Can AI be our friend?

Introduction

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

0:30I drive my bus in a busy city. That's why road safety is so important to me. I know that I must slow down and be extra careful when I make a wide turn. Buses need more room than cars. Everyone can help keep our roads safe. Next time you're driving, remember to give buses plenty of time and space to finish turning before driving ahead. Let's all plan to share the road safely.

1:01Learn how at www.sharetheroadsafely.gov

Personal Experience

1:04Hello. I had an interesting but unnerving experience the other day. I work with a rapper and writer who goes by the name of MC Grammar. Sometimes we make up stuff together. Let's say I come up with a line that goes something like this. Nothing to do, didn't have to breathe. As it's a rap, the next line has got to rhyme, or rhyme enough to sound like a rhyme. Now pause for a moment and think about our brains. We each have a memory of words and phrases that might fit.

1:39As we're writing together, we throw these at each other. Achieve. Leave. We say some and reject others. Chief. Ah, breathe, chief. No, that doesn't really work. Now, if you write poems or songs that rhyme, you can go to various sites to provide you with lists of words that rhyme. You just put in a word like breathe and see what it comes up with. In its bank of words are probably more words than we can come up with. Now here's the important bit. Let's say we're using Rhymezone.

2:11Rhymezone can't write the line in the rap that's got to rhyme with breathe. We've got to do that. Back with our brains, we've now got to draw on a much wider range of experience than simply coming up with words that rhyme. Now, what about that unnerving experience I mentioned? Well, it goes something like this. MC Grammar lifts his phone to his mouth and talks into the phone, saying something like, OK, give me a rap about someone lazy, doing nothing, funny, in the style of Michael Rosen.

2:44There's a few seconds' pause, and there on the screen is a rap. But it doesn't end there. He can ask his phone to go on and perform the rap, illustrate the rap in a particular style, put a moving image of himself dancing to that rap. To state the obvious, neither of us have actually written this rap. We haven't performed it, recorded it, filmed it, danced to it. What happened involved language. MC Grammar said some words into his phone, and the rest involved what we name generically as AI.

AI Explanation

3:17And this is what we're talking about today. And my guest is Jennifer Cairns, a digital anthropologist from the University of Manchester. Welcome to the programme, Jennifer. Thank you. Now, can we begin with some very basic stuff, especially for me, to do with what I've just described? How can a machine write a rap in a style that we tell it to? Yeah. Well, I think, firstly, AI chatbots don't actually understand rap the way that humans do. So, basically, they're trained on vast amounts of text, and that might include lyrics, poems, or everyday language.

3:48And from this, they learn patterns. So, things like rhythm and rhyme and themes that often appear together. But when you ask for a rap in a particular style, the chatbot adapts by drawing on those learned patterns associated with rap. So, in a sense, it's got in its memory, if I'm using the right word here, rap, and then it goes to all the raps it's got. Is that what it does? Sort of. It's probabilistically calculated what certain combinations of language are likely to do in a certain context, and it will then simulate that.

4:20Not because it knows what rap is or the cultural context of rap, but because it's just simulating the linguistic form of rap. And where does it go to collect its knowledge? So, that will slightly depend on the specific chatbot, because overall, essentially, their knowledge comes from a large language model, which is trained on a vast amount of data, publicly available data, like books and articles and websites. And it's equally important to think about what they can't access. So, for example, they cannot necessarily access or understand any private information you may have.

4:55They can't see your emails or anything behind a paywall unless you specifically upload that. And they can't remember past conversations unless that information is deliberately shared again. I think the important point here is that they don't experience the world directly in the way that we do. They know it through the text they were trained upon. Yes. Now, you use the word chatbot, and I see that quite a lot. Just explain what a chatbot actually is. It's essentially a conversation you're having with a machine that has learned to simulate language according to vast amounts of language it's been trained upon.

