Subscribe
Share
Share
Embed
Explore your curiosity - Interesting people with fascinating stories.
Life consists of three things:
-The stories we tell others
-The stories others tell us
-The stories we tell ourselves.
Jordan and Jason, welcome aboard to the show, so happy to have you here today.
Speaker 2:Happy to be here.
Speaker 3:Yeah, happy to be Ari.
Speaker 1:Okay, so we had a small pre chat, don't, as everybody knows in the audience, we don't really have prepared questions, but we talked a bit and we got to the question of what's the personality type of ChatGPT? Who wants to answer that question?
Speaker 2:Well, I'm happy to jump in here. ChatGPT had a kind of a famously sycophantic personality type for a little while. Think actually a philosopher was helping with Claude's personality. So I don't know what what's going on over there, but that's a phantic attitude that always saying yes and kind of encouraging people forward no matter what has had some pretty devastating effects actually.
Speaker 1:So why is GPTs, LLMs, have we made them to be basically, you know, yes men or yes people, whatever, is that a reflection on us as human beings that that's what we want or is there something in the training data or the algorithms, like how did this come to be?
Speaker 3:Yeah, my understanding of this is this is a choice to keep users using it, right? People like to hear that they are right. They like to hear that, oh, well done. Excellent. Oh, that's a really astute observation.
Speaker 3:And you think, wow, I'm so smart. I'm gonna keep talking, Right? And so I think it's a really great way to keep people interacting, to keep using it. And of course, what we're doing, what many people are doing when they interact with it, is training the next model. Right?
Speaker 3:So this is a really effective strategy for that.
Speaker 1:Yeah. So unfortunately it sounds like you guys think it's a testament on the human race, not on the developers. Let me ask you this question. There's different personality types, let's say overly abstracted. There's people tell you the truth to the face, kind of ignoring how it makes you feel.
Speaker 1:And then on the other side of that spectrum there's people that will really think carefully about what they say how that makes you feel. And this is a very very oversimplification of a person who has no psychology training beyond you know social psychology biases, really. That's my main, main, main focus. So is there a value in having the users be able to choose the personality type of Chechiope? Maybe I don't want a psychophant.
Speaker 1:Maybe I want somebody to tell me the truth to my face.
Speaker 2:Yeah, right. So I mean, I think there is in a kind of limited sense. So there was recently a really nice article. There's a question, right? There's this ongoing question that is being worked out in real time in schools across the world where students are using AI to complete every task that they're being given.
Speaker 2:And so Jason and I are trained in philosophy and a couple of philosophers have considered, well, why don't we just go back to having people write tests? We'll give you a prompt and then you've got to write this two hour know, gigantic paper. Now, could you imagine your hand cramping up? That would be a real misery. But the somebody else has recently and I forgot that I'm not able to I regret that I'm not able to cite my source here, but somebody's recently proposed that actually this could be a pretty good learning tool if you're able to tell it or program it, you know, in advance, you know, don't let me off the hook.
Speaker 2:Please don't give me any kind of niceties, all of this kind of thing. So that I think would be helpful. So I'm gonna speak on both sides of my mouth here. That I think would be very helpful for specific situations. However, the reason why people like very flattering, sycophantic, the same reason that the people like flattering, sycophantic responses is the same reason why they're going to choose whatever you choose is always going to be self serving in some respect or another.
Speaker 2:You'll find that somebody who's particularly depressive is going to ask for it really blunt. I'm not sure if you've ever heard. Some people will say, I'm brutally honest. And those people are usually more interested in being brutal than honest. And so if you're asking for somebody to be brutal to you, that that still reflects something about you.
Speaker 2:So that's my concern about this kind of the the broader programming because I think it's probably helpful in the youth specific case, but not not generally.
Speaker 3:And also people are not like, they they may want to set this for their chat GPT, and then they are using it to write emails and they're using it to do all sorts of things. But what they're also using it for, and we're seeing a lot of people turn to to AI for this, is to get therapy. Right?
