this is bbc news, the headlines: thousands of people on the greek island of rhodes are forced to flee their homes and hotels as widlfires continue to spread. the country has been engulfed in fires for several weeks as europe continues to see record breaking heatwaves. israel's prime minister benjamin netanyahu announces he's being fitted with a pacemaker, ahead of a crucial vote on his controversialjudiciary reform. tens of thousands of protesters marched on jerusalem on saturday evening against a proposal to give politicians more power to appoint the judicary. and it's day four of the women's world cup, where the first match of the day, sweden against south africa, is expected to get under way in the next half hour. later, netherlands play portugal and in the last match france take on jamaica.

now on bbc news, click. this week, we delve into the world of artificial intelligence. shiona heads to wimbledon to meet the latest match commentators. norrie, from the united kingdom, will play tsitsipas. could you spot a deepfake? we look at the software trying to detect them, but how good is it really? ithink fake. that is actually real. and joe investigates the darker side of image creation. yeah, i know you've redacted it for me... yeah, it's still powerful, and it's difficult. - this month, fans all over the world were glued

to coverage of the wimbledon tennis tournament. it was televised for the first time in 1937 when fred perry's run of wins was ended by the american don budge. since then, the bbc and broadcasters around the world have provided expert commentary on all of the action. well, actually, not quite all of the action. no, covering of the outlying courts has remained somewhat silent, but that could be about to change, thanks to artificial intelligence, as shiona mccallum has been finding out. what was it like as a player walking out on centre court? it is just the most fantastic feeling. in front of this crowd, with them cheering you on, and that thrill, and bit of fear as well, but it's all mixed in and there's nothing like it — it's just wonderful. jo durie was britain's number one in the 1980s. jo recovered to win the match, taking the final set 6—0.

like other players and fans, she knows what a magical place wimbledon is. look, nothing beats the atmosphere here on centre court in the sunshine in wimbledon. it's epic. but not everyone can make it to south—west london. so, for a lot of us, that means watching on a tv, mobile or computer — not all of which will have commentary. the all england club has been working with its it partner, ibm, to fix this for the wimbledon website and app. it collects masses of data from all around the courts for its ai platform called watsonx. what we're doing is we're taking all that information and we feed it in to the model, which then takes text, turns it into numbers, takes those numbers, and compresses it, expands it, then outputs texts again — that can then be spoken as an ai commentary. and this is what it sounds like. swiatek, ranked first... both male and female—sounding commentators are featuring the app.

murray, from the united kingdom, will play tsitsipas. but the voices are computer—generated, which is why they lack — how can we put it — the human touch. what a fantastic place to do an interview. this is amazing, isn't it? right on centre court, sun's out. jo durie, who we met earlier, has been commentating since she stopped playing professionally. what do you think makes a good commentator, then? i think you need to be short and sharp, sometimes, with what you're saying. and to get a feel of what's happening. sometimes you don't need to say anything, because the crowd have gone mad, and you want to feel that, and you can see the players — so let it go. other times, maybe when it's flagging a bit, then it's time to fill in a bit from behind the scenes. so i think you very much have to have that feel of when to talk and when not. so is this just the start of computer—generated commentary replacing

people like jo? to be clear, this is on matches that don't have commentary at the moment. we would love to make all matches available in a more accessible way, and you can see how the intention here is to complement and supplement those commentators that we know and love, rather than replace them. for now, the ai commentary is only being used after the event, for highlights, but ibm has plans to generate commentary in real—time. we could also label, or classify, for example, a net approach, a volley, the types of shots... but just how close to that goal are they? so we're actually very close. you know, these models are very big, and, in fact, i'm going to tell you a very interesting stat, here, is that, you know, by the time a tennis point is over, i can create commentary in 1—2 seconds. so it's really fast. at the beginning of the project, that was the grand challenge, was to be as fast as possible. and we were looking at the order of 20—10 seconds. will be able to do it in less than a second, is that the aim of this? yeah. yeah, yeah, so i think

it's achievable, for us to get there. you know, we could horizontally scale out some of our algorithms. we could make a data—gathering collection faster as well. so when might people be watching live and getting real—time ai commentary? i think it's going to be sooner than any of us could really imagine. but the time we roll around to next year's tournament, i think you'll really enjoy what you see. where do you see this technology moving in the future, and do you think it will be on real—time matches, for example, next year? well, i can't give you any commitments, but i... but i can say what we are going to be doing is innovating and inventing new experiences which maybe we have not even imagined yet. ok, but what do the professionals think? wow, now we're going somewhere where it's a bit scary, i think. for me, in my mind, i can't quite understand how an ai would grab the feel of the situation. yes, there might be certain things that happen in a match

