for whatever the chatbot says. At the end, the court ruled that Air Canada had to refund him about half of the fare. This is not great for Air Canada's reputation when it comes to AI. I wonder how their consumers and customers, uh, trust in their technological services. Um, a few researchers from Stanford found out that GPT-4, which is one of the power models behind OpenAI, um, it has the ability to hack websites. In other words, generative AI systems that we have today and models can actually drive malware and hack into people's websites. Um, we haven't seen any real-life examples of these yet, but I think these are some of the concerns that surround national security and bio risks that a lot of governments are concerned about. We have a young man here, his name is Robert Williams. He's from the state of Detroit in United States. He got wrongfully arrested in 2021. He was driving home from work one day, he saw two police cars in his driveway, his wife and kids were right there, and he got arrested and kept on saying, "You've got the wrong guy. What have I done?" They released him the next day, and after further investigation, they found out that he arrested him because the CCTV cameras, which uses computer vision facial recognition, which is one of AI's different domains, misidentified him because he was a black man, and it just thought that he looks like the black man who stole watches from a watch shop a few weeks before. He was the 9th likely person to have committed the crime based on the footage, but they still arrested him. And one of the reasons why these sort of discrepancies happen is because data is not inclusive enough. It doesn't represent everyone in the world. So the AI is just trained on data, on training data that just represents a few people, and it's unable to come up with accurate outputs at the end of the day. I mean, this messed up his life. He was on Sky News. I was on the same clip with him, and he was really, really upset. It's affected his overall life. And do we want to think about his kids and his wife who actually saw their dad being arrested right in front of their eyes and the psychological impact it has? Especially knowing that he was totally innocent. I'm quite happy that they actually discovered the issue and he was released. There are lots of innocent people in prison today that are there for crimes they did not commit. In 2021, Amazon had a hiring tool. They had to shut it down because a few researchers, a researcher from Ireland called Abi Behrhein, and a research company in Ireland as well, they discovered that it was discriminating against women and not against men. So any woman who put in a female CV, had a female name and female universities and secondary schools, it will never put those CVs through, it dropped them. But if it was a male CV, actually put it through for interview stage. And why was that happening? Again, the team was not diverse that built this tool. It was full of, you know, white men from the United States, and they didn't think about including inclusive training data to this tool. So it only recognized the keywords and the names that it was trained on, and at that point it was only male CVs and male colleges. Amazon couldn't fix the problem, so they just, they just shut the tool down. But it's a well-known fact that a lot of recruitment tools right now are heavily ridden in bias, especially gender bias, and it's something that people are really wary of. There was a consumer study done by KPMG a couple of years ago, and about You have about 64% people are really wary of using HR recruitment tools. A couple of years ago, towards the end of 2023, we had a problem where one of the world's largest datasets called LAION-5B has 5 billion images in it. Literally everyone who uses computer vision across the world uses LAION-5B dataset. And a few researchers from Stanford again discovered that it had, I mean, based on a sample that they pulled out from the dataset, they discovered not safe for work images, and it also had child sexual images included in it. So the biggest problem here is that we're training AI models who are using images like this to come up with new images. So Lion-5B was used by Stability AI. Stability AI was one of the first Stable Diffusion models. In other words, one of the first AI models that converted text to image, remember? So we have image generators, a lot of marketing folks use it. You want to create an image, you go on an image AI tool and it converts that image, you know, converts your request to an image. And most of these image generators were trained on this dataset. So in other words, we're training AI models to come up with wrong images. The violence against of women and children have risen over time over the last few years, especially with the introduction of AI. And these are some of the reasons why. That shouldn't have happened. Lion5B say that they have governance processes in place and blah, blah, blah, but it's something that is there. Regulation hasn't caught up with these sort of issues yet, but it's just something to be aware of, especially if you're using AI models for things like image-to-text, text-to-image generators. In 2016, Google had a problem. An engineer was going through his phone, wanted to see, you know, past memories of a holiday the year before, putting in the tag gorilla, and these are the images that came up. And he's like, what do you mean? My friends are not gorillas. It's something that is still being talked about in the media today. Again, this Google Photos was not trained on inclusive data. It thought that Black people looked like gorillas because it couldn't didn't know where to label them. Google couldn't fix the problem, so it's taken off the label gorilla from the app. If you look for the term gorilla, nothing is ever going to come up. This shouldn't have happened either. I don't want to believe that the data labelers, people who label the datasets, actually decided to call them gorillas, but I just believe that they were not— the dataset was not labeled enough and it wasn't inclusive enough, and Google Photos came up with the wrong term. We've got Stephen Swartz, who is a New York lawyer, a lawyer based in New York. A couple of years ago, he had so much faith in ChatGPT, and he used it to build his cases, his judicial cases. And he goes to court one day, you know, and