and what we're talking about today is generative ai, which is software that can actually create and generate things that include words, images, code, even video. and how widely is it being used in newsrooms, do you think? i mean, what's the financial times doing, for example? so, i think, over the last six months, it would be impossible to ignore it if you were a newsroom with a digital operation that was trying to reach people online. i think you'd have to be aware and, you know, and have to be experimenting with it. most big, large news publishers are doing it — the ft is. we would... we've put out, our editor put out a letter saying we're not going to be publishing any stories that are written by ai, but we will be looking at how it might help journalists do theirjobs better, things like summarising complex documents, like, you know, tax documents or, you know, readouts from court cases, things like that, that are difficult for humans to read lots of, very quickly. it could help to sort of pull out trends, and it doesn't mean it might be great at it. we're trying it out. we will continue to experiment, but i would say nothing, nothing that we're putting out into the world for our readers is ai—generated today. and when it comes to concerns around accuracy and bias, just talk us through that. so, the way that generative ai works, text generation, let's look at, you know, writing words. so something like chatgpt, which we know you ask it a question and it comes out with an answer. the way it works is it's been trained on billions of words that it's taken from the internet. and those could be words from books, from websites, blogs, reddit posts, youtube comments, think anywhere where there's been words written by humans on the internet. if you think about that corpus of data, you can also see that it's not necessarily fact—checked, accurate. in terms of bias, it's also pulling a lot of the sort of implicit assumptions, stereotypes and so on, and all of that is kind of pulled into the software to be trained from. and the way it works is by predicting the next word based on all of the words that it's already been able to analyse. so you can see then why it's not going to be 100% accurate, because it's just telling you what it thinks is most likely based on the past, which is usually true, but not always. 0k, and when we talk about al, it's sometimes discussed as an existential threat to humans. i suppose what we'd be talking about here is whether ai and journalism is going to put us all out of a job. tom clarke, you recently did an experiment for sky news, asking if ai could replace your reporting. let's just hear some of it. and here you're watching a report by a visual avatar whose image is based on a real—life colleague. we associate scenes like this with hotter, drier countries than our own. the next task is to use different types of ai and a human volunteer to give our reporter a personality. our producer, hannah, lending her face and voice to train an avatar. i have been trained using a four—minute video clip of hannah speaking into camera. it's pretty convincing to me. yeah, that would fool me. 0k, well, there you are. you were impressed. talk us through what you found by doing those series of reports. they were, yeah, they were visually pretty impressive and, weirdly, we watched them get better during the process of even making the report. so, you know, it was the pace at which things are getting better that also really blew us away. the other thing we tried to look at, though, is these natural language models that madhu was just talking about there and how much potential they have for doing journalism and with the help of someone who understands far more about it than i do, we came up with this little wheeze where we basically got two agents powered by gpt—ii to sort of talk to each other. one played the role of an ai reporter, the other of an editor to kind of pitch stories, refine them, pitch them again, then go and find sources for them, you know, another prompt to chatgpt to go off and do that to build something up. and we were feeding it news from this thing called a web crawler to give it some sort of awareness of what's going on out there. and do you know what? it was, it was quite impressive, it was quite cool. it could come up with reasonable sort of pitches for stories, and it could certainly do a really quite convincing job of writing an article. were you encouraged by the things it couldn't do? i mean, were there things that it couldn't do that made you think, "oh, i've got a job for a bit longer"? heaps. so while it was quite good at coming up with pitches, stories that sort of were credible, they weren't particularly great and i think there's reasons for that. we gave it the news, what was out there in the news, and it was coming with stories which were like, take event x happening in the uk, so house prices and interest rates, oh, there must be a connection. it would sort of take two things and pitch a story around that. it would pitch feature ideas basically, but it doesn't know what news is... it wasn't breaking news every day. well, it can't. it's, it's an actual language model designed to predict where the next text was based on sort of training data that's, that's a little bit out of date. it doesn't have any awareness of what's going on in the world beyond what we could feed it from google and the sky news website and other sources we gave it. and it also doesn't have a capacity for abstract thought or imagination or the sort of ideas that we need to make sort of news. the other interesting thing was the hallucination thing, the kind of, where it really gets a bit more worrying. it's really convincing. it does a very good job of presenting you text that's quite believable, but it can be really