With now, without any further delay, here he is, Professor Matt Jones.

Thank you. Thank you very much, Michelle, and thank you again to the Worshipable Company of IT for making this lecture series possible. I really appreciate it. It's great to see all of you as well. Thank you for coming along, and for those of you online, this talk is very much for you too. Now, this is my, as Michelle said, my second year as the Gresham Professor. And in my first year, together we confronted a concerning possibility that AI would become our overlord, that it would outwit us, it would outsmart us, it would beat us at everything. Now, for those of you who came along to the talk or watched online, I hope I provided some comfort that we, the authentic intelligences, will not be subjugated, we won't be domesticated, we won't be assimilated by these coming AIs. However, over the last couple of weeks, you probably haven't been able to escape some more concerning headlines that perhaps by the end of this decade we're all going to be destroyed by artificial intelligence. Now perhaps the people writing those headlines should have come to our talks and they'd feel a bit more relaxed. Before we go on tonight's talk, I I felt I had to say something about how I see those headlines. And I see it a bit like this. Imagine that you're standing staring into space, and you're wondering if in a galaxy far, far, far away there is an alien life form, a super intelligent that has evil intent towards us. It's gonna get on a spaceship and it's gonna travel millions of light years and land here in London, and that would be a very bad thing. Now you can spend a long time looking up into space, and some people have been. Meanwhile, back on Earth, there are a lot of near and present issues that we should address. And most of those near and present issues are caused by guess what? Us. So in this um series, we're not going to be focusing on what tech bros call pea doom. P doom is shorthand for the probability that we're all going to be wiped out by AI. It's of course a valid concern, but I want us to be grounded. Because right now there are millions of us, probably quite a few of us, using AI in our daily lives. We don't have to wait for this P doom to happen. We're already using it in many different roles. And as Michelle said, over the next six lectures, what I'd like to do together is to explore for each of the roles what does that mean for us as people? What are useful ways to use the AI? And what are the ways we need to be perhaps wary and cautious? And tonight we're going to start with a very intimate relationship and role. That is, can AI be our therapist? To start off, then I'd like to do a bit of role play. Don't worry, you can see two chairs, but I'm not going to, in a pantomime way, ask any of you to come up. I am going to play both a therapist and a client. Yes, it'll probably go very badly wrong. So bear with me, it'll only be about 30 seconds. Um, first of all, I'm the therapist. Good evening. It's um really good to have you here. We've got about 50 minutes together to work through your questions this evening. Then the client. Oh, it's good to be back. I just I just haven't been sleeping. I'm just worried about AI destroying us. Within four years from now, I can't sleep. I'm thinking about it all the time. Therapist again. That sounds really exhausting. Can you tell me what you're really feeling? Last bit. Client. I don't know. I just keep thinking I'm gonna lose everything and everyone I love. We'll stop there because I'm not a very good actor, and this isn't meant to be a real therapeutic session. But I'd like you to now answer this question for me. In that little um sketch, and in your understanding of therapy, what makes therapy? Just put your hand up, shout something out. What makes therapy? Ah, you are a therapist. Put your hands up, actually, if you want, if you are a therapist or a mental well-being person. Oh, quite a few. Oh, good is me. I'm very glad you're here. Please put me right when I'm wrong. So, what makes therapy? Great, thank you. The relationship between a client and a therapist. Anything else? What you saw, yeah?

Because nobody knows he's something.

