The problem with AI

0:00This BBC podcast is supported by ads outside the UK.

0:30Singapore, where business events can create lasting impact. At the BBC, we go further so you see clearer. With a subscription to BBC.com and the BBC app, you get unlimited articles and videos, ad-free podcasts, the BBC News Channel streaming live 24-7, plus hundreds of acclaimed documentaries. From less than a dollar a week for your first year, read, watch and listen to trusted, independent journalism and storytelling. It all starts with a subscription to BBC.com and the BBC app. Find out more at bbc.com slash unlimited. Hello and welcome to Tech Life on the BBC World Service, the programme that explores technology and the way it's reshaping our lives.

1:16I'm Shona McCallum. This week, artificial intelligence under the spotlight. Can you trust it to give you facts that are actually correct? And even when it is right, is it biased? In a moment, we'll hear from experts with deep roots in big tech on why those questions matter more than ever. Also today, how digital twinning is being used to preserve and restore a UNESCO World Heritage Site and the video games that are deliberately boring. So why are millions of us completely hooked? It's just so peaceful. It's just delightful. It's what I would imagine people get from colouring books or knitting.

2:16Let's start today with artificial intelligence. Do you use an AI chatbot at work or at home? And when it gives you an answer, do you actually trust it? We'll be looking at bias in AI later in the programme, but first, accuracy and how much faith we really put into these tools. A new survey for UK startup Unlikely AI suggests trust is far from absolute. It found employees now spend almost as much time checking AI's answers as they do using the tools themselves. And among business leaders, just 41 % say they trust AI completely. So who's building these systems and how reliable are they? Unlikely AI's founder and CEO is William Tunsil-Peddo, he helped create the technology that became this.

3:08That's Amazon's voice-activated assistant Alexa. He and his team are now trying to address some of the fundamental problems of AI, like making it more trustworthy using what's known as a neurosymbolic approach. He can now explain what that means. The way I think of it, there's two types of computer software. There's the traditional computer software that we've been using for decades, the software of your spreadsheet, of databases, of algorithms that works very reliably, can explain what it's done, and you trust it. And then the other type of software is machine learning, what's often referred to as AI, artificial intelligence.

3:50and that type of software makes big complex statistical guesses to do things, to predict things, to classify things and that type of software is solving problems that have not been solvable with computers for decades. So it can understand what's inside images, it can understand what's inside video and the big breakthrough in the last three and a half years is it can understand language but you know the downside of that type of software is that it's statistical it's wrong some of the time. It can't explain how it gets results in any way that makes sense to people, but it does new things, new capabilities.

4:30So if you use the word neuro to mean AI, to mean the machine learning, the statistical models, and you use symbolic to mean the traditional, reliable, trustworthy software, the idea of neuro symbolic is to blend the two so you get the best of both worlds. You get the capabilities of the LLMs, of the chat GPT, of the advanced AI solutions, but you also get the reliability and explainability and accuracy of the symbolic world. And that's what we're trying to do. We're trying to create software. We have very, very deep tech in that class that addresses these problems. And why did the industry move away from symbolic AI in the first place?

5:12Well, it's to do with the new capabilities. AI can look inside a photo, for example, and tell you what's in it. If you try to write a computer program to do that, you will not succeed. That's an incredibly difficult thing to do. So if you want to solve these problems that haven't been solved with conventional software, you need statistical models. You need machine learning. You need systems that can take data, learn from it. And as you expose it to data, it gets better and better. And because the human race is It's been generating astronomical amounts of data. And because computing has become more and more powerful, these machine learning models have slowly, over many decades, started solving problems that were not previously solvable.

5:54But at a cost of trust, at a cost of the models being black boxes and not being completely accurate. And you report that trust is low in AI. Why do you think that is? It's low for the reasons I've mentioned, particularly in regulated industries or in, you know, what we call high stakes industries. If you're running a business, it can have huge amounts of money associated with it. It can have a brand that's worth huge amounts of money. If you automate your business with statistical machine learning only, and it goes wrong, it can cost you money, it can damage your brand, it can result in you breaching regulation.

6:33breaching regulation can get you fined it can get you you know it can get you in all sorts of trouble so a lot of machine learning solutions that business leaders are trying to adopt for their businesses are not not succeeding not making it into production because of these trust issues because of these technology limitations that exist and your systems are designed to say that i don't know the answer to that why is that so rare in today's ai it's rare because of the machine learning, what the machine learning world does. So when a machine learning model is trained, it's making a statistical guess at the result.