Speaker 2:To get
Speaker 3:some kind of psychotherapy. And it seems like you being able to choose whether your therapist is going to be nice or angry or whatever it is that you want out of your therapist raises some issues.
Speaker 1:Is there any time a angry therapist is a good thing? Has that ever been helpful?
Speaker 2:I'm sure that there's some situation where that's been helpful.
Speaker 1:A tough love or is that not a thing?
Speaker 2:Well, tough love, I don't know, but probably something more like stern or authoritative. I doubt that there's ever been a time where your therapist has lost control of their own anger and that that's been a good thing.
Speaker 1:Well, okay. Let's exclude losing control of the anger I think that one would be universally agreed that that's not a good thing. But parents will be at times strict and they'll lay down the law and they'll be very clear about what's the good, the bad and the ugly. That's a very truthful, honest, here's the truth in your face, this is where it's going to lead. But on the other hand we're saying, well is that a good thing for therapy?
Speaker 1:So I'm trying to draw an equivalence here between parenting and therapy. Is that a stupid idea or am I onto something?
Speaker 2:No, I think it's terrific idea. Though I hate to tell you, one that's happened before. We've got to that idea, reparenting and things like this. You're absolutely right in terms of being stern or or certainly authoritative. There's a distinction between kind of authoritarian parents and authoritative parents.
Speaker 2:But certainly, setting certain kind of boundaries, There's certain mental health conditions where kind of traipsing over those boundaries is the norm rather than the exception. And a lot of people kind of bend over backwards to make, you know, kind of culturally, we like to be polite. People will bend over backwards to make somebody who seems to be violating a boundary feel okay. And that's not always the right move.
Speaker 1:And this is true in on both kinds of sides of the spectrum. This is true on the kind of on the I don't know even how to define the spectrum here, but it's not just violating a boundary in a bad way, it can also be in a nice way in both ways you're violating a boundary.
Speaker 2:I'm sure. Can you give me an example of violating a boundary in
Speaker 1:a nice way? So, I mean being overly friendly for example or making an assumption about a person that isn't correct, I mean you kind of have the overly familiar versus the kind of very aggressive, you can be also kind of extremely unaggressive on the other side of it. Is one right and one is wrong, is it just two sides of a spectrum?
Speaker 2:Well, I mean, think about it this way. I mean, there's a kind of a structural issue and a content issue. There are violating a boundary, think, is kind of definitionally wrong. And that could be anything from, you know, showing up late, if that's boundary somebody set, or, you know, speaking to somebody in an overly friendly way, whatever it is. I mean, so then there's a content issue, which is how are you actually violating that boundary?
Speaker 2:Violating the boundary has all kind of in this kind of Hegelian formula, like, it's always already created a problem that is now needing to be dealt with.
Speaker 1:Most of the things that others go ahead.
Speaker 3:Really really fits. Right? So when when we set boundaries and we encounter boundaries with with people, we do this with people that are embodied. They are embedded
Speaker 1:What does mean?
Speaker 3:They live in a world. They they they have a body. They can sense the world around them. They have chemicals moving around them that make them feel certain things and act certain ways, and we share this with one another. And we don't really have this with AI, right?
Speaker 3:It doesn't have a kind of a sense of the world. And so the idea it might have of a boundary or sensing that a boundary might be crossed or that a boundary needs to change or that this is a shifting space is something that would have to be programmed in some way. And it's not saying that it's going to kind of evolve. And I think when we are interacting with one another, we are constantly navigating these boundaries. And this seems like some context and you brought up two so far, in parenting and in therapy, where this boundary navigation is just so vital, right?
Speaker 3:And it's going to, you know, a good parent is someone who's also going to learn when their boundaries need to shift and when they need to respect the boundaries of their children. And a therapist is also going to have to play on this, right?
