Chatbots changing minds

0:00Welcome to Tech Life on the BBC World Service, the programme about how technology affects all our lives all over the world. Today, how much can AI influence our political opinions? A new study suggests quite a lot. Actors who are interested in engaging AI to persuade on their behalf would certainly have no problem finding models that would serve their purpose. We'll be hearing from a political strategist and from you about the research. And we look at the web beneath the waves on which so much of our life depends. Internet cables are surprisingly fragile. and when they break, it's a disaster. The Taiwanese have gotten increasingly conscious of the fact over the last few years that cutting these cables is a potential act of war that China could wage on Taiwan at any time.

0:52And they suspect that maybe some of this has already happened. And two unlikely groups have teamed up in Mexico to help conserve salamanders, scientists and nuns. It's a very unusual experience. So it's amazing, really. So the nuns in the monastery care for these salamanders and they absolutely love them.

1:29as this year staggers towards its end we begin with a topic that is i suspect going to become a bigger and bigger issue in the coming 12 months the power of conversational ai chatbots to persuade. Politicians and campaigners of various kinds are always trying to convince us to support their party, their cause. Come election time, a chat on the doorstep with a political candidate, a leaflet through the letterbox are tried and tested approaches here in the UK. Well, ladies, my name's Bradford. Do you know me? Yes. How are you indeed? Very well, thank you. Well, I hope you'll be able to come out, will you, on Monday?

2:08I will be able to be out on Monday. Anybody else in the family? Yes, my husband and my daughter. Good. Three votes. A little bit of archive there. But now a major study by Britain's AI Security Institute and the Oxford Internet Institute has revealed the power of chatbots and other conversational AIs to influence our political beliefs. The research has looked at over 90 ,000 AI conversations with nearly 80 ,000 UK-based participants. To explore what they found, I spoke to the paper's lead author, Kobe Hackenberg, and then took the tape to Hannah O 'Rourke, a political strategist who is a member of the UK Labour Party, to get her reaction as a tech-savvy campaigner.

2:52We used to call this type of thing a stop-the-tape interview, but you can think of it as a radio version of a reaction video. Hello. Hi. Nice to see you. Yeah, good to see you too. want to come through yeah we're in an old building uh in the heart of hatton gardens which is the jewelry district of london uh so what's really interesting in this building is we've got ai companies and jewelers in the same building it's lovely we're in a very sort of victorian i know it's not victorian but it feels victorian sort of little corner here yeah it feels a bit like the scene from the matrix lots of stairs walkways walkways tiles yeah i'll just take you through so why don't you introduce yourself to begin with cool hi i'm hannah o 'rourke i run campaign lab which is a network of progressive technologists data scientists and campaigners who are interested in campaign innovation and how it's changing so we've been speaking to some researchers who've done a big study into how persuasive how politically persuasive AI can be and I thought I'd get you to listen to it and as we go through it maybe tell us what strikes you about it from your perspective as a as a digital campaigner involved in politics.

4:18Okay let's have a listen then. So these new AI systems can do something which is really novel which is that they can engage people in open-ended, multi-dynamic conversation. And this presents the possibility that in these conversations they would be able to persuade people and change their attitudes and beliefs on things like political issues. How persuasive are AIs compared to sort of typical forms of persuasion? I think what we find in our paper is that they do seem to be fairly persuasive and that we observed that across hundreds of issues that we tested, the AI was able to move people's attitudes substantially in just over the course of a nine-minute conversation.

5:06Sorry, yes, you wanted to butt in at that point. We'll stop the tape. Yeah, so I think what's really interesting about this study is when he describes this novel capability that AIs have, this is also what we do on the doorstep as campaigners sort of every day, and this idea of having a conversation. And yes, AI does mean that you can scale this capability. But I think there's lots of stuff that happens in a conversation between humans that may be happening there that an AI maybe can't mimic completely. I think it's interesting to say that this is novel. But yes, I'm interested to hear more. You know, we tested 19 different models and, you know, basically all of them were able to achieve this persuasive effect.

5:47Did you learn which model was the most persuasive? Yeah, well, so actually what we find is that these frontier AI models, from most of these leading frontier AI labs, are all sort of really persuasive. The really interesting thing we found is that these very small models, which are accessible to anyone, can be trained to be just as persuasive as some of those frontier models. That has some implications for safety, doesn't it? Because there's a proliferation of small models, if you like. Yeah, exactly. I think even if all of the big labs were to put on, you know, prevent their models from persuading, as it were, you know, actors who are interested in engaging AI to persuade on their behalf would certainly have no problem finding models that would serve their purpose.