7:10And as you give it more and more data, the guess gets better and better, but it's always guessing. So it doesn't say, I don't know, it gives you a result. And the hope is that that result is correct 87 % of the time, 92 % of the time, even 99 % of the time, whatever it is, but it's never 100%. So it doesn't know when it doesn't know, it doesn't have an understanding of how confident the answer is. All it has is a very sophisticated equation, essentially a formula that calculates something with very large numbers of very large amounts of data to make that formula get better and better. In the symbolic world, it can say, I don't know.

7:49If it calculates it and can't calculate it, it will say, I don't know. If it does calculate it, the answer is typically correct, unless there's a defect. So do you think we've accidentally rewarded AI for sounding confident rather than being correct? So if you're talking about LLMs, if you're talking about ChatGPT, yes, I think that's right. I think the data it's given to train on typically is expressed confidently. And when it's right, of course, and it's confident, and it turns out to be right, you know, that's fantastic. But obviously, it's incredibly trust damaging when ChatGPT or another of these models tells you something very confidently that turns out to be invented or wrong.

8:26That's, you know, so-called hallucinations, where it invents something and expresses it like it's fact. And what would you say would be the main solution then to trusting AI outputs more? I think setting expectations about what you're trying to achieve. So if you're trying to use AI in a business, I think having a clear understanding that the system won't be 100 % perfect and is this application something that can deal with that. But I think the long-term solution is technological. There are solutions, they're very difficult, but there are solutions to these problems. And the neurosymbolic technology path that Unlikely AI is on is one of those approaches.

9:03But I think long-term, it's about technology. It's about making the software work really robustly and really reliably and really explainably. So it can be used in a way that's completely trustworthy and completely reliable. And that was the CEO and founder of Unlikely AI, William Tunstall-Peddle.

9:28Now I'm going to take you on a journey to the only non-royal country house in England with the title Palace to find out how tech is helping its restoration.

9:46Welcome to Blenheim Palace, home of the Dukes of Marlborough and the birthplace of Sir Winston Churchill. Now nearly a million visitors come through these gates to see its splendour every year and now it's undergoing its biggest and most ambitious project to date. I'm here to see how technology is helping protect this treasured landmark. Inside, there are centuries of history. But the building is more than 300 years old and has suffered severe water damage over time. The massive roof, which covers an acre, began to leak after heavy rainfall. Those leaks caused water damage and mould issues on the artwork, especially the Great Hall ceiling, which was previously regarded as the palace showstopper.

10:41Dr Sterling McKinnon is the project manager for the Heritage Innovations Laboratory Oxford. This is a physical iron cramp and when these are exposed to moisture, when they're embedded in the masonry, they can swell up to I think 12 times the original size and that causes the fracturing and that causes it starting to fall off the palace. The water damage and decay have come at a cost, significant repair work. And at the heart of that work is the Digital Twin, a sophisticated virtual model of this UNESCO World Heritage Site. It has been, as you can see here, sliced and diced into tiny little bits.

11:23That's Clara Saliba. She's the AI Insights and Data Analyst here at Blenheim Palace. If there is a tiny piece of stone that breaks off of the palace, it is recorded here, which is important not only for visitor safety, but also predictive analytics. It allows the team to upload various bits of information about each individual element of the building. A lot of people are like, why are you doing AI at Blanham Palace? And I mean, it's this big, gigantic 300-year-old stone pile. But we actually have quite a lot of data. Every time someone comes and visits us, every time someone comes and buys a sandwich, every time someone has a cup of coffee and returns their reusable coffee cup to us, I get a line of data.

12:05And so it's really important to understand who our visitors are, why they're coming, so we can do things like effectively plan exhibitions, effectively staff the cafes, that sort of thing. The digital twin combines centuries of archives with real-time data from sensors across the palace. Handwritten maintenance records and old photographs are also being digitised using AI. And what this will help us do is visualise repairs in the palace over time. All this means there's a single source of information for the building and its operations, helping the team track repairs and even model the long-term impacts of climate change.

12:47By analysing the records with large language models, Blenheim's team now has the palace's long and complicated history at their fingertips. There are a lot of things at Blenheim Palace that are one of a kind. A lot in the staterooms is invaluable and is intangible cultural heritage. And so we want to make sure that that roof is repaired before there's a gigantic flood. We're now 109 steps up onto the roof and to the heart of this restoration project. You can probably hear some of the works going on up there. Now the aim is to protect the heritage and make sure that lots more visitors can enjoy it in the future.

13:26So what will success look like for you once the scaffolding's down in October and the project is finished? That's a really good question. Amy Akinola is a junior project manager here. It's always funny because we talk about once all the scaffolding comes down, no one potentially would notice that anything has been done. But for those who've worked on it from the very beginning, I think just having a palace that will withstand storms of maybe 100, 200 years' time without any leaks and the protection of the paintings underneath, I think, will be fundamental to our success. Now, it's lots of different people that work on a project like this.