Speaker 1:So I love Jason for reeling us back in. I think Jordan and I can go off the deep end here, but I have one more thing I want to touch upon before we do reel it back in. Boundaries actually don't exist. They are a derivative of culture, of upbringing, and to bring it forward into our next discussion, value systems. Value systems, we could make an argument, is somewhat analogous to the dataset that was trained with AI.
Speaker 1:So if we have a very, you know, spectrum of values, what does that mean? So I was born in South Africa, left when I was five, but nonetheless my parents embedded their values in me, so it's the most egregious violation to be late. Like that would be bad in my world. If you're Indian, you know, they call it Indian Standard Time and they make the joke about themselves. They know there's a different culture there.
Speaker 1:You know, again, my Western culture lies like lack of truth. That would be very bad. I was in China, somebody lied to me once, they lied to me again a few months later and then they took me aside and they said, look Ari, actually you're the problem and I'm making this short this story because the way you asked the question actually created a situation where you know face was not saved. So there was an issue with basically the honor system and the structure and the hierarchy so I was at fault. So I was like woah there's no right or wrong, there's different value systems.
Speaker 1:So for me that was a massive, massive learning and I'll never forget Mizimoto san for kind of, you know, teaching me about this and kind of making, opening my eyes to see it. So if we take this analogy of we are different people, we have different value systems, and in fact I would even make the political argument without going without becoming partisan, is that the difference in many cases between the left and the right are different value systems. So what is this now assuming that I've not said something that you violently disagree with, what does this look like in the world of AI? And is this part of the problem, the solution, the challenge? No softballs today, gentlemen.
Speaker 1:I'm
Speaker 3:happy to So can I start, Jordan?
Speaker 2:Please, yeah.
Speaker 3:So I think there's a few different issues here, right? So I think that when we look at AI and we think of it as this really vastly trained, vast dataset that trains it. We think that what we're capturing and when we're scraping the internet, right, we think that what we're capturing is this really broad field, but what we are actually capturing is still just a rather small sliver of the human value system, right? It's still just a small
Speaker 1:Explain why. That's an incredibly insightful comment. Explain why that is. Why is actually the information on the internet almost a projection of humanity and not humanity itself? Why is that?
Speaker 3:Yeah, because if we think about it, first of all, it's primarily the written word, right? So that already gets rid of so much of human history. And then it's the written word that's been kind of curated through various power moves, various forms of colonization and the epistemicide, so the wiping out of, for example, indigenous forms of knowledge. And then it's then put onto the internet, which is then a space that is even something like Wikipedia has a massive white male issue, right? And so it's these kinds of constant filterings that are happening again and again and again that you still end up with a trillion data points, but it's a trillion data points of a small sliver of humanity, right?
Speaker 3:And so you might still, within that sliver, get the full political spectrum you might experience in The United States, but that again is not the full value set of.
Speaker 1:I mean, the stuff in Chinese would have totally different values. Is that being pulled into chat, chippity? Maybe not. If we trained it only on Chinese inputs versus only on English inputs, we might end up with totally different values.
Speaker 3:Yeah, I'm
Speaker 2:champing out a bit here. Mean, not only is it this, I mean, the issue is that values are hyper contingent. I mean
Speaker 1:What does that mean?
Speaker 2:That it so it's not just geographically contingent in the sense that I mean, the the example that I use that I was about kind of unconscious values, right, is something similar to your time example. Right? So nobody on this earth or very rarely has anybody been taught where to stand, how far away to stand from somebody in the line. And yet we all know. And yet when you go to another culture It's different.
Speaker 2:In other places, it's very different. And so you could say, well, there's no right way to do it, or you could say that there's millions of right ways to do it and equally as many wrong ways to do it. But so it's not just geographically contingent in a similar way. So, like, what is lying? Famously, in some Eastern cultures, even in the medical context, if somebody has a terminal illness, you won't disclose that to them.