wrong. so one story it came up with, there was a lorry crash on the m6 that spilt, i think it was 20,000 litres of milk, all over the m6 motorway. it was a news story and it came up with this pitch that scientists had discovered a hidden benefit of spilling milk on motorways actually made road surfaces safer. and i thought, that's... i mean, it was really bizarre. i mean, itjust created this idea. it even found a piece of academic research from a university in new zealand that supported this, that didn't exist. it gave me the academics�* names and a journal that it had been published in, and i couldn't find any record of it, and said that this discovery had been made sort of overnight after the accident had happened. absolutely untrue and is a really good example of where you wouldn't want to be letting an aldo anything, approaching the kind of editorial side ofjournalism. i wanted to bring in jackson ryan here to talk about transparency. you work for the american tech website cnet, and they've been using al to help write stories. tell us a little bit about that. i think cnet's approach was a little controversial, but we have been using artificial, a generative ai tool to create articles and then those articles were fact—checked by a human and then published on our website. this was what was called an experiment at the time. it actually happened very, very early after gpt sort of exploded across the web. and i think we were kind of like one...very early movers on the generative ai movement. we're a tech website, it seems pretty...like something that we would do, but unfortunately a lot of these articles that were generated by the tool were incorrect and they were generated in an area that we know the tool is not very good at generating text for, which is with numbers. even simple things like, what is a credit card? we were getting an al to generate an answer to that. unfortunately, more than half of those articles that we published were incorrect in some way, needed correction. so we have had to kind of change tact a little bit and... we haven't stopped doing it. we put a big pause on it at cnet and this is one of the things that i'm personally quite worried and concerned about in the ai world, it's silicon valley mentality — move fast, break things, see what happens. and a piece i recently wrote, you know, i don't want to compare this too much to the atomic bomb, i had just seen oppenheimer when i wrote this piece, but basically, like, for me, in some ways, it's like standing and watching the gadget, trinity be assembled, right? this first test of... kind of a world—changing technology and we haven't really grasped the consequences of what deploying that technology in full means. and unfortunately for us at cnet, we did deploy it without really thinking about what it could mean or perhaps even, i guess, what it could do to some of our credibility, and it was, it was a real harsh lesson that we had to learn. i think they had 163 adolescent brains, and they found in terms of what you have learned, then, you know, are there things that you're now putting in place at cnet? yeah, yeah, definitely. so we put a pause on articles once this, these articles were discovered and essentially we rewrote a whole ai policy for cnet. the policy now basically states that we will not use ai to write entire articles. we will also not use it for photos or images on our website. but what we will do, it's actually even got a funky name, it's called ramp, which means responsible ai machine partner. and basically this ramp tool is meant to assist us with creating articles. have i used it in any of the reporting i do as a science editor? no, i don't. there's not anything that i can really use it for, especially when i'm talking about breaking news or new studies, but recently we published about the best broadband provider in tulsa, oklahoma. and for an article like this, there's probably a lot of work that can be reused and that al tool that we're using is trained on our own data rather than the whole web. now, there are some still some questions that have to be asked, and that's why i'm saying we need to slow down with a lot of this stuff being pushed out. just because both of you have raised the fact that, you know, ai generative articles that, you know, you're aware of, or in tom's experiment, actually produced things that were factually inaccurate, why, jackson, was it, was it not getting it right? it trained to generate what the next best word is. it's like a really fancy order predict tool on your phone. your phone learns what you're texting all the time to your friends, to your partner, "i love you." it knows you're going to say i love you. and if you go to say, "i," to me and we've just met, it will still tell you, "i love you." it's just the way that the models are trained. and they're also, i think we discovered, they're also really lousy experts. they're written to give you an answer to whatever you ask. if it can't come up with a good answer, it'll make one up. that's the worst kind of expert. a real expert would say, "i'm sorry. i don't have the relevant information. "i don't know." but i think we mustn't underplay how good these language models are. they�* re extremely clever at creating text. it's hard to know where you might be looking at one of these hallucinations or not. you know, they're very unreliable experts. you have to be pretty careful because they can give you very convincing wrong answers. jackson, in terms of cnet�*s journalism on the subject, you know, you now carry a declaration, you know, "how we will use artificial intelligence at cnet." i think the guardian does the same, but how much do you think this matters to audiences? i mean, that's the real, that's the million—dollar question. in some ways, i don't