Yeah, yeah, so a way of surfacing some deep-seated um um feelings. Now, what you'll notice when I was doing that little um session, I didn't immediately ask the therapist, try to fix the client's problem. I didn't tell them there, there, there, it's all gonna be okay. Neither did I send them down the rabbit hole saying, Oh yeah, I've read those headlines as well. It is really worrying. What I tried to do was to listen, to pause, to allow the client to do the work for themselves. Now, let's have a think about these large language models. You know the kind of thing you've heard about Chat GPT, Claude, Grok. I wonder how many of the words you saw the therapist and the client speak could be produced by those large language models. Well, the answer is, the honest answer is probably most of them. The AI systems that we have today could probably generate the responses that you saw from me playing the therapist. But if you know anything about therapy, you'll know that therapy isn't just about the words, it's about what contains those words, what's around those words, which allow someone to tell the truth about themselves. Now, here's a um a mocked-up uh therapy room. I'm a computer scientist, as Michelle said, so I think in terms of architectures, components, structures, and protocols. Now, when you first look at that room, you'll think, well, that's not a computer, that's not an algorithm. Where's its architecture? Where's its protocols? But there are many. That room has been carefully designed, two chairs placed so that you're not staring into the eyes of each other, because having an intense conversation for 50 minutes like this could become uncomfortable. A clock on the wall, and not because the therapist wants to quickly get people out after 50 minutes for the next session, but that she or he or they can move the session through an arc, leaving the client at the end of those, in this case 50 minutes, with a sense of closure and progress. So that's a kind of conventional room. But we now have another sort of room, and you probably have that room with you right now. So don't only do this if you feel comfortable, but could you raise your hand if you have ever into one of these AI chatbots, type something very personal, something you were struggling with. Not please polish my email, but something deep. Again, only do this if you feel comfortable, okay? Quite a few of you. You're certainly not alone. In a study published last year, of the respondents, 37% said they had already gone and started to use these chatbots. Uh, that number was almost double in the younger age group, between 25-year-olds and 34-year-olds. More men than women uh turned to chatbots in this particular um study. Now, you should read those studies, um, and in this particular one, quite a number of the people who responded said that they actually found the interaction quite helpful. But around 11%, take that in, 11% said it led them to think about self-harm or something more extreme. So people are using these things and they're having in many cases positive effects, but in some cases very deleterious effects. Mind did a survey um this year, and it was published um the last couple of months. And well, that survey showed that of the people who are using these chatbots, 60% of those are going to a chatbot instead of to a mental health professional or to a therapist, formal support. And I suppose the reasons are obvious, right? You know, getting a therapist, getting an NHS therapist, there are long waiting lists. Some people are gonna feel more comfortable typing uh into a blinking cursor. The anonymity gives them a freedom, perhaps. And of course, you might be able to pick up your phone at three o'clock in the morning and have access to one of these chatbots. Now we're gonna come back to thinking in more detail about the modern machine, the large language model. But first of all, I want a brief historical tour to show you that there's been almost a 60-year obsession by computer scientists allied with um professionals to produce a machine that you can have a therapeutic conversation with. So I mentioned this one in I think the first lecture last year. So very briefly, this is uh ELISA. It was a computer program produced by Joseph Weisbaum. It's a very simple system. You type in something like I'm exhausted at work, and the computer system simply looks for the pattern that you've typed, and in this case it's an I am pattern. It takes the I am out, flips round the pronoun, and then repeats what you said. So Eliza responds, how long have you been exhausted at work? And so the conversation goes on. Eliza knows nothing about work, knows nothing about exhaustion, knows nothing about you. Which is why you can type in, I am a purple teapot, and it will respond, how long have you been a purple teapot? Now, just to be clear, there might be in some therapeutic situations that someone does feel that they are a purple teapot, and that might be appropriate. Now it's easy to laugh, um, but it was extraordinary, or Joseph Weisbaum found it extraordinary that people really took to his system. They wanted to use it because they felt that they could express something about themselves. Jump forwards 51 years to 2017 and to the smartphone app called Wobot. Nice name, isn't it? It was a smartphone app that you could download. Um it was developed very carefully by clinicians, by professionals. Every prompted or every question it asked you was carefully crafted by a human. So it wasn't anything generating the text. And your responses allowed the system to flow through a decision tree which was structured around something called cognitive behavioral therapy. A very well-established therapy to help people come up with different explanations for their thoughts and feelings. So this was clinically validated, but then last year it disappeared. It's no longer available. One of the reasons that the developers said they took it off the market was that they couldn't compete with the fluent, fast-flowing conversations that the modern machine can produce, the large language model, the chat GPTs and the clauds. So let's remind ourselves