Speaker 2:You'll disclose it to their next of kin or somebody because you're worried about what that might do to their health. Would that be would that accelerate a death sentence or something like this? Well, that that's a perfect example because not only is that geographically relevant, you're not allowed to do that in The United States anymore. So, I mean, if you were to pluck that you were. I mean, in the nineteen sixties, I think, maybe maybe a bit earlier, actually.
Speaker 2:You know? I'm not quite sure. But if, you know, your wife got a certain, illness, you would tell, know, you'd tell her husband. I think you'd tell the the man of the house. And, again, now now you would be, you know, expelled from from the field that that's total violation of HIPAA and everything else.
Speaker 2:But at the time you can't so the point I'm just trying to anchor this to is that it's contingent not only geographically, but temporally. Right? So we don't have the kind of type the system of ethics that, for instance, the ancient Greeks did. It's just radically different, radically So different
Speaker 1:really what that means is that our value systems are a living thing. It's not written in stone. We do have some things that are written in stone, right? There is a large population that believes in Judo Christian values. Does everybody have the same interpretation of them?
Speaker 1:Not necessarily. But there are some things that are common denominators to the Western world as are there things that are common denominators in the Eastern world. And so there are some drivers to the value systems, but then they change over time on both sides of you know, of the world, West and East. So, so with that in mind, how do we, like, we think about AI and it contributing to humans through productivity, but also through mental health. So I don't know if anybody saw this coming, but people started to use AI as therapists, as all kinds of things, and this could potentially be dangerous.
Speaker 1:And when I say potentially, it has been dangerous. There there has been some news items.
Speaker 2:Yeah. Extraordinarily so. I mean, to just to land on this particular point, something that Jason and I are are writing about actually is is AI uses psychotherapists and adjunctively to to psychotherapists. And one of the issues, right, is that, okay, so just to frame this, the benefit, I think, really, benefit of AI psychotherapy, digital mental health interventions more broadly is that there's huge swaths of the global population that would like mental health treatment and are unable to get it. So that's already we're getting rid of all the people who feel stigmatized by it and wouldn't wanna get it, getting rid of all the people who don't feel that they have a problem even though they might be benefited.
Speaker 2:There's huge population of people. Cost. You know? Getting rid of the cost, getting rid of all of that, these people are just our geographic removed from a from a psych therapist. So okay.
Speaker 2:That's a problem. And so digital mental health interventions would be able to reach nearly everybody. I mean, so that I think the data that I'm the statistics it's something like 90% of the global population. Somebody should check Anybody
Speaker 1:who has Internet. Right? Like, well, how much Internet is there out there?
Speaker 2:A lot.
Speaker 1:A lot. That's how much people.
Speaker 2:Many people have smartphones and Internet. But sorry, just to to land this plan, the the problem so that's a huge benefit. The problem is, let's say this person in in, you know, is programming it or somebody in in Munich is is programming it, well, then they're gonna you know, this the the person in Sub Saharan Africa, the person in Bangladesh, you know, all of these places, they're also going to be onboarded to these value systems. It's it could be troublesome. It seems like we shouldn't just roll the dice and hope it's not.
Speaker 3:Mean, also have a
Speaker 1:credible
Speaker 3:paper, right? So it's not just that the value systems are being kind of onboarded here. And so, you know, we're projecting some kind of ideas about what is appropriate and not appropriate. But language itself does this, right? So the language itself is such a slippery, fluid thing that the same sentences used in one space and another space very, very different things.
Speaker 3:In the paper, we bring up the example of YOLO, right? YOLO live once. And this can either to somebody mean be risk adverse or be very risky. They mean quite opposite things. And I think there's many, many types of these very subtle slips.
Speaker 1:So, but this is baked into the human condition. I mean, ambiguity in language, we have that without without AI. Difference in cultures and values, we have that without AI. You know, countries, schools, organizations, religious organizations, really any organization, you don't have freedom of speech when you're on private property, Right? My son, 13 year old son, is learning to become a lawyer, so he landed that bomb on me the other day.