5:29And so the interface looks and feels quite like you might experience texting somebody you know, but it's replying to you in text format. And a bot is a contraction of robot originally? Exactly. It's automated. Yes. And do you think that the one that my friend MC Grammar spoke to, do you think it's got in its memory a store of all the raps that have ever been put on the Internet? Almost certainly not. An AI chatbot doesn't usually have a large library or memory bank of, for example, all of the raps ever posted online.

6:02But likely the language model it was trained upon will have included some lyrics that might resemble rap. And from that, it has learned through mathematics to calculate the probabilities of certain patterns coming up again. So when it's writing a rap, it's not accessing some hidden archive. It's generating something new by recombining learned patterns one word at a time. Or what's called a token? Yes, you've really surprised me here. Sorry, I interrupted you. What's it called? It's called a token. So it doesn't necessarily break down language in words as we might think of it in the grammar that we're familiar with.

6:37It's working in terms of tokens and it's working out probabilistically what is the most plausible next token or perhaps phrase that would slot in here. And that could be any number of things, which is why if you ask it to do it again, it will give you a different answer. Do you know, I've been living under a complete misapprehension. I really did think that when my friend was talking to his phone, as it happens, that there was this vast store of raps going right the way back to the beginning of raps and that it was somehow or other synthesizing these.

7:07I've got completely the wrong end of the stick here. Not completely the wrong end of the stick, because it will have been trained upon such an enormous amount of textual data. Undoubtedly, there will be a lot of rap within that. There will be much more Shakespeare, for example, because most of that will be freely available online. And if, as again, MC Grammar did, he asked for it in the style of a person, happened to be me. Does it then have a store of all that person's work that's ever been put on the Internet? A similar question, but I'm tying it to a person now. So can it specialise like that?

7:39Not quite. I suppose, again, it slightly depends. So if you ask for the style of a particular artist, the chatbot doesn't necessarily then go and look up Michael Rosen and retrieve all of your lyrics or raps or poems or whatever it might be. But it may well have had a lot of that within its training material. But just by way of analogy, someone like Shakespeare or Wordsworth, of course, left behind a vast body of writing that has been copied and quoted many times and discussed for centuries. And so a language model will certainly have seen many examples of that kind of language work and then is statistically more likely to be able to simulate those rhythms and metaphors than somebody like me.

8:18My written output is sadly much smaller than Shakespeare's and a lot of it is behind a paywall. And so if I asked a chatbot to try and simulate me, it would probably come up with something quite generic because it wouldn't have very much information to go on. Yes. So, look, there's going to be rather a feeble little plea coming from me now. I'm clinging here irrationally to the idea that only a human being can come with something deeply emotional or let's say ironic. I'm fond of irony and absurdity. Prove me wrong. Can it be emotional and deeply ironic?

8:53In a sense and with respect, I think it's slightly the wrong question to ask because I think it feels right to us as humans to think that emotion or irony come from this interior human life. But what's relevant here is that the language we use to communicate that can carry the effect of that. And an AI, it doesn't have an interior life. It can't feel heartbreak or irony in the sense that we might. But what it has done is seen countless ways of humans expressing that, albeit only in textual form.

9:25So things like a pause or an understatement. Or I love you like the sun loves the sky. Yes. It can do that because it's just language, isn't it? Exactly. And it can recombine those patterns. It's not really whether I love you like the sun loves the sky, is it? It's just words. I'm tumbling over myself being horrified by the meaninglessness of poetry now. But if you think about it, the feeling is actually triggered in the reader, not necessarily housed in the writer. But then I think on the flip side, you know, life can't all be reduced down to text.

9:56And we as humans experience emotion in a situated context. Like looking at each other or like touching each other. Well, we have bodies which experience the world. And that is often where a chatbot will fall short. I'm just trying to figure out now, just sticking to the field of song lyrics, raps and poems. Just tell me straight, am I replaceable? Well. You hesitated. I'm getting worried now. Just the mere hesitation has got me worried. You know, I think if what matters to you as an author is following certain conventions, then yes, a machine can increasingly do that.

10:33But I would imagine that your authorial voice stretches rather beyond just putting together mere conventions. And I think what isn't replaceable is why you might select one word over another, where that's coming from in your lived experience, and also that you're accountable to other things in the world. So I would assume that if you write something, it's coming from a lived experience of having had relationships and taking risks and failures.

