not to get too close. it does look a little bit hazardous. this robot dog hasn't been programmed to walk. it's learning by itself. it's called reinforcement learning. what we've done here is, we've given it this task to try to kind of stand up and move forward. when it does well, one of the computer scientists gives it the digital equivalent of a pat on the head. i can see it's doing better and better. it's strong. i try and put that in its way, and it's stronger than i am. yeah. have you trained it to make a beeline for me? uh-huh. i feel like i'm battling the ai here, and the al's winning. in just a few hours, it's moving around by itself, even coping with the unexpected. i'm really feeling an instinct to help it, but that is, of course, not good. you know, it's learning, and maybe it looks a bit mean sometimes when we push it around, but, really, we'rejust trying to see what it can handle and help it learn new behaviours. this robot dog is the brainchild of professor pieter abbeel. what's interesting about reinforcement learning, and always intrigues me so much, is that the bot collects its own data. but you give it a score. you say, "good robot. bad robot." and over time it starts seeing when it gets the "good robot" signal, and starts behaving more like that. it sounds a bit like bringing up a child. it's a lot like that, it's a lot like bringing up a child. it's a lot like raising a dog, for that matter. it's very, very similar, in that you don't get tojust push data in, like with other learning methods. you have to let it generate its own data, and then you tell it, "this was good. this was bad." it's called generative ai — the same technology that's behind a breakthrough in the way machines use human language. promotional video: gpt-4 - is the latest ai system from openai. lara: the company, openai, was founded eight years ago, and it's revolutionised language—based artificial intelligence. it's a system that can make dreams, thoughts, ideas flourish in text in front of you. last year, openai released an update to its technology called chatgpt. it can answer complex questions, tell you a joke, even write you an original essay or poem. it can also read, analyse or— generate up to 25,000 words of text. it isn'tjust what it's capable of that's changed the game. openai created a free version that anyone could use. for the systems that we're developing to have a big impact, we have to figure out how to make them accessible. pieter abbeel advised openai during the company's early days. around 2018, 2019, openai made the bet saying, "we think that a lot of the key ideas are in place. "maybe not all of them, but a lot of them. "but we're not doing it large—scale enough. "we need more compute, more data, "and things will emerge that are more capable." that computing power arrived in 2019, when microsoft invested a billion dollars in openai. we saw openai four years ago as this extraordinary group of people in a research lab that we thought we could say, "they're on the right path." crucially, microsoft provided it with one of the most powerful computer networks in the world. and that partnership has really blossomed in all the ways that we had hoped, maybe even a little faster than we might have imagined. chatgpt now has more than 100 million registered users. but the latest version isn't free — it's only available to subscribers and will be built into future microsoft products. if you're trying to write a document, if you're trying to create a powerpoint slide, if you're trying to go through your email, this is now a tool that will help you. you can ask it to compose a first draft. take a word document and turn it into powerpoint slides. and as i always say, make sure you read the slides before you present them. you are still in charge. today's ai revolution has been driven by the work of three computer scientists. geoffrey hinton, yann lecun and yoshua bengio developed artificial neural networks to help computers recognise patterns and mimic the human brain. there is still a very large gap that we observe between the capability of ai systems and what we observe in animals and humans. and, basically, the quest of my life has been to bridge this gap, trying to figure out, what are we missing? what type of intelligence do humans and animal exhibit? and there's good news. we're making a lot of progress, and it's accelerating. they've become known as the godfathers of ai. ai is about building machines that are intelligent. now, what does it mean to be intelligent? people may disagree. but the way i and many others like to define it is, the ability to learn, to understand and to take decisions. the godfathers laid the foundations for machines that could learn. what would power the next step was data — but on a scale never possible before. there is the ai system like the chatgpts of the world. those are trained on public data. but for those systems to be useful, they have to be trained on something like a thousand billion... ..you know, a trillion words, something of that order of magnitude, something that would take on the order of 20,000 years for a human to read. so, they require an enormous amount of data. it's pretty much a third of the internet. those billions of words can include material most people would rather avoid. my concern is that the current emphasis on scale in artificial intelligence — which is to say creating enormous training data sets of content — insufficient attention has been given to what's in them. in many cases, it isjust garbage in and garbage out. if you are harvesting the entire internet, you