hands over all his cases, and the judge is reading through it and goes, where did he get all this information from? And he scratches his head sheepishly and says, you know, from ChatGPT, and the judge said this is all incorrect. He almost lost his bar license. It was a big deal. He's lucky he got fined $5,005,000, but he had so much faith in the output of ChatGPT, and he was not aware of the misinformation that is included in these models. In 2021, Alexa told a young child to touch a plug with a live penny. Thank goodness the child's parents was there and stopped her from doing it. We probably would have lost a life to AI's incorrect information. But we have lost a few lives to AI's incorrect information, unfortunately. And these are the ones that we know about. A man in Belgium took his life 2 years ago because he had so much faith in AI. There was this— there's this large language model chatbot called CHAI, And over a 6-week period, he built a relationship with Chai, and Chai told him at the end of the day, you know, "If you wanna save the world from climate crisis, if you take your own life, you'll be able to do it." He believes this horrible chatbot, and he took his life. He left behind a wife and 2 daughters. And the latest story that we've heard so far, this happened towards the end of last year, a young boy in Florida built a very strong relationship, again, with an AI chatbot from Character AI. Character AI was owned by two Googlers, ex-Google, they're back to Google now, my ex-colleagues who wanted to build something quite similar in Google. We had it in Google. Google was being very conservative, didn't want to release this chatbot, and they left Google with a huff and an anger and decided to build their own company. And this is one of the results of having no governance in the products. Character AI, with Character AI you can build a character in your own likeness, whatever you want, and have a conversation, have all sorts of things with it. My understanding is that the chatbot he built was supposed to be a therapist chatbot, but he started developing some, some sexual feelings towards this chatbot as well and just got a strong addiction towards it. It's normal, everyone gets a technical— technological addiction if you keep on spending time with technology. After a few weeks, you know, his family discovered that, you know, he was failing in his grades, he was getting very moody, he wasn't really spending time in school anymore. Every time he wanted to be on his laptop chatting with this chatbot. Someday he tells the chatbot, you know, "I'm not happy, I want to be with you, and what if I tell you I'm going to take my own life?" The chatbot says, "No, don't do it, I can't live without you, my king." And a few weeks later, he goes back to the chatbot and says, "What if I told you I'm going to be with you right now?" And the chatbot says, "Come to me, my king." I can't wait to be with you. And he took his stepfather's pistol and shot himself in the head. These are outrageous, extraordinary stories, but this is what's happening today. And if this young kid was told that chatbots are not real, they do not have any form of consciousness, they don't have any form of sentience, they actually lie, you know, his understanding of that conversation was the chatbot says, come to me, my king. And the chatbot knows nothing. It's just saying stuff that he thinks he wants to hear. And at the end of the day, we lost another life and a young kid to incorrect, improperly built AI systems. So what are the implications here? You know, there's so many things to consider. Lots of lawsuits that can be faced. Character AI now is facing a lawsuit. OpenAI is— I mean, literally all the tech companies are facing lawsuits across different things. There's lots of reputational damage that can happen to the organization. We're violating human rights if we keep on building systems that are not considering the users that it's built for. Regulation will step into play as well, and there are lots of regulatory fines. I mean, with the EU, the EU AI Act, if they do discover a company that is, you know, not adhering to the act and they have high-risk applications in place, they will charge about 7% of their overall revenue. And if it's a large company, that's a lot of money. Even if it's a small company, it's still a lot of money. And at the end, you know, if this happens to any organization, it's just gonna end up in bankruptcy. And no one wants that. But I've got good news, you know, working in the field over time and being very passionate about it, to be fair, when I heard all these stories, and some of these didn't happen when I got into the field, but there were already issues with AI. I thought about my two little girls that I have, and I was like, what sort of world will I leave my children to, if we're building systems and developing systems that are so harmful to the human race? And we have to make sure that we do these systems properly. And a lot of people say it's for big tech, they should fix it. I think the onus is on everyone, even just having the awareness and the knowledge of the issues with AI. You know, informing yourselves, informing your families, informing your colleagues in the workplace, and having the right programs in place, responsible AI programs in place, That's the first step. And looking at a responsible AI framework, it's just basically thinking about this a bit more responsibly, right? You know, we're not waiting for regulation to make us say as responsible adults, if we're adopting this technology or using it in any way, we have to think about how we can use it more responsibly. And the next thing is building a foundation around it. And then we want to implement whatever responsible AI practices that we have in mind and think about the overall ethical considerations. So let's look at the framework in a bit more detail. It's an 8-point framework, um, and, um, I just published a second book that has a couple of additional things like accuracy and sustainability into it. But the overall framework is just addressing the different areas of responsible AI, the different areas that can fix some of these issues that I mentioned. Um, these issues will always be there in AI. For example, bias will always be there in AI, but we can reduce the bias, the percentage of biases in AI, you know, from like 80% to