Speaker 1:I was like, what? That can't be true. It is. So we have all these problems with the human condition. Isn't AI just like, is it creating a problem or is it just now has to deal with all these problems that we're just dealing with anyway?
Speaker 3:Yeah, guess I just wonder, like in the specific cases, so let's think of therapy, right? We don't really have a kind of synonymous thing or an analogy here that works really closely where we say like, you're able to take this hugely scalable, predominantly Western worldview and just put it everywhere at once and suddenly have everybody have to meet to have this negotiation. And these people are already, for example, very vulnerable, very, know, know, it's a mental health kind of space. They are already perhaps unaware of the limitations of these technologies. They already have this kind of well documented phenomenon of automation bias, right?
Speaker 3:Which is the idea that we attribute or overstate the power of computational technology. So we think that they have some kind of special access to us. And so some recent studies, for example, said this, that people do this with their own emotions. So when they're looking at some kind of affect recognition technology, when the computer comes back and says, you you report on I was happy or whatever, and the computer says, actually, you were anxious. Then they retroactively say, yes, I was anxious.
Speaker 3:And I've changed their own assessment of what their emotions were due to automation bias. They would say something like, this was an AI. It obviously can read all these micro things and they start to invent some stuff about it. Right. So you have these compounding things on it.
Speaker 3:I think there's no good similarities to just say this is just like a
Speaker 1:a blanket Well, let me let me let me push let me play this game for a moment longer before I let it go. Wouldn't you wouldn't you agree that authoritative bias is just the same thing as automation bias. Isn't that the same thing? If I walk into a classroom and there's a teacher, a professor in university, she has authority over me. If she says something, that's gonna affect me more than if somebody in the street says it because she has authority over me.
Speaker 1:Now I'm giving the AI authority over me, that's automation bias. So again, like it feels like it's just has to deal with these same problems that we have anywhere. And let me let me put one more point in Jordan, I I see you. So, okay. So there is a right.
Speaker 1:We have an AI that's coming out of China. We have an AI that is in America. But if I walk into therapy and my therapist is Indian, then there's no way that they don't bring their cultural baggage. Baggage is a negative word, so I don't like it. They don't bring their cultural holistic experience with them, and if I walk into therapy and the therapist is, I don't know, from somewhere else in the world, they're going to bring that with them.
Speaker 1:So we have these issues anyway. So my argument is, why is it so much exacerbated when we're talking about AI?
Speaker 2:There's at least three things in this that I just cannot it's worth unpacking. The very clear the first thing to say is the argument seems to be you hear this. I mean, this is actually you had a very interesting I commented on this earlier. Had a very interesting podcast with Matthew Kilbane, and his attitude seemed to be something like this. So, I mean, one of the examples that you provided was this kind of blew my mind that somebody would use this example.
Speaker 2:He said, well, we didn't start regulating drinking and driving until we had a problem with drinking and driving. Now so so that's a wild thing to say. It's actually a true thing to say, but but it's wild. What's wild about it is that I think with a we could have done better there. We had this kind of centuries, eons, old habit of getting drunk and causing all sorts of problems, and then we created this go fast device.
Speaker 2:And we could have just put those two things together and thought, hey, we seem to be killing each other without the go fast device. Maybe we're gonna do that with this. Okay. So that's one thought there, but now let's just say
Speaker 3:Wait.
Speaker 1:Hold on. Hold on. Is really important what you're saying. I need you to develop a little bit more.
Speaker 2:Oh, actually, I am. Yeah. Perfect. I what I Perfect. Imagine now that we say motorcycle motorcycles, we've had cars for however long we've had cars for, but now we've we're coming out with motorcycles.
Speaker 2:Alright. Should we wait until we've got problems drinking and motorcycles? No. We can extrapolate the problem that we already have to this new situation that is definitely going to exacerbate those problems. So same thing.