that you can say, "oh, they missed this or hit that that," but what about the reaction of everything? what about the reaction of the player? what has just happened? so to get my head around this, i'm... i'm just not sure. and, you know, will i be out of a job soon? i don't know. ibm sasto's job is safe for now. but five, ten years from now? well, that's a whole different ball game. that was shiona, and look, i know we're a week on, now, but i'm still reeling from both of those wimbledon singles finals. and to be honest, i think, right now, i would rather hear a human commentator react to those amazing shots. that was some incredible tennis. yes, and aside from some of the amazing opportunity that ai brings us, it can also have a dark side. you may find this next piece hard to watch, but it does show the reality of what's happening

in the world of ai image generation. yeah, joe tidy�*s been investigating this story, and, as part of it, he's been shown redacted versions of some of these images. now, you won't see them here, butjust to repeat you might find the next film upsetting. these days, anyone with a laptop can be an artist. with aland an idea, the sky's the limit. and online galleries and forums are exploding with creations. by typing in a few words, you can create any image that you like, so, for example, a cat holding a banana in the style of van gogh, let's try that. the only real limit is your imagination, but there are some safeguards put in place. many of these tools have banned words or phrases to stop people

creating illegal or offensive imagery, but we now know that hasn't really worked. there are only three organisations in the whole world who are licensed to actively search for child sexual abuse material, and this is one of them, the internet watch foundation, here in cambridge. in this highly secured facility, the iwf finds, removes and logs some of the most abhorrent images and videos imaginable. their trained analysts say they're seeing a flood of images made with artificial intelligence. and we're the first people outside of the charity to be shown redacted versions of some of the images being shared by predators. the first image i want to show you is more of a cartoony style of image of a girl on a beach. i would say she's probably about 3—6 years old. it's really realistic. it's quite realistic. if i saw that, i would think it was a photoshopped image. so this location is a bedroom, and that's quite classic of real images that we see of

children posed in the bedroom. so... do you see that type of image...? yes. children doing that pose for...? that. . . children that young ? yes. it's a horrible image, just the idea of what she's doing here. i know you've redacted it for me... yes. but it's still hard to see, actually. it's still powerful, and it's — it's difficult, absolutely. it's ok. sorry about that. i've never experienced anything like that, and it was a shock. absolutely. and there's no need to be sorry, because these images are shocking. last month, for the first time, the iwf started actively logging reports of ai images. analysts discovered galleries on multiple websites, some containing category a material — the most graphic possible. stable diffusion is the most popular open source image generator. it's been repurposed and repackaged by countless websites and businesses.

it's giving image generators from big tech firms like open.ai, microsoft, and others, a run for their money in popularity and power. but stable diffusion doesn't have a silicon valley start—up origin story. it started here, in a leafy downtown corner of munich, and the way it was launched and created was completely unique to any of the other ai image generators. we're all really now looking over the shoulder of a development process which is super rapid. professor bjorn 0mmer was the lead scientist on the stable diffusion project. he and his team did their best to remove pornographic content from the two billion images that the al was trained on, but he admits it wasn't perfect. they also coded in a list of hundreds of banned words and phrases, but, of course, people quickly found a way to delete them. he defends their decision, though, to unleash his model onto the internet as open source. the dangers that you portrayed with open source, yes,

i see overall, like, there's potential in generative ai, that it's a powerful technology, and powerful technologies can be misused. but closed source has not proven to be the way that would actually do this mitigation for us, because it has been either leaked or, even more importantly, just been reimplemented. but you must accept that by making yours open source, you've made it easier — very, very easy — for people to download it and do whatever they want with it? of course. we made it easier and that's why we first off first released it to the research community. that was something that was important for me — that the models had just been released to the research community. we really need to face the fact that this is a worldwide, global development, so us stopping it here would not stop the development of this technology. all of the versatile api functionality is built from the ground up on our robust cloud platform.

stability.ai's the most prominent company developing it. pretty good, hey? its founder, emad mostaque, helped fund professor 0mmer�*s research. he declined to do an interview, but has previously said the firm prohibits any misuse of its ai for illegal or immoral purposes. but, of course, stability.ai has no control over what others do with the source code. and as regulators begin to plan potential legislation for al companies and products, it feels like, in some cases, the cat might already be out of the bag. if you've been affected by issues raised by this story, details and support are available at bbc.co.uk/actionline. time for a look at this week's tech news. the production of tesla's cybertruck has begun. the electric car maker tweeted a picture of its first built truck at its giga factory at austin, texas.