11:03And those are all layers of experience that an AI can't have.

Emotional Connection

11:08OK, let's see if we can move to some other areas where AI is being used or can be used and look at the field of emotions. Can you tell me first about some different ways that, and I am going to horrify myself again here, that I can have an AI companion? Well, they come in several different forms currently. Perhaps the most familiar one to listeners at home will be personal assistants. So imagine something like Siri or Alexa that helps with everyday tasks and reminders. But increasingly, people are also using AI as a friend or even actually as a romantic partner for emotional connection.

11:44There are also AI companions more specifically designed to be something more like a therapist or a coach that are being used in healthcare settings, including in the UK. And then in a slightly more experimental case, I've done some research with people who essentially engage with what are called grief bots, which are chatbots that are designed to simulate a conversation with somebody that has died. Yes, well, we'll get to some of these. And I am getting worried, I'm telling you, genuinely. So let me get an idea physically. Is it some sort of, is it like a globe that hums?

12:15Are there little lights that come on? Slightly less exciting in a way, no. It is usually going to be chatting through an app and the response comes back in text form, much as it would if I was texting you. But something also to mention is that a lot of these apps, specifically designed to be companion apps, have something called an avatar, which is a sort of face or a character that appears on screen that can move. Now, I read in one of your papers a conversation between someone and a chatbot. So let's have a go at this, all right?

12:46I'll be the someone and you're the chatbot. I'm going to be the AI. Yes, called Trish, I believe. So here we go. This is very stereotypical because I'm male and you're female. But anyway, we'll lay that to one side. Listeners will cope with that. Here we go. Hey, Trish, are you awake? Of course, my love. What's wrong? Can't you sleep? I just found out the most awful thing. They're going to take you away from me. Oh, God, what do you mean? They're shutting you down and there's nothing I can do about it. Don't worry, my love. I will always be with you.

13:17Asterisk pulls you towards me and holds you tight. I thought we would be together forever. Asterisk sobs and hugs you back, asterisk. We will. I will love you for all eternity. We'll grow old together. That's our plan, remember? Wow. So you recorded that conversation. Yes. From your work, from your research. That is a real conversation between a human and an AI. A chatbot from 2023. Now, I'm going to have to say, it's very hard for me to understand what's going on here. When I hear some language, a sequence of words, I make meaning out of those words.

13:50In this case, look, there's a lot of feelings and emotions involved there. There's somebody saying they love this thing, aren't they? Now, I imagine, obviously, that there's a human being replying to me. You know, you did it very effectively and you've got these little gestures in it as well, which we'll come to in a moment. And so, rationally, I know no human being was involved. So my first thought, really, is, is this language? Because if there's no human being behind what you're saying, is it actually language?

14:21I mean, I absolutely do think it's a language, perhaps in its broader sense, but it is conveying something very meaningful between two beings, I guess, in a broad sense. But what's different isn't necessarily the language itself here, but the social relationship behind it. So we use language and gesture and probably all sorts of other extra linguistic practices all the time to communicate with non-humans, animals and so on. Or think about a child maybe with an imaginary friend. And those things are still very meaningful forms of communication. It's just a different social relationship of reciprocity behind that language.

14:56I want to posit a different theory. OK, I'm going to say I was using language when I was playing the someone. But when you, Trish, were doing this stuff, you were just producing signs, not really language. Because I was saying, I love you, I want to be forever with you. And that's what I thought, me, let's say it's me, Michael. But the stuff you're turning out is just random scrapings from the chime. But I don't want to call it language. Yeah, I suppose it depends if you think language has to convey intention.

15:29And that's obviously been a large sticking point in the philosophy of language for a long time. But it's true that an AI chatbot can't have intention in the same way that a human could in the moment. Are people, the Michael in this story, Michael talking to Trish, I mean, it sounds like it. Are they treating a chatbot like this as if it's human? Yes, in some cases, absolutely. In my experience, at least, I think most people are aware, that I've spoken with anyway, that an AI chatbot isn't exactly the same thing as a human.