6:35Yeah, I think that's a really good point. I think there's two ways to think about that, though. One is that it means that this capability is not just like in the hands of like big tech, but is more democratized. More people can use AI for persuasion. and in some ways that's a positive thing. I think also there's a deeper point here which is like ultimately in a world where you don't know if you're talking to an AI or not what it will come down to is do you trust the source of information that you are receiving? The democratisation is also going to make the control of say people spending on these systems on what they say, on who's behind it that's going to be a real tough challenge for the authorities isn't it?

7:13Oh yeah definitely and I think as we've seen with things like social media It takes kind of regulators a few years to catch up to like where actual practice is. And I think that's going to be a real challenge. We tested a bunch of different persuasion strategies. But actually, across all of these sort of sophisticated persuasion techniques, the thing that worked the best was just telling the model to use lots of facts and evidence. But interestingly, the more persuasive the model became, using this sort of fact-heavy approach, if you like, the less the accuracy of those facts. Consistently, we observed that all of the things we could do to the model to make it more persuasive systematically resulted in the model being less accurate with the facts that it used in persuasive conversations.

8:02And I think there's two possible explanations for this. One is that the untrue facts themselves were more persuasive than the true facts. The other explanation is just that as the model was trying to deploy more and more facts, the quality of the facts just degraded over time as the model sort of reached the bottom of the barrel of the facts that it knew and was able to deploy. So the facts just became less and less accurate. I mean, that's quite alarming, isn't it? Because that suggests that if you're a, how can I put it, a responsible, ethical campaigner, and you take care to make sure your AI isn't producing untrue facts, if we can live with that mental concept for a while, but producing falsehoods that seem true, you're going to be at a disadvantage when it comes to persuading people.

8:56Yeah, I think that's a really interesting finding. I think it's quite alarming and it was also something that I don't think these models are going to be able to solve anytime soon. There's a definite problem with things like not being able to put limits on what something will say, depending on what kind of goal you're giving it. I mean, persuasion can be good and it can be bad. How do we make sure that what you've learned about persuasion is used for good things? persuading people to do things that are beneficial and not for bad things such as you know increasing social division social unrest all the things that we've seen people use misinformation for if you like we've had lots of you know communication technologies that have transformed the way that we form political opinions in the past from you know television radio to social media now ai will again, transform the way we form political opinions.

9:54I think the key is to not be naive to the risks, but also, you know, we shouldn't be hopeless that this is going to be something that we can't respond to and overcome. The one thing I would end on maybe is just that these are all, these, you know, these results we have showing the persuasiveness of AI models are in, took place in a survey experiment. It's not clear the extent to which these results will generalize to the real world context where people might not want to engage in a very information dense conversation with a chatbot. And so an important direction for future research will be seeing how these effects that we've observed in the lab translate to the field when campaigns actually start trying to deploy chatbots to change people's minds.

10:38Do we need to worry about influence campaigns? I mean, we've seen big influence campaigns, sometimes orchestrated by nation states. I mean, are you optimistic that we can resist things like that? I'm not completely optimistic, certainly not. I think we are heading into very challenging times. But I think it's one of those things where the tools have changed, but the battles are the same. And it's just unfortunate the tools we have now available to us can do things at scale quicker, faster. But the fundamental sort of battle for who do people trust, Who do people want to vote for? You know, what has legitimacy in people's minds is very similar to the ones that we've always fought.

11:23Hannah O 'Rourke listening to Kobe Hackenberg there with a little bit of help from some old fashioned tape sound effects.

11:35You're listening to Tech Life on the BBC World Service with me, Chris Valance. And as always, we want to hear from you, our listeners. And this week, we put a question about politics and AI on the World Service Facebook page. Our producer, Imran Rahman-Jones, has left his throne of power behind the studio glass and is here to tell us what you said. Yeah, after you spoke, Chris, to Kobe and Hannah about the use of AI in changing our politics, we thought we'd hear some of our listeners' views as well. And what did we get? Well, we did have lots and lots of responses. I'll start here with Saif Alahi who commented, AI will reshape politics by transforming communication, influencing public opinion and increasing both governmental efficiency and the risks of misinformation and manipulation.