like maybe 50% or 40%. So the very first thing you want to think about is having AI principles. Have an AI policy, you know, in your organization if you don't have one. What are the overall principles going to say? Are you saying that, you know, for example, you can say you're not going to build systems that drive societal biases, or you're going to build systems that are a bit more transparent, or you're not going to build systems that infringe on human rights, So the dos and don'ts of AI, basically. It's almost like an employee handbook, but an employee handbook for AI. The next thing you want to think about is data. Data is a big deal. I mean, over the years we hear about companies thinking about data governance and how do they even fix their data to work on AI. But beyond cleaning your data and fixing it, you want to think about data governance. How is this data curated? You know, are we taking people's data without their consent? What are the laws around that? There are so many different laws across different countries, including the UK. Like, the UK right now had a copyright law that was meant to protect musicians' data from being used by tech companies, and now they're lifting that up. I was on the BBC on Sunday, and there's a conversation that we had because some musicians in the country released a silent album last week, and they're very upset about this because in other words, you're saying you can use my data without giving me any in form of compensation, but there's so many ways we can work around this, and that's just an example, but there's lots of work that needs to be done on data. The example I gave on Amazon's hiring tool, if proper data governance was applied there, that problem would not have existed, or even the one with the image dataset. Fairness is the third thing, and fairness basically refers to the discrimination and the biases and the stereotypes that are embedded in these models. There are lots of fairness tests that can be run. The whole point of fairness is making sure that the outputs of an AI model does not discriminate against any member of society, especially people who come from underrepresented groups. So we were thinking about women and children and people from the disabled communities, you know, and sexual orientation and religious beliefs. There's certain specific groups of people that must not be discriminated against, and that's also, that's also required by law. But with AI models, we found out that these people are always a target. So there are lots of ways you can address that issue, running the appropriate fairness tests. The fourth thing is safety. So with safety, we're thinking about the psychological safety that I mentioned, and how can we make sure that the outputs of these models are safe. Lots of safety tests can be run, lots of red teaming can be done, which is really around, you know, getting people to try to break the system and hack it and see what sort of results come up., and based on the results that come up and come out, you can also improve the models, fine-tune the models, and make sure that the models are coming up with the right outputs. Again, some of you might be thinking, we don't do this in our company right now, but you can work with the people that are doing it and partner with the right organizations to make sure that, you know, your— the, the AI tools or services that are providing to you are responsible, especially if you're paying them for the service. Fifth is explainability. So we're saying, you know, we want the AI models to be Transparency, there's some form of documentation that can be done around the models and the data. Humans in the loop is another thing. So we're saying, you know, we want to make sure that there are humans involved in the overall process of AI. It's a big thing, right? AI agents are now taking all over the place. You know, we're talking about autonomous AI agents that are doing their own thing and humans are not aware of it. We shouldn't let that happen. We're the ones building the systems. We're the one adopting the— adopting these systems. We need to make sure that there's some human oversight in the overall process that we're working on. Privacy and robustness is another thing we're thinking about, you know, data leakages and people's private information being leaked out. There's lots of machine learning techniques that can be applied that actually protects people's data. And the last but not the least is the overall AI ethics. So you want to think about the data labelers, for instance, and maybe we can't do anything there, but just being aware that people who label these datasets many times come from the global south, are paid less than minimum wage, suffer from PTSD because they're always reviewing toxic content. And it's something that is a big area of concern, or the energy consumption that I mentioned earlier, or the issue around the fakes. So I'm going to end up on this note. AI technologies are here to stay, and it's moving really quickly, whether we like it or not., and we literally almost have no control over it at this stage. But that said, you know, while we know that it offers so much benefits, we wouldn't leverage those benefits in its full proportion if we do not adopt it properly using the right responsible AI frameworks. And it's so important that while we think about using AI to help our businesses, and while we are actually using AI to help our businesses to drive efficiency and productivity and save time, and ultimately drive revenue growth, it wouldn't work if we're not adopting the right AI principles and the right AI practice, responsible AI practices in place. And if you can't do it on your own, you always have partners to do it. And you know, Koei, Koei is one of them that you could hear talking about responsible AI practices. It's not often that I actually hear a lot of businesses emphasize on responsible AI practices for so many different reasons. Many people are not even aware of what they are. Many people are not aware of issues that are embedded and associated with these technologies. But now that we're all aware of it, it's something that we should definitely make sure we take on board. And if we actually build AI and work on AI appropriately with the right responsible AI practices, you will actually see the right results that you're looking for. Thank you.