14:04It's not just your regular builder, is it? No, no, absolutely. So we work with many heritage skills professionals. So we've got stonemasons, we've got lead workers, we've got carpenters, and each one of them have probably worked on a building like this within the heritage sector. Now you are a UNESCO World Heritage Site, so clearly lots of eyeballs on this project waiting for that kind of reveal. Do you feel the pressure? Yeah, I think there's always going to be an element of pressure. I mean, even if the world's eyes weren't on us, we would want to do the best we can because we all love this building.

14:39How useful has the digital twin been then? So by documenting this project and digitally archiving it, we hope that whatever works may happen in the future, they'll have a full understanding of why we made the decisions we made and where we made them as well. A final word from Dr McKinnon. As an academic who did a lot of research that was looking critically at things like 3D visualization, specifically as they relate to archaeology and heritage science, I've been really impressed and excited by all the work that we've happened here. It's been some of the most innovative use of the technology that's been available, and it's already making a big impact on the physical preservation and protection and conservation of this building.

15:22From centuries-old craftsmanship to artificial intelligence, Blenheim Palace is proving that heritage and innovation don't have to be at odds. It's 2009 and we're in the German mountains. a man straps himself into a car on the world's most dangerous racetrack. He whispers to himself, It's time to put my balls on the dashboard. as he starts the engine. In 15 minutes, he's in an ambulance, unconscious. In 15 years, he's a billionaire. This is Toto Wolff, Formula One's most powerful team boss and the breakout star of Drive to Survive. This week on Good, Bad, Billionaire, how Toto Wolff made his billions.

16:01Listen wherever you get your BBC podcasts.

16:16This is Tech Life on the BBC World Service with me, Shona McCallum. We'd like to know how you're using AI in your life and whether you trust it with the information it gives you. Does it matter if a chatbot isn't 100 % accurate? Our contact details are coming up in a moment. But first, we've received a question on WhatsApp from our regular listener to the show in Cameroon. Hello, BBC Tech Live. I am Nakombe Motte from Yohonde in Cameroon. If internet comes from satellites in space, why do we still have a map of undersea cables and threat of sabotage that can disrupt internet access to millions?

16:56Thank you. Well, hi Nakombe. It's nice to hear from you and it's a good question. Although satellite is an increasingly popular way of delivering the internet to people, the vast majority of international data traffic moves around the world using undersea data cables, and they are vulnerable to accidental or intentional damage, which is why it's good to have satellite as another delivery option. If you've got a question for us or an idea for a future show, please do get in touch. Our address is techlife at bbc.co.uk. You can send us a WhatsApp text or a voice note. The number is plus 44 330 1230 320.

17:37I'll repeat the details at the end of the programme.

17:45Now for our next item in today's show and... Ah yes, producer Tom has chosen this exact moment to vacuum the Tech Live studio floor because nothing says making a radio programme like some aggressive housekeeping. Don't worry, he'll be finished in a second. Although, honestly, he might want to stick around because what's coming up is this. Yeah, exactly this. We're diving into the world of mundane job simulation video games. Things like power washing, pool cleaning and mowing the lawn. Not so much heart racing action, more deeply satisfying chores. and somehow millions of people find it relaxing, even joyful.

18:29A peaceful alternative to high-stress gaming and apparently to untidy studios. Right, Tech Life's Laura Kress is here to explain it all a bit more. Hi, Laura. Hi, Shona. Nice to be here. And in a tidy, clean studio. Yes. Tell me about some of these mundane games then that have caught your attention. So you might be surprised to learn there's quite a lot of these games now. So there's Lawn Mowing Simulator. There's also Electrician Simulator. And there's also one called Power Wash Simulator. And that's made by Future Lab. They're based in the south of England. And it sold over 17 million copies.

19:05And its sequel has also just got nominated for two BAFTA Game Awards, which are awards in the UK. And essentially, you're putting on your power washing suit. You're cleaning different locations. It might be a roller disco. It might be SpongeBob SquarePants Home. that they've worked with all these different brands for new content. So I was fascinated to learn more. I spoke to one of the fans of the game, the comedian and games journalist Ellie Gibson, about why she enjoys it so much. It's just so peaceful. It's just delightful. It's what I would imagine people get from colouring books or knitting.