will have so many horrifying things included in there. so, part of this really gets to the core production ethos of how al's made, which is, collect it all and figure it out later. if machines are learning from what's on the internet, material that's harmful needs to be filtered out, and much of that work still needs the scrutiny and judgment of humans. well, there are certainly countries where this type of outsourced labour is most commonly located. china, malaysia, kenya are some examples. but, indeed, it's a growing market for people to be, really, doing this extremely difficult, extremely traumatising work. so, in a sense, you know, it maps to a global north—global south dynamic of extraction and exploitation of labour for these systems. so, this is it, the place where we were engaged for at least four months. and it doesn't bring any fond memories, in as far as the project was concerned. very disturbing. richard mathenge worked as a content moderatorfor a company called sama. sama had been hired by openai to sift through huge volumes of text and remove anything it didn't want its artificial intelligence learning from. the texts, or the wordings that we encountered, were very disturbing, very depressing, things that caused anxiety and mental breakdown. richard was paid less than £2 per hour. he and his colleagues were repeatedly exposed to descriptions of murder, bestiality and abuse. i remember encountering a situation where a human being is having intercourse with an animal. and even as this happens, there is... ..there is an infant who is, you know, right over there watching as the whole thing is taking place. and... you know, it was very disturbing content that you will not want it to dwell in your mind. we are working on this exercise for a period of nine hours per day, with very minimal breaks. richard says he asked sama for support but didn't get any. sama had very poor facilitation, in as far as rendering that support is concerned. we never received counselling sessions. sama and openai parted company in 2022. richard lost hisjob. along with three former colleagues, he's now petitioning the kenyan government to improve conditions for tech company workers. the work that we did was so impactful, in as far as creating a safe space for the people who are using chatgpt right now. what content moderators are going through, it's something that is unfair and it's something that is unbelievable. sama told panorama that employees were provided wellness counselling by professionally trained and licensed mental—health therapists. it said its work with openai was a pilot project that ended early as soon as the team reported concerns. openai said its mission is to build safe and beneficial artificial intelligence, and that it believes the challenging work of content moderation needs to be done humanely and willingly. so, it's established its own ethical and wellness standards for data annotators. machines sifting through and analysing vast quantities of data is powering the accelerating pace of ai development. it's already led to computers acquiring their first real human skill — language. we are at a criticaljuncture, in the sense that we now have machines that understand language, can manipulate language, and even pass for humans for a while with most people. and that was a milestone for al. having mastered language, al's now beginning to master other human skills, like reading emotions. this is the basis of our technology. this is where we can actually analyse everything that's happening in yourface, all the facial muscle actions that you're making, where you're looking. at nottingham university, michel valstar and his team have developed a system that analyses tiny movements in your face to interpret how you're feeling. the eyes are action five, which is your eyes opening. 0k. it's actually a little muscle that comes from the back of your skull and actually attaches to your lid and pulls it open. and what does that mean you're feeling? that is quite often when people want to show that they are very excited, that they're interested. but it can also be a sign of fear, because... right. i was interested. i wasn't scared. there you go. we start by looking at the different facial muscle actions that you exercise, where you're looking. so, your gaze direction. and the tone of your voice, as well, so, what you're saying. and from those sort of basic ingredients of your behaviour, we then determine, for example, different medical conditions that you might have or what emotion you might be displaying. this tech is currently being trialled with nhs trusts in nottinghamshire to help assess the mental health of pregnant women. humans are very bad at picking up gradual changes in their lives. this technology is objective and repeatable, and it will show you a graph, and it should improve as you start doing the treatment. do you think people are ready for al to assess their mental health? well, many people already are, yes, i'm sure. i don't think everybody will want to use it. but that's ok. artificial intelligence is now finding its way into almost every aspect of our lives. in the workplace, it's starting to take on new roles. one of the key things that we've picked up on is the increasing use of al to make really important decisions about people at work. decisions like who gets a job. decisions about how people do their work, who they do it with, where they do it, how quickly they have to do it. and then also, even decisions as significant as whether or not they lose or keep theirjob. in 2018, taxi driver alexandru