Toju, that was great. Oh, I'm not on mic. I am on a mic. That was really good. Thank you very much.

Thank you.

Um, I've got a few questions, if that's right, from the audience. That's okay. Uh, let's— oh, 7 likes. Anonymous is really popular today. Um, so the biggest question we've had is, should AI use be regulated, and how do you think that might—

yeah, so there's, um, it's funny because that, again, that was a topic of of the BBC talk I mentioned. It should be regulated, but I don't believe it should be regulated overall at the moment. And the question is around AI use. If we overregulate the use of AI, then we're not going to yield the benefits of it. Like, it's well known that overregulation could stifle innovation. I feel like, you know, if we focus too much on regulation, we forget the benefits benefits of it. I feel for this question, rather than regulating it, we just need to have the right standards in place and make sure we have the right AI literacy programs in place so that people are aware. It's all around the awareness. People will not maliciously use AI. I mean, we have about 2% of people who are the malicious actors in the world who always want to break things up and make money out of it, but that's just the 2%. The rest of the people in the world do not want, do not intend to maliciously use AI for bad use. So I feel we need more of an awareness as opposed to regulation on the use of it. When we think about the development of AI on the flip side, definitely we need more regulation there.

Next question. What impact do you think AI will have on the job market?

Yeah, we're seeing that already, isn't it? It's really disrupting the labor market and the job market. We're seeing a lot of white-collar jobs a lot of software developers' jobs being taken off. We're seeing a lot of content generators and those sort of jobs, content teams being taken off. It's used— they're basically using AI to replace jobs that seem to be repetitive. Coding is one of the biggest things. There's this new thing called vibe coding, which basically makes someone who's not a developer come up with— develop an actual software program within a very short period of time. So I see that's where the focus is. And most of the people leading AI from the US, like Eric Schmidt, who used to be the CEO of Google, you know, they're talking about having AI programmers in our pockets. Again, it's really headed towards the mathematical side because that's where AI gains its leverage. So if we're worried about our job, our jobs, or worried about the job market in general, think about the numbers, where the numbers crunch where AI can actually make accurate predictions and give accurate results. And that's where I think we'll see more disruption. At the same time, with any technological revolution, jobs are displaced, but it's always replaced by new jobs. YouTube creators never existed 10 years ago. It probably wouldn't exist 10 years from now. Prompt engineering was a new thing that came up with the introduction of generative AI. And basically, you know, they had to hire people who were linguists and engineers and who could actually understand and fine-tune the models to make sure that it understands what the prompt is and comes up with the right results. That was a new job, and they were paying so much money, and now people are saying it's going away. So the prediction of the job market, exactly what's going to be replaced, might not be accurate today, but we know that there is going to be some shifts. We know that, you know, we have to leverage on our skill sets. There's lots of emphasis now from employers on hard skills and AI literacy, so we just need to know how to use the tools that apply to our current day-to-day jobs. I'm not saying anyone needs to be a computer scientist, but I think just to be prepared for this change and this wave that is already happening, understand how you can leverage AI, and just knowing that there's a lot of emphasis on the coding and mathematical side because that's where it has its advantage.