12:27And there were a lot of people on the Facebook page who posted similar thoughts saying AI is a tool and whether it's positive or negative depends on how we use it. And somebody called Mike replied to that comment and he said most of our older generation has no clue when they are viewing something that is AI and can influence them to vote or not. Well, I think it's worth mentioning that the participants in the study we spoke about were mostly younger, with the vast majority under the age of 50. So it's not just us olds who can be duped by AI. Yeah, I'd say we've definitely seen lots of examples of people of all ages being unable to spot the difference between AI and the real thing.

13:11But I'll finish with this from Gatwetch, who said people rush to AI chatbots for answers instead of thinking critically and getting to a solution themselves. and that's something that I've heard a lot actually from researchers and politicians who are worrying that chatbots will have an impact on our critical thinking skills whether it's for studying or even thinking about who to vote for. Well thanks very much, thanks Imran for helping out with that and thanks to everyone who got in touch and if you've got any thoughts on what you hear in this episode please do email us on techlife at bbc.co.uk or send us a WhatsApp message or even better, a voice note to plus 44 330 1230 320.

14:04Now, four years ago, a volcano erupted, causing devastation across the South Pacific, including in Tonga, a country made up of over 100 islands. The International Federation of the Red Cross estimate that more than 80 ,000 people have been affected by the eruption. The last communication that we received was straight after the eruption and then it's a total cutoff. Right now we are trying our best through our satellite phones, but they haven't picked up any signals. Recovery efforts were made even harder after debris from the volcano damaged an undersea cable, the only cable which supplied the country with internet.

14:49Communications to Tonga remain badly affected, but not knowing of what happened to loved ones is creating unbearable concern. That story from Tonga opens a new book, The Web Beneath the Waves, all about the importance of the networks of subsea internet cables connecting the planet. I spoke to its author, Samantha Bromanian. He told me about the most unexpected consequence. of the internet outage. I think the most surprising anecdote I heard concerned a woman who had kind of gone off the grid almost entirely. She didn't rely on the island's traditional electric grid for power. She had a solar panel installed in her roof, and that was the source of all her electricity.

15:33But a month or so after the internet gave out, she noticed that the solar panel just wasn't working anymore. And she couldn't understand this because it didn't seem like that was connected to the internet at all. But then she discovered that the solar panel, like so much other infrastructure these days, tries to automatically update software on the air every so often. And when it doesn't do that, it just breaks up. And this thing happens to Teslas, it happens to printers, and it also happens to solar panels. But it was just another reminder of how even unexpected elements of infrastructure in our lives ultimately depend on the internet in some way or the other.

16:13I mean, it's an idea, it's a metaphor that you explore, but, you know, so much of our existence hangs on a thread, if you like, and that thread are these cables, and they are quite fragile. They're very fragile. In the deepest ocean, these cables consist of tiny, tiny glass fibres, no thicker than a human head, and maybe you bunch six or 12 of them into a cable and they just have a couple of extra protective coats so that the final product is about as thick as a garden hose. And that's in the deepest part of the ocean where the cable is most vulnerable. As it comes closer to shore, there are other dangers.

16:55So for example, fishing boat trawlers sometimes trawl the bottom of the seabed and they might snag a cable and break it. A ship anchor might snap a cable. So as the cable comes closer to show, companies tend to cloak it in ever increasing layers of armor. That fragility is of great concern when it comes to areas very much dependent on the Internet. And the world depends on their Internet in some respects, where there is a lot of geopolitical tension. And one of those areas is Taiwan. A court in Taiwan has sentenced a Chinese citizen to three years in prison for damaging an undersea communications cable in the Taiwan Strait.

17:42There have been a number of such incidents in recent years off Taiwan, which it blames on China. Beijing denies involvement, saying they are common maritime accidents. The island of Taiwan depends on these cables entirely for its connection to the outside world. And so the Taiwanese have gotten increasingly conscious of the fact over the last few years that cutting these cables is a potential act of war that China could wage on Taiwan at any time. And they suspect that maybe some of this has already happened, but they're acutely aware of the fragility and precarity of all of their 15 cables that land up on their shores.

18:21Can we defend these vital cables? And, you know, the act of sabotage doesn't need to be particularly unusual, could be dragging an anchor. Yeah, that's right. And I think, you know, one of the problems with data cables out in international waters is that we don't really fully understand the legal rubrics that might apply to such cables and to acts of sabotage in international waters. You know, ships are flagged in one country, crewed by people from another country, maybe sailing from a third country to a fourth country. So whose jurisdiction does this cable sabotage fall under? And of course, cables are also fantastic troves of information.