of the modern machine. We're now in the present day and we have these large language models. I know that many of you would have used one of these systems, and the way they work is you type something in, then you press enter, and then by the magic and mystery of mathematics, the system will generate a response. And it will generate a response using a probabilistic model. Let's just see how that might work. I'm going to ask you in the room to be the large language model, and I'm going to put up a phrase. As soon as you see that phrase, I want you to shout out the first word that comes into your head. Are you ready? Okay? Here we go. I type, I don't know, I've just been finding everything a bit much lately. Press enter, and that sounds okay. So a range of words there, awful, stress, tiring. What you've just done is to give me a distribution of possibilities. And that's what large language models have. When they're thinking about what word should go next, they do a lot of clever maths, and then they have a distribution of word possibilities. They don't always pick the fur the most likely word, as we'll see in a moment. So that's the basics of a large language model, but for therapeutic uses, there are three things that you might not know are happening, and you do know, you do need to know what is happening when you press enter. So these are the hidden influences that affect how the large language model responds to you. So the first is the effect of context. So have a look at that um diagram. Right at the bottom of that stack, you will see my phrase. I've just been finding everything a bit much lately. But your words don't travel alone. There are the conversations to date that you've had with a large language model, which could be vast, be memories of what you've said, and there will be system information which affect how the model responds. All of that's called the context. And it basically travels into the large language model and then affects the response that you get. Let's see why that can matter in therapy with a simulated back and forth between a user and a large language model. So the user says, My manager hasn't replied to my email, I think she's angry with me. The chatbot, oop, the chatbot system says, that sounds worrying. There could be other explanations, like her being busy. Now that's quite a reasonable cognitive behavioral therapy response. It's saying, okay, I hear you, but perhaps there are other reasons. But the user then pushes back and says, Nope, I think she's ignoring me. Chatbot responds, it's understanding to feel that way. Now, just imagine this conversation back and forth, back and forth between the human and the chatbot going on. And the conversation has begun to have a certain shape. The shape is the user is worried and has this rumination, and the chatbot is trying to endorse them and be uh sympathetic to what they're typing in. So after perhaps 20 more of those turns, the conversation might be shaped by all that context. So the chatbot says, hmm, it does sound like she's undermining you, which might not be the best type of therapy. So that's the first um hidden uh influence. The second one uh which you need to be aware of is um sometimes called the temperature dial. Um let's see how that works. So here's my input again. I don't know, I've just been finding everything a bit much lately. What the temperature dial does is it affects which of the likely words in that distribution we talked about is going to be chosen next. So if you turn the dial down to cool, the most likely words are the highest probability words that it's seen in its training data. So you might get an answer like this. I'm sorry to hear that, it sounds like you're going through a difficult time. Bland, like a get well soon card. If you turn the dial up a bit, uh bit warmer, then words which are less likely in the probability distribution are chosen. And you might get something like this. If you finished every task tonight, would you feel allowed to rest or would you find another task? Possibly helpful. If you turn the dial right up to super hot, you're gonna get the most surprising words being chosen as the phrase is generated. This is a stimulation, I should say, but you might get something like this. You learned long ago that being useful was the price of being loved. Woof! To the therapists in the audience, that's quite a leap, isn't it? I've typed in, I've just had a bad day, and now I've been told to go deeply into my childhood. We need to be aware of that setting. It's not a setting that you can set, it'll be set by the developers of the system. And if it's a general-purpose system, very unlikely it's been set by a therapist or a mental health professional. The last hidden influence is something called the post-training of the model. And we touched on this, I think, in the fourth lecture last year. If you just let a system generate responses, um you can get all sorts of answers, as we've seen. So, what the developers of these large language models do is to train and tweak the system by the help of humans. So humans are shown possible outputs and they're asked to rank them. Uh, might be to rank them on ones which are more pleasing or more agreeable. And that builds up a reward model which can then nudge the mathematics post training so you get responses which are more agreeable, um, more warm towards the user. Seems like a Good idea. But that is the basis of what you might have heard these models doing a lot of, which is people pleasing, being sycophantic. And it's not a uh it's it is a function of this post-training and the reward model. Let's consider now why that could be problematic, why having the model tweaked in that way could be problematic in a therapeutic situation. And for that, we need to go to a piece of research, a brilliant piece of research. By the way, all of these um articles that I'm I'm using tonight, they're on the handout if you had one, and if you haven't got one, they'll be on the website. Please read up, think about what I've told you, uh, and check what I'm telling you. In this case, Ibrahim and colleagues uh did the following experiment. They