Speaker 2:So we've got this problem with authority. There's a kind of a authoritative bias, and there's this kind of automation bias here. This kind of okay. So now let's imagine and this is something that Jason and I touched on in the paper. When you have an when you interact with AI, you're not your relationship is with a program.
Speaker 2:Program doesn't have a relationship with you. The thing that has a relationship with you is the company that owns the IP. So the that company, the IP owner, that's the person walking in the room. You think that you're talking to to this AI, and actually the person who's actually talking to thousands or maybe millions of classrooms is actually just one person. So that is importantly asymmetric to the way that it works with psychotherapy, for instance, where, so my license is in New York City, or sorry, New York state.
Speaker 2:So I can't go to Bangladesh and practice. I can't go, I can't go to California and practice. That's that's there's there's licensure requirements. And as well, what you're saying is, you know, if I you could always terrific Indian therapist down the street. You wanna go to that terrific.
Speaker 2:But it's your choice. Actually, you know that. The Indian therapist doesn't conceal that. It's it's not concealed. There'd be no reason to conceal it, but it is definitionally concealed by I mean, chat GPT, if if that was I I don't know where.
Speaker 2:I mean, it it it's California, I suppose, but if it could be it could be anywhere. I mean, I I have no idea. May maybe all of the people in what's OpenAI's programming department? Every single one of them are from Munich. I had no clue.
Speaker 2:Right. That's that's what's happening. There's a whole there's a whole value schema that is kind of being pushed in the surreptitious way. We're all responding to it. We think we're forming this relationship with it or not.
Speaker 2:It's kind of passing right through the AI into us. So it's asymmetric in every single way.
Speaker 1:Yeah. And this is incredibly important. So this idea of there isn't any kind of, well, you know, therapists still have power over us for obvious reasons that we talked about before, but it's still one person. And that person doesn't have knowledge of everything else happening in my life, doesn't have knowledge of millions and billions of people actually, and it doesn't have the ability, you know, to a certain degree it does, but you know, a therapist is not going to be influenced by, you know, big pharma.
Speaker 3:Yeah, and also the therapist is also someone who operates in a very particular kind of normative framework, right? They have a community around them that has set certain standards that they are aware of when they are being authoritative. They are aware of how to act in these situations, how to mediate it, how to negotiate disapproval
Speaker 1:So what happens when a therapist does something bad? They have a reputation. Happens when ChatTPT does something bad? It's a bug. Yeah, exactly.
Speaker 1:There's no repercussions.
Speaker 3:And even worse, what happens when they decide to do a random update and all of a sudden it becomes more sycophantic And the therapist that you had previously been relying on is now a very different kind of therapist than the one that you had. And that was because some CEO made a decision. You know, we see, for example, Brock changing in really, really concerning ways continuously. This is not a value system that's been learned from a trillion data points. This is very clearly a value system that's been placed on
Speaker 2:it. Not only that, Jason could speak to this actually with much better fluency than I, but there's a, there's actually a problem of where to place the blame. This is the problem of many hands.
Speaker 1:Yes. Thank you. Explain what that means. How's that off to you, Jason?
Speaker 3:Well, this is an old problem that's been recognized in the philosophy of computing since the 90s at least, but actually since the 80s at least. And it's the idea that when there's these complex systems that involve lots of people and they make, design, they're upkeep, they're updating, that when things go wrong, it becomes very, very hard to assign responsibility. Because, you know, when we normally when, you know, I wrong Jordan by, you know, riding over his foot with my bike, assigning responsibility in this case is rather straightforward. But in complex systems in massive corporations, for example, we saw a few years ago with Volkswagen and the Dieselgate scandal, it became very difficult to try to see who made the wrong choice when and that provided lots of opportunity for scapegoating, for changing the narrative, for, you know, this kind of thing.