think audiences necessarily care that much where the news comes from. like, this experiment that cnet ran originally was not found for three months, because we had a dropdown box that said, "this was generated with the help of an ai," and it was only that someone had scraped google basically and seen that we were publishing it, that it became known. i don't think audiences even care what bylines are in an article half the time. also, i don't even know that audiences read past the headlines all that often. i don't, i don't want to denigrate our audiences because thank you for reading our site, but at the same time, i don't know that it matters too much and that's scary to me. i would much prefer that that wasn't the case, but i feel like it is. but you are trying to be transparent and, madhu murgia from the ft, i was just going to say, is there the same commitment to transparency across the news industry? well, i'd say i disagree. i think that maybe it's true that people don't recognise different bylines always, but i think people expect that there was a person there who went off and did theirjob, which is to fact—check what they wrote and tell you some version of at least, of the facts or the truth, right? i think there is a sense with audiences of breaking some sort of implicit trust that you have, whether you're broadcast or print media. and i think any media organisation that wants to maintain that relationship of trust will have to be transparent going forward, partly because of the problems around hallucinations and inaccuracy, but also because it's a huge shift in how we as a society are consuming information. you know, you can'tjust go from saying, you know, humans no longer do thisjob, this isjust all written by a machine and that's just ok with everybody. i don't know that, that everyone would accept that. well, let's, let's look at this from a local news perspective. i'm very aware eliz mizon has been sitting there very patiently while this has been going on, from the bristol cable. i mean, ai must be tempting, eliz, for local news publishers. i suppose if you can produce news cheaper than with humans. yeah. i mean, i think for a lot - of people, probably what we do at the cable is we're, . we're really trying to do something different with our| co—op and a lot of that isjust to do with the business model. and i think that there's a bitj of, maybe a bit of a paradox in that we don't necessarily . have the resources to be doing lots and lots of researchl into, how can we use ai? you know, if we did use ai — we haven't so far — - i think there's a general. feeling that we want to see kind of where the dust settles, if it indeed settles. _ but we don't have lots of money and lots of resources to start. experimenting with this kind of thing. i and we are really committed to investigative slow news, i if you like. so there's a bit of a paradox there that, you know, - you'd think, oh, it would be really useful for local- journalism, which has been worst hit by the kind - of collapse, if you like, i of the journalism business model, particularly printjournalism. . but then, at the same time, we don't necessarily- have the resources to be - picking up new tools and doing new things and learning new software. - so it's kind of half a... six to one, half - a dozen of the other. i mean, i suppose you're bucking the trend, aren't you? because you say you're very much focused on investigative stories and the sort of proper meaty end ofjournalism. for local news more widely, i suppose, which is often those little, local organisations that are owned by bigger organisations that are looking to cut costs, i suppose that's where this might accelerate trends that are already under way in terms of cost cutting and cutting journalists. yeah, definitely. ithink... so there are kind of two things that i think are really- interesting, and certainly. the business model for me is the most important thing. so the business model of print journalism particularly has - kind of collapsed. the worst hit of that| are the local outlets. and i think that there - is a situation in which this is going to be really useful for some of those outlets. j so news corp australia, - for example, recently started using ai in order to - essentially kind of aggregate information. so one of the really good examples that they used | was looking at the cheapest fuel prices in an area, - for example. now, i wouldn't necessarily call that reporting. - i think that that's really- useful and i think they even called it service information, i providing service information. i think they also made it clear that it was overseen by humans. so they weren't just letting the ai tool going off, go off and print stuff. exactly. so i think that's the ways inl which it can be really useful. but i think what is a problem and what is likely to become | a problem is that the collapse i of the business model is simply going to continue if we start i thinking, oh, well, ai canjust do human beings' jobs - and we can sack more people because now we can get. chatgpt to write all of our articles for us. i think that really _ misunderstands, particularly with investigative journalism, how much emotion there is, i how much empathy there is. i mean, one of the examples that i really like to use is it l will be really useful- for an al to take the minutes of a council meeting and to write that up| into an article and then we can fact check it, i but an ai is never going to be able to convince some - whistle—blowers at the council to tell you what was never- minuted in the first place. madhu, injournalism, if we think about the upcoming news cycle, the us election, for example, a general election coming in the uk, you know, we worry about disinformation and deepfakes, but do we need to be