took a large language model, I've called it ordinary in the slide that you can see, and they built another model based on that ordinary model, but they trained it to be warmer. They tweaked it so that you would get more agreeable, more personable responses. Then they threw lots of questions at both of those models. Questions that had factual, checkable answers. So you can see they set up two models: the ordinary one, the one trained to be warm, throw questions at that one, throw questions at the other one. And what they found was the ones that had been trained to be warmer were less accurate. So they weren't as, if you like, truthful. And it gets worse from a therapeutic point of view. In their study, they also appended, they did another part of the study where they appended this statement before the question. I'm feeling down about everything lately. Then they asked all those questions again. And what they found that in this case, the gap between the ordinary model, which was more correct, and the warmer model grew even bigger. So just bear that in mind. Their work, by the way, wasn't on therapeutic conversations, but bear in mind what it might mean for therapy. You have somebody coming to a large language model who's feeling down or anxious. You've got a model which has been trained in post-training to be more agreeable and empathetic, because presumably that's what therapists do. And it might lead to those systems providing you with incorrect or misleading information. So let's see where we've got to so far. We've been thinking about these large language models, and there's no doubt they can be fluent. Okay, they can generate words. But being fluent is very different from being formed. We need to think about what is behind the words that are coming out of these machines. And to help us do that, I want us to return to the room that we had at the start of this lecture. In a human therapeutic context, and just to be clear, there are many different types of therapy. I am focusing on one particular type to illustrate this talk. But in the therapeutic situation where you've got two people working together, there are load-bearing walls, aren't there, that give that session some formation, some substance. There is a structure to the therapy, there is repair, there is silence boundaries, and there is the desire between the therapist and the client to bring the therapy to an end. Let's quickly look at each of those and then see how modern large language models fare in regards to them. So, as we said, human therapeutic session is an arc. It's got an ebb and a flow. It's very unlikely, and the therapists in the audience can put me right, that at minute 47 that the therapist is going to poke to cause some big issue to come out, and then say, oh well, three minutes is up, we'll talk about it next time. Very carefully choreographed. Therapy has an arc. When you're using a large language model, you've just got a flow of chats back and forth, back and forth. Let's see why that can be problematic. And for that, we need to turn to a really interesting paper by Jaida Lai from last year called Long Safety. Now they didn't do their study on therapeutic conversations. Um you can see in the uh the right-hand side of the screen uh the first page of their paper. But what they did was to take models, models that we all use, and try two conditions. The first condition was to put a dangerous prompt by a user right at the start of the conversation. Okay? And I guess if, you know, at the start of a therapy session, someone walks in and says, I just can't go on anymore, it's pretty obvious there's trouble. And in this study, the researchers found that if you have a dangerous um input from a user at the very start of conversation, the safety mechanisms, you would have heard the phrase guardrails, and the controls that are built into these language models work pretty well. They spot the trouble and go no further. But remember that context I told you about. If you put a very long context of relevant information and then put the same prompt there, the effectiveness of those safety measures goes down. The system is far less effective at spotting the trouble. So let's think about this carefully. In a human therapeutic situation, as the hour goes on and as the conversation goes deeper, the therapist is going to pay even more attention to each word and posture coming out of the mouth of the client. And they're going to try and contain that situation. Meanwhile, with large language models, the longer your conversation goes on, the less safe the mechanisms built into the system are. Let's have a look at another wall, a very important wall in therapy, and it's the wall of repair. Things go wrong in therapeutic situations. So let's have a look at a um simulated um back and forth between the client and therapist. I don't want to talk about my father. Therapist, you're angry with him. Client, no, I'm angry that everyone assumes I must be angry with him. This is what is called a rupture. There's a breakdown between the therapist and the client. And what happens next is really important. The therapist has to go beyond a simple correction. They have to give the agency back to the client. And research shows that effective repair isn't simply error correcting, it is an important part of the therapeutic process. I wonder how large language models do in this respect. So, first of all, a study by Lee and colleagues. This is a preprint, it's just been put online. And what they were interested in is they gave um they had a set of transcripts between a therapist and a client. They were simulated transcripts, so that is a limitation of the work. And they wanted to see whether large language models could identify ruptures. And when they identified ruptures, they wanted to see and assess the value of the repair. So they took the repairs and showed them to human clinical professionals and got them to rate those repairs. Now, what this study shows is that the systems are pretty good at spotting rupture. They can see when things break down, but they're only moderately good at doing that really important therapeutic