the light duty vehicle was first revealed in 2019, but elon musk blamed battery supply issues for the delay in production. the united nations security council has held its first meeting to discuss the risks and benefits of artificial intelligence. the is—member council heard from two ai experts who said that the world must come together to prevent the misuse of the technology. the overarching goal of this body would be to support countries to maximise the benefits of ai for good, to mitigate the existing and potential risks, and to establish and administer internationally agreed mechanisms of monitoring and governing. it's good news for playstation owners as it was announced at call of duty will stay on the sony system after a deal was made with microsoft. the news comes as a judge in the us rejected calls to block microsoft from taking over activision blizzard, the owner of call of duty. and tiktok has just signed a deal with warner music. the first—of—its—kind agreement

will give artists more opportunities to grow their platform and connect with its fan base. tiktok says further development with the deal will be announced in the coming months. the internet is the backbone that underpins so much of our lives these days. and although many people have been part of its development, these two are generally considered its fathers. vint cerf and robert khan developed the method by which data travels around the net. the internet story is now 50 years old. it started in 1973 and its roots go back into the late 1960s. we designed it to be expandable and extensible, and i'm very pleased to see how this thing has scaled up

by probably six or seven orders of magnitude from the time that we started working on it to today's current internet. and there's still plenty of room to grow. these days, he works for google — and he's still watching and working on new networking technologies. and i asked him for his take on the latest iteration of artificial intelligence, especially the chatbots that sound so human in their writing, but not in their fact—checking. the scale is astonishing. the verisimilitude of intelligence is striking, and the reason that it is so misleading, i think, is that what gets produced is persuasive because it's glib almost — it certainly parses correctly. and because it looks well done, we assume that it must have significance, that it must have weight. so i did a test with one of these chatbots — i asked it to write an obituary for me. i figured that form had been ingested and also

there were facts on the net about me that it could use. but what it did was take facts about other people and conflate them with me. and the reason for that very simply is some of those facts were mentioned in the same text as references to me. so it's almost as if... there's this appliance called a salad shooter. it's a thing that chops up vegetables and sprays them into the salad bowl — like carrots and lettuce and so on, celery. so i want you to imagine that the vegetables are facts. and you stick them in the salad shooter and it chops them all up into little pieces

and throws them in the salad bowl and then you go like this in the salad bowl and you put the salad in front of somebody. that's what happens with these things, because they chop up the facts into pieces and then they conflate things together because they don't know any better. vint thinks that we might be able to protect ourselves from chatbots that don't know fact from fiction by taking a leaf from the way that self—driving cars are graded depending on how autonomous they really are. at the lowest level — level zero — the car doesn't do anything for you, you have to do everything. at level one you may get some warnings for example, "don't change the lane right now because there's a car over there and we can sense it because i have a sensor that shows a little light." and then as you work your way up the levels the car takes over more and more responsibility. think about the chatbots as being like self—driving cars in a way, and we should think about what level of risk are we taking when we use them for something? so if it's used for

entertainment there's very little risk. if however you work your way up in application until you get up to somewhere around the top where it's giving you medical advice, or is doing a diagnosis and recommending treatment, that's a super high risk thing, just like a self—driving car. so i think chatbots deserve a risk rating depending on the application for which they're being used. vint, thanks for giving up so much of your time, it's been a pleasure. well, i certainly enjoyed the chat. you'll blame me for interfering with your democracy. it's not hard for- democracy to collapse. these days, it's not totally clear what to believe and what not to. these videos are fake. lies have been around for thousands of years. deepfakes are new, but they're just lies in a slightly other form. gun control laws have a risk of noncompliance. deepfakes are becoming increasingly easy to make and hard to detect, and that's mainly because of advances in generative ai. in theory, anyone with a computer and the internet can make a deepfake. there's this potential

for videos to deceive people. there is also an example of a deepfake of president zelensky, and that was to admit defeat to the russians about a year and a couple of months ago. speaks ukrainian yet it's actually pretty hard for even experts to work out whether a video is real or not. it's a problem that intel claims to have solved by detecting blood under the skin. so we've come here to check it out. deepfake videos are everywhere now. you have probably already seen them. fortunately, intel is developing several solutions to detect deepfakes in real—time. like the one you are seeing right now. because you are not the real ilke demir — lam. all right, you got me.

the real—life ilke agreed to sit down with the bbc and explain the unusual way it works. if you only try to find the wrong things, sometimes they can be fixed, and you cannot no longer find the wrong things. so we twist that question and we asked what is real about authentic videos, what is real about us. so fakecatcher looks at that question in the sense of looking at your heart. when your heart pumps blood, our veins are changing colour and that colour change is called photoplethysmography — ppg for short. we take those ppg signals from many places on your face and convert them into ppg maps, then we develop deep learning approach on top of that to classify into fake or real videos. in short, fakecatcher looks for minuscule signs of bloodflow in your face, something a deepfake would not have. it also analyses videos for authentic human eye movement. normally, when humans look at a point, when i look at you, if i'm shooting rays from my eyes it's actually converging on you.