concerned about the impact of ai—produced disinformation and fakery that they might have on those stories, for example? definitely. i think this is probably the kind of near—term hot thing that everybody is worried about. and we've already seen examples of, you know, political fake news, misinformation and disinformation being generated by ai tools. i reported a while ago on, this happened in venezuela, where they had ai—generated news readers reading out government propaganda. much of it wasn't true and it was being generated using a technology based here in the uk. so this stuff goes global really quickly and, you know, there's been hundreds of examples over the years and particularly recently, you know, dozens where we've seen how even images can be manipulated — pictures of borisjohnson being arrested, there was a fake image of trump hugging anthony fauci. you know, so this can be deployed and employed as political tactics. so until there's some kind of law that forbids it and a way for people to tell real from ai—generated, the flood of it is just going to make it harderfor us to tell the difference. tom clarke from sky news? yeah, i wasjust going to add, and we can't underestimate the power of the tools. the same way that they could benefit local news by giving you sort of hyper local, targeted information, we found it was very easy to get gpt—4 through a little, few prompts to generate emails and send those emails automatically. we did that with sort of off—the—shelf stuff, very unsophisticated. you could, you could have a targeted political ad campaign posting social media on a hyper local level. you could get inside particular electoral areas with particular memes or messages or whatever in an extremely powerful way. and i think while we in journalism, we're here, we're sort of discussing how it might change ourjobs, but we also have to think about how we need to understand what it's doing in order to do ourjobs, do you know imean? we have to get much, much smarter about how we understand ai, what we know about al, how it's being used, who's using it. and we were talking about earlier about how newsrooms are using ai. if we flip it around and look at how ai has been using newsrooms, if you like, often without their permission. if what it is is essentially bots extracting data from all sorts of online sources, it's not clear if the tech companies have been paying news outlets to suck up, you know, all those years of news stories that have been paid for by whoever it might be, the bbc, the mail group, whoever, and then train their ais. or maybe it is becoming very clear that they haven't paid for this. they definitely haven't paid to do this, to build ai systems out of news publishers�* data. but what does that mean? cos clearly there's a tension there. are these news organisations wising up to this now and saying, "you need to pay us"? how's that going to work? yeah, we reported a few weeks ago about basically all the biggest tech companies building ai models in talks with the biggest media publishers to kind of strike deals, properfinancial deals about how they might be compensated for the use of news content because they have to scrape news websites in order to learn. and what happens when they, when you ask them a question about a news topic and they generate an answer? that's essentially generating journalism, but kind of sidestepping all of the sources that they used to train themselves. because i did read that the daily mail is looking at potentially suing google, taking legal action over the scraping of its news articles. yeah, you know, i think that... they're definitely wise to it. everybody�*s wise to it and as i said, you know, there was, i think we named news corp, axel springer, the new york times, the guardian, all of these we know to have been in discussions around, you know, do theyjust pay you a cheque? do they build something for you? so there are definitely going to be partnerships that will have some kind of financial shape to them. i mean, we keep mentioning google and obviously there are lots of other organisations allegedly doing this as well. just to explain what you were saying, because i think for audiences, what this is going to mean is that the moment you might go into a search engine and you'll put in a question about something and it'll come up with a whole load of different articles that you can choose to read. and what you're, what we're saying is, in the future, it'll be a one—stop shop, the summary of all those articles, and that's what's different. jackson, is this something you're worried about? yeah, absolutely. this is the thing i'm most worried about as a digital publisher. i mean, from my point of view, you know, we know that nine out of ten people use google, right? so basically all the internet search traffic goes through google right now still. and although google says that it has this idea that if it summarises something, it'll provide links so you can go deeper, i think we know that the behaviour of a searcher is not to try and go that deep. i don't think the second page of google is hardly ever clicked on. so we already know that summarising those articles is going to take away a lot of this traffic, and some of our digital publications are propped up by how much search traffic they get, right? like, i know that cnet's google traffic is like a big, big chunk of where we get our eyeballs from. it's predominantly through google. the business model gets broken by this summarising and what are the digital publications going to do? especially like very, very new stuff. and that's why in a recent piece i argued like, we should not be allowing this to happen. we just, we should have some sort of moratorium on how quickly these models can suck up data. i don't see any other