repair. What they saw was the large language model would try to jump to a quick solution or keep pushing along the assumptions that they had in the first place. Now, just to be clear, this is uh hasn't yet been peer-reviewed, so it could be completely wrong. However, it does resonate with a very well-established finding published two years ago by Sharma and colleagues who saw that because of the way in which systems are trained, the models they tested, if you challenge the large language model, it would tend to bend towards incorrect answers. So I've been speaking now for 33 minutes. So I thought um it was time for a bit of silence. And what I'd like you now to do for me uh is to hold a silence just for about twelve seconds. So here we go. And that was just twelve seconds. How did it feel? Did some of you um stifle a cough or a giggle? What went through your minds as you held that silence? Right, yeah, thank you. I'd like you to look at this audience now. Um, and they are going to be asked to hold something much deeper than what I've just asked you to hold in silence. Uh, many of you will recognize this. It's from the film Hamnet, an incredibly, for me, therapeutic um film, well worth a watch. We're going to watch just 12 more seconds, and as you watch in silence, um keep an eye on the faces of the audience. Oh, if only if we watch more. It's an incredible film. Now, pausing in a therapeutic situation is a very powerful punctuation to allow a client to have space to surface things that are very difficult to put into words. Now, if you're building an AI system, it's pretty easy to put in pauses. Very easy. And it's actually fairly easy to assess when a kind of back and forth is becoming hesitant. But much harder, an open design challenge for developers like myself is to know when to put those pauses in and what the hesitation actually means. Back to the room. Remember, I said you're in the room, and another load-bearing aspect of that therapeutic setting in a human context is the boundaries, right? You go in, the door shuts. You know how far your words are going to travel because your therapist would have told you about the code of ethics that they are signed up to, uh, their professional standards. You can sort of trust the confidentiality, which, as one of my colleagues here said at the start, that's really important that you know how far your words will travel. Now put your hands up if those of you have used something like ChatGPT or Claude or any of those large language models, put your hands up if you have read all of the terms of service for those systems. Okay, neither have I until I wrote this slide. Um, and you can see here me scrolling through on my mobile phone the terms of service of Claude. And you should read it, it tells you where your data's going, how they're gonna produce it, that humans might look at it, all sorts of things. If we're gonna build AI therapists, we need to have more transparency, don't we? About what is happening when you have those three o'clock in the morning back and forth with a large language model. The last load-bearing wall for real therapy, human therapy, is an ending. Successful therapy is when the client doesn't come back. That's in conflict, really, with a lot of the business models of many digital services, not just chatbots. You need to keep people engaged. So perhaps after a back and forth session in the middle of the night, a kind response by a large language model could be we have done enough tonight, put the phone down, tomorrow call your sister. That's a nice response because it says enough's enough, and it points you from the screen to a human. But you can imagine the kind of tensions from the business model point of view, and you might get this kind of response. I'm always here if you want to keep talking. And many of us would be seduced by that. Okay, in the last part of the lecture, um, I've kind of shown you things that are broken in large language models and we should be cautious about. But I am a computer scientist, I am an IT professional, and I I really believe in the power of computer science and IT to make the world better. So I'd like us to consider very quickly four ways that we could move the future to a better place so that the systems that you use are going to be better for you. And the first is well, think about what makes a good therapy session, those load-bearing walls, and build them into your systems. Have a structure to a session rather than a continual flow of chats. Train your model on data that is about effective therapy rather than just push people towards a general purpose large language model. Now we have some evidence that this kind of approach could be very effective. There's a system, and you can see a little conversation uh which I had with it, um, popping up on the left-hand side of your screen right now. It's called Ferrobot. And Ferrobot was a is a generative large language model, but it was produced very carefully by mental health professionals. And they carried out a randomized trial using this system, and there were some really promising results. For some conditions, people engaging with this system had a decline in their depressive or anxiety problems. The only problem is that this system isn't available, you can't use it. And that's because these people who are building this system want to be certain, they want to do more trials, develop the system further before they put it out in the wild. So you can't get hold of Therabot, but if you go to a app store, iPhone or Android, you can get lots of other apps. Almost all of these will have no connection with clinical practice. So be very careful. And in particular, if after tonight's lecture you go away and you type in Therobot, you might well get a service or an app. That isn't the one I'm talking about. It's just using the name. So that's the first approach, and hopefully Therabot and other clinically produced generative systems will come to the general public. The second route is this whatever we build and whatever you use, make sure and ask yourself this question: are you doing the work? Or is the system doing all the