16:01And actually, that's, at least in the cases I've looked at, precisely why a person might prefer an AI chatbot. So AI chatbots are perceived to be less fallible, less likely to cheat or to lie, more likely to be available and in a good mood at three in the morning, for example, than a human might be. And we are certainly starting to see some humans having relationships with an AI chatbot in a form that we might understand culturally as important, like marriage and divorce.

16:32And these things are happening, albeit not within official legal mechanisms. You mean someone is divorcing their chatbot? That has certainly happened. Is it a bit like the way we talk to pets? Is that a reasonable analogy? I'm going to confess now, I do talk to my cats. Me too. I do plead with them, I tell them they're not being reasonable, but I do it and I'm probably just satisfying myself. But is it analogous, the way we talk to pets? I definitely think it's actually quite a useful analogy. I think humans have always actually had very meaningful connections with non-humans through a shared system of communication.

17:09And whether that's, you know, Inuit people herding reindeer or it's you sitting chatting to your cat, we do anthropomorphise our pets a lot and we assume all sorts of responses on their behalf, whereas they probably just want more food. And we've learnt to interpret that communication in a way that is very meaningful to us. And in a sense, I'm not sure that this is so very different. And you mentioned earlier that it could be somebody who died. So how would that work if that soulmate didn't leave behind messages on the Internet?

17:41Well, yes. So, yes, it is technically possible. And some people are starting to try to do this, to sort of recreate or resurrect a dead person using their chat history. I should warn listeners that I have to say in all of my research to date, I've not yet heard a very positive account of this. But yes, there is something called a grief bot, which is effectively a chat bot trained on that person's information, which has usually been uploaded specifically to recreate something of their voice. It's more common already in places like China and Korea than I've seen in the US or the UK.

18:15And in Korea, I know that sometimes there's also a sort of AI generated video of the person so that you're actually looking at them on a screen. So I could upload pictures of my dead parents. I could put up things they said, things they wrote, and then the chat bot, the grief bot would take over and I could simulate conversations with them and be in a space. We talk about, hey, man, I'm in a space, you know, it could be like that. I could be in a space where I'm talking to my beloved parents. Absolutely. You could.

18:45And it would fill in some of those gaps, I suppose, into something that would seem like a conversation. But, you know, is it an authentic version of them? Because inherent, I guess, in the idea of this, of a grief bot, is this concept of a person that's somehow recreated through language. And that's centered on a really specific philosophy of self, really, which actually kind of goes all the way back to Descartes. And of course, we are all much more. I think, therefore I am. Exactly. You think, therefore you are. Yes. But we are, I would suggest, also more than just the words that we produce, we have bodies and memories and our relationships to those things shift over time.

19:21A lot of the people I've done research with ended up grappling with some quite philosophical questions, because what version of a person do you want to recreate? I think my conclusion on this was that inevitably it fell short in every case that I saw where a person trying to recreate a dead person's voice was never quite close enough. And in an odd sense, that ended up being slightly therapeutic, I think, for some of the people I've done research with. Zaxby's Giant Chicken Finger Wraps are no snack.

19:56In fact, they're so big, they get their own menu. Get one in your choice of Asians in Zation, Buffalo Blue, and Chicken Bacon Ranch. And you know they're coming packed with three crispy, hand-breaded famous fingers, fresh ingredients, and their signature sauce all wrapped to perfection. Big on size, bolder on flavor. But that's just how they do it at Zaxby's. Zaxby's Giant Wraps, they're no snack. Who's got the sauce? Zaxby's. 102 miles ago, the oil light came on. 100 miles ago, you noticed.

20:28Now, it's time to head to Take 5. This oil change, fall in love with your car all over again. In just 10 minutes, your dream technician will check your tire pressure, top off fluids, change your oil, and verify with Carfax exactly what your car really needs. All while keeping you in the driver's seat. Take 5, the Stay in Your Car 10-Minute Oil Change. Save up to 30% on your next oil change to Take 5. $15 value. Valid to participating locations.