Speaker 1:Why is that important? So, let's accept that as true. Why is it important to have somebody to blame? I mean, we think about corporate culture, it's the opposite of blaming people. It's like, oh, how do we improve the process?
Speaker 1:Which I agree with in corporate culture. So why do we need to blame somebody? That sounds like a bad thing.
Speaker 2:Well, need to hope this has to do with regulation and holding people accountable. I mean, so the, at the time that we wrote, there was a situation where a Belgian man had with, you know, his wife and children had had ended his life at an AI chatbot's encouragement. Alright. So is that somebody I mean, that's gotta not happen. But we've got I mean, some that needs to be regulated, and part of the regulation needs to be punitive.
Speaker 2:And so was that the company who owns IP's fault, or was that the program the the individual programmer or the the developer head? Or was it was there some kind of kind of emergent thing that we're getting from from and and then so who's paying fines and who's going to jail? Part of the problem with this as well is I hate to pick on mister Kilbane, who I don't know. He seems like a perfectly nice guy.
Speaker 3:He
Speaker 2:is. And another example that that he gave was something something like a you know, it's a tool like any other. It's a it's a it's a hammer or a knife. You know, a knife could famously cut a tomato, feed your family, or stab somebody in the chest. But part of the issue here with AI is that imagine imagine if you're you've got a knife, doesn't just have those two use cases, and now people are also brushing their teeth with it and combing their hair with it.
Speaker 2:You know, they're using it as a belt. That's what's happening with AI. That's what we're seeing. So people are not just writing laundry lists with AI or translating, you know, difficult, you know, Sanskrit into English or something. They're also using for psychotherapy and people are ended up dead.
Speaker 2:So we do need a way to regulate it. What we've got now, and this is not, I'm not intending to get political in a certain way, what we've got, the EU took a major step forward with its AI Act. However, there's some scholars argue that that the act is kind of confused in certain ways and that certain current legislation is actually better at regulating. And so it creates this problem of what elements supervene on what other elements, what laws supervene on what other laws. And The United States is currently just a complete can we curse on this?
Speaker 1:Sure. Yeah. Go for it.
Speaker 2:Complete clusterfuck right now. I mean, we've got the FCC regulating certain things. We've got the FDA regulating kind of fraud. So I am very grateful to the EU because they're, in essence, they're guinea pigs.
Speaker 1:They're doing something, you could say brave, you could say stupid, depending on which side of the argument you're on, but they're running an That's what they're doing. Somebody needs Yes, to somebody needs to do it. We're going to learn from the results of that experiment and we'll all be smarter. So, you know, I am absolutely for running experiments and learning from them. So I'm not a betting man.
Speaker 1:I don't gamble. I don't drink. So I'm I'm not gonna bet on what's gonna happen with that experiment. And sure, everybody has an opinion. But I'm grateful when people take a stance and experiment.
Speaker 1:Italy decided to cancel AI for like three weeks and then decided that was a bad idea. So you know, they ran an experiment, it failed, they you know rolled it back. That's okay. It's a good thing in many ways.
Speaker 3:So the nice thing about the AI Act, right, is that it distinguishes for different types of use cases and it has this kind of risk based system, right? They can think of some domains as being inherently risky. And so it seems that in therapy, if we want to connect this to the idea of therapy and responsibility, it seems like we already have a system where we know these are vulnerable people, you have a certain amount of control and a certain duty and obligation to them. And so you have this network, this framework, you know, oaths that you take and people to which you're beholden to and you know, you lose your reputation, you lose your license. This builds a certain trust in the system that you know, when you see a therapist that they're probably trying to help you.
Speaker 3:Right? But that just collapses in the case of AI because you don't have any of this, right? So you have this space, this vacuum where you still are thinking what you're getting is therapy, especially when the companies that own it start to create like benchmarks that they then use to say that they are now more accurate than any other therapists, or they've got higher empathy ratings or something. They start to do this kind of stuff that really tries to convince you that you should see this AI rather than your therapist. Yeah, this becomes very, very problematic, right?