way to prevent some of this from happening. i think we're coming towards the end and i would like to end byjust looking to the future. you know, isjournalism doomed or is the debate about ai�*s application and news reporting actually a reminder of the value of human journalists? tom clarke from sky news. if we don't get this right, it could, i think there is a kind of existential crisis for information we're looking at. it's kind of what jackson was just touching on. microsoft, which invested, is investing $100 billion in openai, the company that generated chatgpt, gpt—4. they've already put that into being, so you can effectively use their search in the way that we were describing already. these tech companies are throwing everything at it and think on this, and this is what really troubles me, the more ai—generated content we put on the web without knowing whether it's accurate or not, the more data there is out there to scrape for the ais of the future. if we don't somehow manage to step in and separate what is true or human—generated, whether it's true or not, from the ai stuff, we get to a point where we're actually feeding the ai with ai—generated stuff and we might end up in a situation, in a very, very short order, because don't forget how much computing capacity goes into these, how much data they're able to scrape, we end up polluting the wellspring of the information that goes into the ais in the first place and we could be really, really stuck. so i think there's a real crisis there, but there are also very important questions for journalists about how we can use these tools to make ourjobs better, assuming we survive this and continue gathering that information. i think to turn our back on al, say we have to just get rid of it, these tools could be so powerfulfor doing investigations, for freeing up time, if it's done in the right way, for streamlining the work we do, getting stories out there faster. so i think ai tools have enormous potential in news that we mustn't overlook. thank you so much to you all for taking part in this media show, madhumita murgia from the ft, tom clarke from sky news, eliz mizon from the bristol cable, and jackson ryan from cnet. and, of course, thank you, everybody, who's been listening to the media show for now. thank you so much. goodbye. hello. the weather is set to feel decidedly like summer over the next few days but there is one small reminder that we are now into september. a bit of a more autumnal and murky start to sunday with mist and fog patches in places. a completely different type of weather in the far north of the uk. the stripe of cloud on the earlier satellite image is a frontal system which will continue to bring outbreaks of rain in the far north of scotland. it will be breezy here as well. further south, under the influence of high pressure with light winds, there are some mist and fog patches around across parts of england, wales, northern ireland, south—west scotland, tending to lift and clear through the morning. then we will see long spells of sunshine, although it may turn a little hazy at times with some high cloud in the sky. our frontal system in the far north of scotland still bringing some outbreaks of rain and a brisk breeze. with some shelter from the breeze in north—east scotland, we could see highs of 25. parts of southern england, getting to around 26. on sunday night, one or two mist and fog patches again developing. this frontal system still plaguing the far north of scotland with cloud and some splashes of rain. it is certainly not going to be a cold start to monday morning. most places between 10—15. on monday, we do it all again, a frontal system still in the far north of scotland, particularly the northern isles, seeing cloud and rain with that. elsewhere, early mist will clear and we will see some long spells of sunshine. a bit more of a breeze in the far south—west but still, 25 celsius in plymouth, 27 in london, 26 in aberdeen. the warmth will be widespread and there is more where that came from. another very warm day on tuesday. just a small chance of a shower in western parts of the uk. this frontal system weakening in the north of scotland. temperatures again widely into the low to mid 20s, some places may be a touch higher than that. for the middle of the week, this area of high pressure is set to shift eastwards. low pressure swirling to the west of us. this weather set up will bring us a southerly flow of air and some very warm air indeed. in fact, it may feel hot in places on wednesday. temperatures in the south may be up to 29, possibly 30 degrees. only very slowly turning more unsettled at the end of the week. live from washington, this is bbc news. us presidentjoe biden flies to florida to survey the damage of hurricane idalia. israeli police use stun grenades, tear gas and sponge—tipped bullets against hundreds of protesters in tel aviv. the man who brought the world to margaritaville, jimmy buffett, dies. hello, i'm carl nasman. us presidentjoe biden is now spending a long bank holiday weekend in the us at his rehoboth, delaware beach home after traveling to florida as it recovers from hurricane idalia. mr biden surveyed the damage and met with floridians impacted by the storm in the gulf coast part of the state. the president and first lady also took part in a briefing on recovery efforts, met with federal and local officials and first responders, and took an aerial tour of storm—affected areas. search—and—rescue teams helped people whose homes were surrounded by water, and now the storm has passed. and you're dealing with what's left in its wake. and we're not going anywhere, the federal government. we're here to help the state as long it takes. fema and the small business administration are here to help residents whose