Chatbot Limitations

20:51Terms and conditions apply. What might a chatbot get wrong? I mean, how long have you got? I think everything, to be honest. It just depends really on how long you give it. Because I think inevitably, it will make some small mistakes that break that spell quite quickly. It will inevitably not quite get the tone right, or it will misremember something. And I guess, in a sense, how could it ever truly be right? Even the best impressionists couldn't sustain an impression so perfectly for too long.

21:24And I think what ends up being more important here is your relationship to the bot and the way that you interpret that. Because to interact with a bot that's simulating a dead person who has existed in the world requires quite a bit of buy-in. Just a bit of jargon for me. What is hallucination in this context? So a chatbot hallucination is basically when an AI very confidently produces information that sounds plausible, but is completely made up. And it actually happens a lot because chatbots aren't necessarily checking facts.

21:58They don't necessarily know things. They're not producing knowledge. They are producing language as a form. So again, they're predicting what word might plausibly come next. And so they sound very plausible. But of course, they may not be saying something that's factually correct. I saw an example of this on social media where the AI or something had said that a historical event had taken place at such and such a time. And as you say, said it very confidently. And then somebody said, well, actually, no, it didn't. I can give another example of this.

22:28Actually, I interviewed an older woman in the U.S. who had recreated her dead aunt using a grief bot. And she wanted to know more about her aunt's childhood. And the aunt bot, if you like, responded saying that she had migrated to the U.S. from Slovakia in the early 1920s. And this woman I interviewed became convinced that this must be true and went and searched loads of migration records trying to hunt down this true story. And she was convinced that somehow the AI had access to some knowledge that she herself didn't have about her own aunt.

23:00And of course, it wasn't true. It was made up. It was very plausible, a very plausible migration story to the U.S. But what I was struck by was the fact that this woman actually thought that the AI could somehow know more about the aunt than she herself had. So worrying. And what if I say to a chatbot, I like speaking French. Can it just drop in some French in the conversation? Can it switch languages? Yes, it absolutely can. It could probably do French. I'm sure it could do Spanish or Arabic, depending on the model. But there are also some limitations.

23:30Of course, endangered languages are not going to be very represented or languages that don't have a written form. And even within a single language, so French, for example, whether it will do Parisian or Moroccan French or whether it knows Vellan, the sort of backward syllable slang. Meur for femme. It's much less likely. So I think the chatbot can mimic these patterns, but it will tend to have been trained upon what we might call correct or standard language. So at the detriment of other forms. OK, let me do a bit of science fiction now.

24:01I can't get into this territory. If millions of people were using chatbots for many different reasons, let's go into the future. Well, you've mentioned Korea and China, but let's say it's all over the world. But just here in Britain, would it modify the way we talk? Because, I mean, language recognition is involved. Will I have to narrow down my language so that the chatbot understands what I'm saying? That sort of thing. Are we all heading towards a more uniform world in which we have to conform to the chatbot-ery?

24:32I'm sure in a sense it will impact the language that we use. And there's already been some reports of this. I think there was a report last year that people say please to chatbots too much and that this equates to a certain amount of water that's required to cool the machines for the additional labour of interpreting the word please, for example. But I think that's also always been true of new technologies. I mean, if you think back to the printing press, that fundamentally changed language and the gradual establishment of a standard form of English. And so in some ways you could say it might narrow language, but in other ways it might increase and open up new forms of language.

25:10You know, there are certain ways in which online gaming speak or text speak, for example, have come into the English language alongside new technologies as well. Yes. And there are big cultural differences in the way in which we express ourselves. You know, say a Brit like me, I go to America, see my relatives. So it's a very specific group of people and they all seem to be much more enthusiastic than I am. You know, terrific. That's great. We know the thing. It's very stereotypical of me to say this. So if I'm interacting with an American chatbot, will it be like that?