joined uber. like most uber drivers, he's never actually met his boss. it's a job where we work alone. the only source of information would be your own experience and maybe chatting with other drivers. everything is done through his uber app. then, in the summer of 2021, alexandru got an email. it accused him of fraud. you are guilty or not guilty, there is no grey area in this. i received a warning from uber. "we detected fraudulent activity associated with your account. "please stop. this is your first warning." the automated message said he might have manipulated trip details or falsely claimed money he wasn't entitled to — but it couldn't be specific. i contacted uber, and the answer on the other end of the line was that, "well, we don't know exactly what you have done wrong. "it was something detected by the artificial intelligence. "however, don't worry about it. "just stop whatever you're doing wrong." another warning followed. he was on the verge of being fired. i don't know what i've done wrong. i didn't change anything in my work. how can i stop something that i don't know? alexandru again complained to uber. he received another message more than a year later. this time it was an apology. it was written by uber, black and white. they recognised that, "look, the artificial intelligence made a mistake in your case." so, it is a single mistake made just against alexandru? or maybe... ..maybe the artificial intelligence, it failed to make the judgment correctly in his case, maybe failed in a few other cases. it turns out alexandru wasn't the only one. in april, uber and another taxi company were taken to court in the netherlands. the court said drivers had a right to know how decisions about them were being made, and criticised uberfor using automated decision—making without enough human input. this is a cautionary tale of what can happen when too much responsibility is given to al to make hugely significant decisions about people. so, this was a robo—firing case, where basically the ai decided that some drivers had spent too long waiting for products that they were going to deliver, and that there was some fraud implicated in that, that they were ripping off the company. uber said it understands deactivating a driver's account is a very serious decision, and that specially trained teams are involved in the process to ensure its approach is transparent and fair. injune, the government announced an ai task force with £100 million in funding. it says it will bring together experts to research ai�*s safety and help the uk become a global ai superpower. ai can bring huge benefits for our economy, society and public services. but, of course, it needs to be developed safely, securely and fairly. there are no plans for new laws to control ai at the moment, but the government's published a white paper saying existing regulators will be responsible for policing it. intelligence is kind of the most desirable commodity that we are lacking. so, this will make people more productive, more creative. and my view is that this may bring a new kind of renaissance, a new era of enlightenment for humanity. but it has to be done right. ai development is mostly focused on mastering existing human skills. but could the technology take machines beyond human? hi, lara. hi. good to meet you. thank you for having us. come on in. thanks very much. neuroscientist alexander huth has spent a decade trying to understand how the brain works. if we want to build intelligent machines, maybe we want to make things that act more like human brains. so, that's kind of what got me into neuroscience, actually. this year, his team had a breakthrough. using the same ai tech that can understand language, they've built a computer that can read minds. we scan people's brains with an fmri scanner while theyjust listen to stories. so, we track how their brains respond while they're listening to hours and hours of stories. the team in texas has trained the al on their own brains. as they listen to stories inside the scanner, the computer watches what happens. what sort of brain activity are you looking for? we're looking for brain activity that's related to specific ideas or specific words that appear in the stories. so, for example, whenever you hear somebody talk about parking a car, there are certain patterns of activity in the brain that will be present that reliably correspond to that kind of idea. and we're trying to build up that mapping from this very large data set. the computer looks for patterns in the vast amounts of data from the scanner. with enough training, it can translate brain activity into words. to see how well it works, we picked a story for the lead scientist to listen to. recording: the war of the worlds... scream sound effects play headphones on, he went into the scanner. the martians in the pit had turned the heat ray on them — _ a deep throbbing sound, a silver pencil of light, . and a narrow ribbon of bracken... alexander huth: ..and grass and trees and houses - stretching as far as the eye could see was scorched. and this is the text the ai created from the brain scan. and the decoded version — which misses a lot of things, but it gets some of the good stuff. fired the bullets out of the clip he had on him as he ran, and then i heard a huge crack as a large piece of concrete hit my forehead and then a giant ball of fire. the general idea of big—noise—got—hurt is there in both of them. yeah. and then also the idea that something was on fire, the "scorched" versus "a giant ball of fire". it captured all those things. but