19:39It's you're doing this thing that's quite repetitive and therefore quite sort of calming and soothing. And there is a finished result. in this case you know it's a clean monster truck if you're a knitter obviously get an actual scarf which is quite good isn't it but you get to admire the principle is the same you get to admire your efforts don't you at the end and say I've done this and I think also it's just a very quiet game and actually there's been a fashion recently for a long time really for very sort of loud noisy games so if you look at AAA block loss I really like Assassin's Creed for example but it's quite bloody and action-packed and there's someone's you know there's always swelling cello strings or there's always the sound of someone's head being hacked off and that's all well and good if that's what you're in the mood for but sometimes you know I've got I've got two children and a job and sometimes I just want to be in a very quiet space doing something you know enjoyable.

20:40There has been some actual research about how mundane games like these are helping people's well-being. That's right. Well, in 2023, Power Wash Simulator was the subject of an Oxford University study. Over 8 ,600 players track their mood every 10 minutes with pop-up prompts while playing a custom build of the game. The study found that 72 % of players experience an uplift in mood while playing the game. So I spoke to as well, Kirsty Rigdon, who is the head of Future Lab, which makes the game. And she said she does get lots of players telling her her games have helped them through difficult times.

21:16We have lots of people writing, and this is where I end up getting upset, writing in and saying, you know, how we've helped them, like, get through cancer treatments and stuff. It always makes me really upset whenever I think about it, like, in a nice way. When people talk to us about their experiences with Powerwash, it's really lovely and uplifting. Like, you know, people tell us that we've helped with our anxiety or depression. I think it's just that mindfulness and being able to focus in on one thing kind of like blocks all the other distractions out. And have you played this game, Laura? I have, yes.

21:52And I do understand what they mean when they say it's very soothing. You know, there's no timers, there's no pressure, apart from the pressure washing, to finish quickly. And, you know, someone compared it to the Norwegian genre of slow television, where you watch a very ordinary event in real time. There's something very cathartic about it. I have to say, though, I haven't played it for as long as Ellie Gibson. She has two Guinness World Records for the game. The world record for the longest time spent playing Power Wash Simulator, which is 24 hours, 6 minutes and 33 seconds. And also, as well, the longest time spent playing a simulator.

22:32I have to say, though, she did have breaks and she live streamed the attempt on the streaming platform Twitch. pitch I love being the power wash simulator world record holder it's one of I'm not even taking the mitt it's one of the great honors of my life and I love telling people about it would you ever do the same with the sequel or are you like no that's it I've I've done it I think not I think um I think I've done it I don't think it would be the same yeah but But, you know, the joy of actually not just playing the game, but doing the world record was so immense. And it brought my Twitch community together in such a fantastic way.

23:19I mean, I cried when I finished the record attempt. And not because I'd finished, but because there are all these people just giving up their time and being funny and kind. And, you know, again, it's everything I love about Twitch. Age 24 Hours was just absolutely there in spades. And it was amazing. Well, thank you, Laura Cress, for brightening up our day. Although I am conflicted now between looking for a mundane game that might suit me or maybe just getting on with the task in real life. Yeah, it's a tough call, isn't it, Shona? Well, thanks very much for having me on.

23:57earlier we discussed the problem of accuracy affecting our trust in ai well bias is another area where trust in ai can be dented zara chatu worked for the tech giant meta for over five years and left last summer she's now the founder of code for good now a consultancy with a mission to minimise bias in AI. She's worried about the number of women starting to use AI tools such as chatbots. This is known as AI adoption. One of the key things for me is that there is a real adoption challenge when it comes to AI adoption for women. So women are adopting AI at 20 to 22 % lower levels than men. And because it's technology that is shaped by its user base, it means that the output is impacted by that.

24:46And so I want to build equity in who's using it to ensure that we build accuracy in the output. So are you saying that women could be left behind by this technology? I'm saying we need to intentionally bridge that gap in adoption if we want to see the positive impact of AI, if we want to point it in that direction. So what's at stake if we get it wrong? Well, bias systems, right? So we've seen what happens with bias. AI can scale bias fast. And so I think it's important that we intentionally sort of bridge that gap. And I also think it's important to say that this isn't a problem for women to solve.

25:27The AI gap is not a competency gap. It's a gap even when women have equal access, they're still not adopting it at the same rates as men. And it's really to do with ethics, trust, and interestingly, a fear of cheating and a fear of judgment, a fear that, you know, we're taking shortcuts. And so how do we amend and ensure that women are represented and that there is not this bias within AI? What needs to be done? This isn't a competency challenge. Yes, there is a level of AI training. But I think when we think about that, we need to think about culture. We need to think about how we can ensure that we are building cultures that sort of allow for experimentation, that allow for risk without this sense of judgment.