19:12And that makes the cables and their endpoints very interesting to all sorts of people. The cables and their endpoints, but also the companies that own these cables increasingly and that kind of lay these cables. So there's a remarkable stat in my book that says that, you know, at the time that I was researching this, two out of every three new cable projects was being commissioned and funded by one of four American tech companies. And this is Microsoft, Amazon, Google and Meta. The other part of this is that given all the geopolitical strife between America and China over the last decade or so, one thing that has happened is that the Americans are increasingly trying to edge Chinese companies out of Internet cable projects altogether.

19:59This only sort of cements my conviction that the internet has become narrower and more closed and more polarized and more stricken with geopolitics and political strife. And that's Samantha Bromanian, author of the new book, The Web Beneath the Waves.

20:28The sounds of Lake Patscuaro in central Mexico. 126 square kilometres of water, which is the only home to a critically endangered species of salamander. In 2018, a team of researchers went to the lake to test a tech solution to keeping track of the amphibians which were left. They caught 80 salamanders and implanted microchips into them before releasing them back into the lake.

21:07Now the team has published a paper detailing their findings in the Journal of Zoo and Aquarium Research. we read widely on this programme. Adam Bland, assistant team manager for amphibians at Chester Zoo, was on the trip. The species of salamander that I ended up working with is called the Lake Pat Squaro salamander, and their local name for that is the Echoke. They're quite a unique salamander. They have sort of wide fins on their tail, they're fully aquatic, and they have external gills. So you've been putting microchips in these salamanders. Why are you doing that? So a fundamental question with conservation and conserving a species would be, well, how many are there?

21:51And with some species, that's actually very hard to answer. We know that previously there was a lot of them. So local fishermen and local communities would say, you know, they would extract certain amounts of these animals every year. And that was very clearly reduced since the 1980s. it became almost undetectable in the lake where they thought that the species was actually extinct. And then we know that there is a population in the lake, but to find out how many there are, really, you need to understand them on an individual level because these salamanders pretty much all look the same. So we had to develop a method to identify them quickly and easily, which could be used to manage captive breeding programs, but also help us understand how many animals are left in the wild.

22:39I mean some of the animals were wild and some of them were in captivity in a really quite unusual setting. Yes so one of the most interesting things about this project other than the salamander which is a very interesting species is actually there's a group of nuns in a monastery in Patskoro who who actually have been breeding the species for you know over 20 years

23:06and if it wasn't for these nuns breeding the species there might not be this established safety net population

23:17as actually the nuns historically used the salamanders for medicinal purpose and through noticing that they were declining in the lake they decided to start breeding them in the monastery so as part of our collaboration is we support the monastery with the university in Mexico. And we used some of those animals as part of our breeding programme sample animal to chip. And then we monitored them over a period of months to make sure that after we microchipped them that the salamanders had no negative health effects and that it was a reliable method of identifying them. One thing I should ask is, what does the microchip actually tell you?

23:55When you scan the chip, it has a unique code. When you want to know how many animals are left in the wild, you identify them and release them and then you see if you catch them again so you go out you chip as many animals as you can and you put them back in the lake oh i see so it's like it's like it's like putting balls in a bag if you start getting the same ball back again you know there aren't very many in the bag is it with that exactly yeah right it is exactly that so if you if after five years every single animal that you you found in the lake is chipped it's a very good chance that that's how many are left.

24:29So it takes a few seasons of surveying. But if you find the same animals, you then understand how long they live for in the lake. And if you record where you last find them, you'll also find out if they move to different areas of the lake, which we currently don't know. And what did the nuns make of all of this? Are they pleased to have assisted science? Were they watching you nervously? I don't know how to describe them, but the animals they look after they care for were in your hands it is it was a very you know it's a very unusual experience so um it's amazing really so the nuns in the monastery care for these salamanders and they they absolutely love them and when i was microchipping them we had to you know sit down and discuss what we were doing that they have the salamanders best interests you know they really want to protect and conserve the species it has such a strong cultural value to them and it's really nice to partner in that sense because you know it really just shows that anyone can be involved in conservation projects and it's a really nice example of that and we were microchipping the animals under very close watch uh it was quite nerve-wracking but it was it was great and then they allowed access to make sure we could collect the data from the salamanders in later on so it was really nice collaboration and you know they're really passionate about the species and they really love them.

25:54And that was Adam Bland of Chester Zoo in the UK.

26:01Well, my microchip is telling me it's about time to end this edition of Tech Life. Do get in touch with us techlife at bbc.co.uk or whatsapp us on plus 44 330 1230 320. Today's episode was presented by me Chris Valance and our producer was Imran Rahman Jones.