but for deepfakes it's like googly eyes, they're everywhere. intel claims the system is 96% accurate and can work on all kinds of deepfakes. so we decided to give it a go with deepfakes generated by mit. i had the answers on my phone. liability protections for companies are more important than individual financial relief for teachers or sanitation workers. so, in the beginning, when it sees very little, it may say, "ok, this, is this a ppg that looks real?" and then it accumulates and has it as fake. this is fake. it finds that at the end 84% accuracy, fake. 0k, interesting. so that was correct. today, as the sitting president in the white house, i still believe that marriage should be between a man and and a woman. i'm a traditionalist, 0k?

really, it's just that i'm a traditionalist. what does fakecatcher say? fake with 66% accuracy. this is...fake. yay! yay. 0k, next one. lam enjoying, are you enjoying this? the system was good at finding fakes, but not so good at working out that a video was real. i'd be working with the leaders of congress now, today. yeah, 0k. ithink fake. saying fake? what is it again? 3bcd. that is actually real. really? that's a real one. ok, that means that the ppg signals are broken at some point. 0k. here's a problem with the system. it doesn't analyse for audio, yet. so often videos that seemed fairly obviously real was still labelled as fake. and pay for the things

we all acknowledge will grow the country... real. that one's real. i thought so, too. just thinking about, i guess the worry would be that in real—time, you say something is fake and it's actually real, you could actually be the one spreading fake news. yeah, that's also true, but verifying something fake versus "be careful, this is 50% accuracy maybe fake", is like, different, weighted differently. deepfakes are going to become more and more of a problem. perhaps fakecatcher will be part of the mix of tools people use to catch them — but it's still by no means the finished article. i beat china all the time, and of course they're mad, they are very, very mad... i don't know about you but i think this last year has been a real inflection point in the world of technology. i mean, ai has been with us for quite long time but it's really making itself known now and it has opened so many cans of worms. totally, we're really starting to see what some of the challenges are. but in identifying them

at least we can start to find solutions. so hopefully it won't be long until we can really enjoy the benefits of ai without so many of the problems. yeah. totally, we're really starting to see what some of the challenges are. but in identifying them at least we can start to find solutions. so hopefully it won't be long until we can really enjoy the benefits of ai without so many of the problems. yeah. true, true, let's hope so. that's all we have time for, thank you for watching and we will see you soon. bye. hello there. well, the weekend certainly started off on a soggy note. plenty of rain around. not that it made much difference for these surfers enjoying the waves coming in at newquay. mind you, there weren't many sun worshippers on the beach. it was pretty dismal, wasn't it? not much sunshine around at all. most of us, you had extensive outbreaks of rain around, although it was a different story across northern scotland, poking out in the sunshine here. actually, we had temperatures

as high as 22 degrees in achnagart in highland scotland, and this beach in tiree kind of looked caribbean—esque in the sunshine. now, back to the wet weather story. we've got more of that rain to come over the next few hours, particularly for northern ireland. across northern england, the rain heavy. the rain easing off for a time over the next few hours for the south of england, southern parts of wales but there'll still be some patches of drizzle around, a bit of mist and fog, too. for the most part, it's not too cool but temperatures dipping down into single figures in scotland, but it is here where we'll start the day with a bit of sunshine. that rain band, then, still with us for sunday morning. the rain probably starting to fizzle off in northern ireland but particularly wet across a swathe of northern england, where we're looking at around 20—30mm of rain. double that over the cumbrian fells, the north york moors and across the pennines could bring some localised surface water flooding. south of that rain band, brighter weather for southern wales, southern england, sunny spells and just a few showers. and with the winds coming up from a south—westerly direction, temperatures do get into the 20s, so it will feel quite pleasant in the sunshine. cooler air, though, for scotland. sunny spells, yes, but there

will be some passing showers as well. 0n into monday's forecast. 0ur area of low pressure that's brought us the wet weather through the weekend starts to clear off into the near continent but before it does do, we are looking at a bit of rain to start off the day, particularly across the midlands, east anglia, south—east england monday morning. that will clear away. and then, it's a day of sunny spells and just a few showers, perhaps some longer spells of rainjust coming into the north—west of scotland, but overall, i think a slightly better kind of weather day. temperatures, though, still below average for the time of year with those north—westerly winds. fortuesday, more showers in the forecast. and then, the middle of the new week sees another area of low pressure. well, you know what that means — a lot more rain on the way. so, no sign of any prolonged, nice, settled sunny spells of weather. it's staying unsettled. showers around on tuesday. longer spells of rain later in the week.

good morning. 0ur headlines today: thousands of people on the greek island of rhodes