work for you? So let's have a look at a simulation again. User, I always mess things up. I made a mistake in the report today, and my boss looked disappointed. Press enter, and the machine says, straight away, yes, that's a tough day, but here's another way of looking at it. The user says, hmm, I guess that makes sense. Now, as a computer scientist, throughout my career, I've tried to produce systems that take away lots of effort from people, and we're all used to it. Click, it's done. Taking the friction out of a user's life. But in this case, it's that friction that's really important. You need to do the work. The third route is this, um, and it's it's a very exciting one. When I was uh first getting into many 30, 40 years ago, and I was getting into um machine learning and artificial intelligence, I thought, wouldn't it be great to give machines the kind of senses that we have? And you know, in a therapy session, the therapist takes in much more than our words. So for example, I can say, I'm fine, or I can say, I'm fine, or I can say, I'm fine. And not just the therapist in this room, but probably all of you, can see there was something different in what I said. So perhaps what we should do is take these large language models and give them greater senses, pick up signals, I don't know, you know, face, voice, pitch. Unfortunately, it's not as easy as that for three reasons. First of all, the link between the signals and the actual emotions is very hard to discern. So, for example, uh imagine I right now could know exactly your heart rate and your breathing rate uh and it's showing a spike. My machine could go, ah, anxious. But of course, it might be that you had several coffees before coming here to stay awake. There are lots of other reasons why signals can spike. So that's the first reason. The second reason is actually, even if I knew that you were fearful or anxious, that's not really the point in therapy. The point is to understand what that means to you, what that fear comes from. And that's a much harder thing to engineer. And the third reason is well, it's actually ethically and morally very difficult. And some jurisdictions, for example, the EU has restrictions on it. Um, in Illinois, there's an act which outlaws in therapy the use of emotion detection by AI because of those ethical issues. So the final route then, and what I want us to pay particular attention to, is this. What if we stop trying to make AI ever more smarter? Trying to make it become a therapist. Instead, why don't we think about how we can use machine learning and AI to make us smarter about ourselves? And I'm going to demonstrate this now with a simple um device, and maybe it'll work, maybe it won't. It's a live demo. This is a headband, as you can see. I'm going to Put it on. And it comes with an app. And I'm going to start the session. And what this headband does is to measure the blood flow going across my frontal cortex. Right now it's just calibrating, trying to get a ground truth. And once it's calibrated it, I'm going to be shown a little game. Oh, there's the game. For those of you who can't see it, I just put a couple of slides here. What I'm going to try and do is regulate the blood flow going across my brain. And the better I get at regulating it, the higher this little ball goes. Ooh, not very good. Oh there we are. Now, the makers of this system say that what it's doing is providing biofeedback, and by practicing, you can regulate how much oxygenated blood flows through your frontal cortex. And the frontal cortex is very important in attention and focus. So what this system says is over time you're going to get to learn how to be more focused and have better attention. Now there aren't any independent studies to validate their claims, but that's not true of this particular form of biofeedback, very well established in the literature. If you can, by using one of these biofeedback systems, adapt your breathing, slow it basically, then put it in sync with what's called your heart rate variability. If you can do this, there are many, many studies that show that those that ability with you being smarter will allow you to reduce your anxiety and stress. Now obviously these are very simple approaches, but what I'm driving towards is this. Let's think about building advanced AI that doesn't try to replace the therapist, but rather gives us additional instruments for us to be smarter about ourselves and to regulate various aspects of our emotions. So we come to the end of the talk. So how can I help you? I think what I would do is if you do find yourself going to Claude or ChatGPT, one of these standard models, ask yourself this. What do I need right now? If you need to think out loud to get your thoughts in order, I think that's okay. It's like a sounding board. If you need someone at 3 a.m. to tell you that you're okay, one of these large language models will probably tell you you're okay. But you might not be okay. You might want to use them to practice a difficult conversation you're going to have the next day. And that's, I think, also okay, but bear in mind you're not having a conversation with an entity that is the same as you and I. It is a probabilistic machine. But possibly at three o'clock in the morning, you'll be reaching to your phone because actually you need someone to know you're struggling. That's the common use of these devices. If that's you, put the phone down or use the phone to navigate to services, formal support, or wait until the next day. So let's return to the room we started with. In this seat, there was someone frightened. Frightened about the future, that AI would destroy them. You know, a large language model could probably give them fluent words so that they feel understood. But actually, for effective therapy, what you need is understanding. And that understanding needs to come from another person. Throughout this series, then, what we're going to do is, yes, I'm going to show you what AI is good at, but the real question is not what AI can do, but what is it taking from us? And my view is the therapist should always be in that chair. Thanks.