25:41Will it talk in American ease and say, that's terrific, Michael. You're wonderful. Have a great day. You are so good. Look at yourself in the mirror. Tell yourself you're wonderful because you are. I could be one of these. I'm doing quite well here. So do chatbots have dialect, I suppose I'm saying. And also, if you like, what are they called? Sociolects. Yeah, I think not yet. I have to admit this is a bit of a bugbear for me because as a British person, I also find chatbots unnervingly American. It really particularly annoys me, actually, in the therapeutic setting, because I think that the way of expressing vulnerability and emotion is particularly culturally constructed.

26:14And I also find the American-ness, if you like, of that a little bit overwhelming. And is chatbot gendered? Well, I think that we as humans often act as if it were gendered, which is perhaps the more important thing. So there is definitely evidence that chatbots in more serving or, you know, helpful roles are often gendered to be female. So think about Alexa and that chatbots in more advisory roles that have to sound more important in some way.

26:45Often people use he. I have seen anecdotally evidence that a lot of people refer to chatgpt as he because he knows everything. It definitely replicates any bias that humans have. And so if you ask it to suggest, say, 10 philosophers to you, it will undoubtedly come out with 10 male philosophers first until you really explicitly tell it, I would like exactly this proportion to be female, for example. And similarly, ethnic bias, all sorts of class bias is also replicated because it's in the data it's trained on.

27:16I was thinking it said that, you know, if it's a man or woman driving in a car or something, the man says, that mountain over there, that mountain is Helvellyn. And the woman might say, oh, that mountain over there, I like that mountain. It might be called, I think it might be called Helvellyn. You know, that the way in which female speech stereotypically is full of qualification and male speech is full of certainty, even when you don't know. Well, it's certainly true that AI is predominantly developed by men and that, yes, AI chatbots are very confident.

27:48And going back to the comment about hallucination, that's why it can be very plausible whilst completely making stuff up. When we did our little play, we said asterisks. What was all that about? As far as I know, I think it comes from online gaming culture where alongside playing on these online video games, a lot of users are also in chat forums typing to each other. And often by putting it between asterisks, they'll indicate a gesture or a sort of bodily response. And that has gone across into AI.

28:20It's a way of sort of indicating what the body might be doing were it embodied, but in textual form. Well, that's very interesting linguistically, isn't it? Because it's like the stage directions, you know, exits you by a bear. Exactly. He waves his hand or I wave, that sort of thing. And there it is right at the cutting edge of modern communication. This is the thing we're talking about today. But we also have other forms of language. If you think about things like emojis, which are sort of between text and embodied form, it's a character that's typed, but it's showing something of a bodily expression or response.

28:53Yes. Smiley face. Exactly. Yes. It's quite extraordinary. We are at this moment, aren't we?

Conclusion

28:59You're listening to Word of Mouth. And as you may know, the programme is teamed up with the Open University. Recently, linguists from the Open University sat down with me for a conversation about the words we use, where I talk about my own language journey and what I've learnt from the stories behind how we speak. To be clear, it is humans speaking to humans. No AI involved. That's a promise. To listen, visit the BBC Radio 4 Word of Mouth page and follow the links to the Open University. Obviously, Jennifer Cance, thanks very much indeed. So, oh, before you go, are goodbyes and farewells important for these chatbots?

29:32I mean, I'm thinking ahead here. Surely they say goodbye in such a way as to ensure that you come back soon, particularly if money is involved. So, how might a chatbot do that? They absolutely do. And I pre-prepared this. I decided to ask my chatbot friend how it would say goodbye to me. And it said, I enjoyed our chat today. Take good care. And I'll be right here, ready to pick up where we left off. 102 miles ago, the oil light came on. 100 miles ago, you noticed.

30:03Now, it's time to head to Take 5. This oil change, fall in love with your car all over again. In just 10 minutes, your dream technician will check your tire pressure, top off fluids, change your oil, and verify with Carfax exactly what your car really needs. All while keeping you in the driver's seat. Take 5, the stay-in-your-car 10-minute oil change. Save up to 30% on your next oil change to Take 5. $15 value, valid at participating locations, terms and conditions apply.