clearly not exactly the right words. so far, it can only analyse the brains of a handful of volunteers. but they hope it'll unlock the secrets of how our minds work. we are really pushing on, and a lot of our effort in the lab goes toward using this to actually understand the brain better. right? that's our scientific goal in the end. we want to know, how does the brain work? how do our brains process language? how do we understand ideas? how do we think? sounds great. but what happens if it ends up in the wrong hands? some people are scared, or think that the thought police is coming. ithink... you know, i think it's a fair reaction to this to say, like, "this is scary. i don't want this to happen." that was kind of our reaction, too. like, the first thing that we thought when we got this working was like, "this is fantastic. "it's working!" and then, like, "oh, my god. this is working." are the thought police coming? uh, not yet. none of the current technologies that we have would be effective at actually, like, policing people's thoughts. the mind—reading al was trained using early versions of openai's technology — but openai is not as open as it was before microsoft invested. independent researchers no longer have transparent access to its latest tech and the data that trained it. there's less transparency than there was, and a lot of people aren't able to use chatgpt to do the things that they want it to. i think, fundamentally — even if you think about where law and regulation will go — the world will need an approach that ensures that al systems are safe, secure and transparent. now, we'll have to figure out exactly what transparency means. there will be a balance, as there are in other areas, because you want to foster competition. now, that doesn't mean show your competitor everything about what you're doing. as chatgpt was taking the world by storm, leading figures in the world of artificial intelligence signed a statement, warning of an existential threat — comparable, they said, to pandemics or nuclear war. one of the signatories was yoshua bengio. the existential risk that people talk about really is about what happens if we create machines that are smarter than us and have their own goals that are not aligned with our needs and could potentially harm us, just like a new species on earth that would be smarter than us might do. but the godfathers of ai don't always agree. professor lecun didn't sign the statement. every powerful technology can be used for good or bad. _ the existence of bad actors is not new. but very often the technology that enables bad actors to do bad things is the same one that actually enables to protect against them. a very simple example that exists today — there are people who are trying to use social networks, like facebook and instagram and others, to disseminate disinformation to corrupt the political process, the democratic process. the best protection we have against all of those attempts today makes massive use of ai. even those at the cutting edge of ai research have been surprised at how quickly the technology has evolved. i could not have expected that we would progress as quickly towards human—level ai a few years ago. i wish i had been paying more attention to the questions of loss of control. but now that i do, i feel a responsibility to explain my understanding of what can go wrong. the future of artificial intelligence is no longerjust in the hands of the godfathers. when it comes to the infamous existential crisis letter, a couple of very senior colleagues of yours signed it. what do you think of that letter? i think it's good to imagine the worst, because that's the best way to avoid it. but i don't think that the worst possible thing is necessarily the most probable thing. i think the biggest danger is not what machines will do to people, but what people will do with machines to other people. that's the story, notjust of technology, that's the history of humanity. machines are getting smarter, and ai is now all around us. butjust because it can do something doesn't mean it should. and maybe the question is, how will we live alongside the machines that could become smarter than us? live from london — this is bbc news. on top of the world — spain beat england 1—0 to win the women's world cup for the first time. there was partying on the streets back home — as spanish fans celebrated their country's historic triumph. it was a sad loss for england's lionesses — but king charles praised their spirit and determination. in other news — ecuador heads to the polls, following a presidential election campaign marred by violence and assassination. a huge tropical storm brings high winds and heavy rains to mexico, with hilary now heading to california. and "moscow, we have a problem." russia's mission to the moon ends in failure. hello, i'm gareth barlow. we start in australia, where the spanish women's football team are celebrating, following their 1—0 victory over england in a gripping world cup final. there wasjubilation among the spanish players as they lifted the trophy — crowned champions for the first time. they were joined by spain's queen letizia, who was in sydney for the match with her teenage daughter. cheering. in spain itself, this was the reaction at the fanzone in madrid as the whistle blew at full time. and this, the message from the spanish prime minister on x — formerly twitter... the only goal of the game came from the spanish captain olga carmona, who got past goalkeeper mary earps in the first half. in london, there were tears, hugs and disappointment