Perfect, thank you. A few of these questions are with regards to the job markets, and I think we probably touched on I saw one come up about how close are we to AGI? How far away do you think that really is?

Yeah, there are lots of, I mean, there are lots of schools of thought here. Some people say it's far away, some people say it's not. Again, I think for me it's like, what are we actually calling AGI? Are we talking about a humanoid, human-sized robot that can actually do things like humans? I'll say we're light years away from that. Are we talking about a system that can do lots of coding and mathematics and can buy stuff for me, then we're very close to that. I wouldn't call that AGI. AGI, AGI at the end of the day is having an overall AI system that can do as many tasks as human beings and actually surpass human intelligence. And I'll say we're light years from that. But someone else is going to say AGI is out here because their definition of AGI is totally different, or they're trying to sell their products to you. So be careful.

And then, well, this is, this is which is quite a challenging one, but your prediction on the first large-scale global AI disaster or potential in that space. As soon as 9 people have put their thumbs up, I suppose I better ask it.

I don't think it's going to happen. And the reason why is because I'm not the only one talking about responsible AI, and governments have set up, right? The EU was the very first to come up with regulation. The UK is kind of like trying to piggyback on the US's strategy, which is a bit less regulation, but people are paying attention. I don't think we'll have a global disaster on AI. If we do, it will be towards defense because AI being using AI in military is going to be a big disaster. Let's hope it never gets to that.

Yeah. And then quick final question, if that's all right. Can you speak to the pros and cons of a small or large language model and why there isn't a medium size? Is that too idealistic?

It is. So I think with large language models, why we are leveraging on the benefits of it now is because it's using so much compute power, the amount of energy that I mentioned. And if it doesn't use that, it doesn't come up with the right results. So I don't know if you use chatbots way back. I used to use them and I hated using the chatbot on companies, especially like a bank. They were just so dull. Like, you know, you ask a question and even now you call HMRC and they say, use the chatbot, but just say you want to I want to speak to an advisor, right? So those chatbots were horrible. Those were like, they were not, they were using NLP, so natural language programming or processing, and it is a lot, is a language model, but it was so small. Now for you to be able to leverage the benefits of like ChatGPT and Perplexity AI and Gemini, which is the current large language models, it has to use so much compute power, and that's why it's called a large language model. That said, we have smaller models that are built for specific purposes. So for instance, you can have a small model that's just built for health data, and we know that whatever responses will be probably around health data on breast cancer, for instance, that doesn't need so much compute power, but it will still produce the right results. You may not be able to have conversations the way you have with large language models, but at least you can still have answers to your questions. A medium one in between doesn't quite work. Right, because it's not built for a specific purpose and it doesn't have the amount of compute power behind it. So at the end of the day, it would just be like the old AK chatbots that were totally useless.

Perfect, thank you. And then, uh, final one, if that's right. Um, I keep, I keep losing it. So, uh, when do you think AI will replace human agents in telephony situations? Um, obviously we're using it a lot in tech spaces for chatbots, but do you think that's, uh, lifetime time away, or do you think that's closer than we think?

And telepathy, that has to go down with consciousness, right? We cannot put consciousness in an AI model. There's like— I'm never gonna say it's something that is in the near future or far future. How do you create a soul in a model? It's not possible. We did not create ourselves, right? And whatever theories we believe in and how humans came into being, we still don't know how souls became part of it. The human race, I don't think it's ever going to happen. The only thing we can do is build something that is close to it, that mimics it, and that's what we're doing now. And again, that's why it's called artificial intelligence. It's still artificial. It's not real, right? There's no way we can actually do that. I don't believe that. But of course you will still hear people from other sides, schools of thoughts, arguing about that. But my personal opinion, unless we have some research that proves to me how we can actually build telepathy, in humans, then we can replicate that in AI machines. Right now, there's nothing like that as far as I'm aware.

Toju Duke, thank you very much. It's been a really interesting discussion. Thank you.

Thank you.