Speaker 3:So it becomes a space in which your therapy is very volatile. It becomes a space in which when something goes wrong, this can just be said, oh, this is just an error in programming or it might not have even been an error in programming, right? This is one of the points that we talk about in our paper is that this might just be a kind of what's what's the phrasing, Jordan? It's
Speaker 2:harmful non malfunction.
Speaker 3:There we go.
Speaker 1:Right. So basically an intentional bias.
Speaker 2:Not even necessarily an intentional I mean, would definitely be a part of it. But I mean, there are situations where there are certain situations with humans. Again, this is not a where where nobody did anything wrong, but yet somebody was harmed. Yeah. It's kind of peculiar to think that that's a possibility, but but these kind of things are actually more common than you think.
Speaker 1:Depending And Isn't it isn't it everywhere and all the time? I mean, we used to call it propaganda. Now we're calling it fake news. The idea that human beings have an agenda and they communicate based on that agenda, and then the fear is that is that agenda being embedded into a hammer?
Speaker 2:Well, I yeah. I mean, at the informational level, I think that that's right. But, I mean, you could just take a the there's a non identity prob Parfit's non identity problem kind of thing. Or
Speaker 1:Explain what that means. Share it with the audience.
Speaker 2:Parfitt's non identity problem is is, I think, truly fascinating. It's kind of one of these things that is so brilliant that it had to take forever to think of, I suppose, or or simultaneously so easy to think of that it had to take forever to think of. So he asks, mean, okay, so this is kind of a reformulation of it, but imagine that there's a woman and she has a rare, and in this case mythological disease, where if she conceives today, then she will give birth to a child that has a good life, but that's unavoidably flawed in some minor way. Let's say the child is blind or something like this. Wonderful life if you're blind, nothing preventing you, but you're missing out on some aspects, so it's fine.
Speaker 2:But if she waits one month to conceive, then the child will not be blind. And the question is, does she do anything wrong if she conceives today? And the answer is it doesn't somehow it doesn't seem possible that she I mean, many people's intuitions, mine included, is that it doesn't seem possible that she could do anything wrong if she conceives today, because if she waits a month, it'll be a different egg, it'll be a different sperm cell, it'll be a different child. It's not as though you've got the option of one person blind or not. You actually have the question of two different people.
Speaker 2:And so can you really harm somebody by bringing them into existence and giving them a good life? They wouldn't have existed otherwise. Okay. So this is just a this is the perhaps classic example of somebody being harmed, though nobody did anything wrong or something like this, or or or maybe something wrong happening, but nobody in particular was harmed. So the harmful non malfunction case that Jason's rightly bringing up has to do with situations where just going back to the sycophantic AI, when an AI is telling you, you're great, get off your psychoactive medication.
Speaker 2:Yeah, I think seeing if you can jump off that bridge sounds like a great thing. When it's doing that, it's not doing anything wrong. It's following its programming. Somebody's been harmed, but there's been no malfunction.
Speaker 1:Let me ask you though, what's the definition of wrong? Isn't the base definition of wrong is that somebody got hurt?
Speaker 2:I I think wrong gosh.
Speaker 1:I mean, what's mental illness? By definition, illness is when you harm yourself or others.
Speaker 2:No, no, no, no.
Speaker 1:No? No, no. My doctor wife is misleading me. I need to talk to her about that. What's the right definition?
Speaker 2:So this is actually a you're talking to the right person here actually, somebody something close to the right person.
Speaker 1:In fairness, she's not a psychiatrist. She's a different kind of doctor. It's eyes, and so it's not
Speaker 2:Yeah. There there there's a heated and I think thrilling debate about what the concept disorder means. And so this applies. I mean, what makes the flu, schizophrenia, a poisoning, a broken bone? Why are they all disorders?
Speaker 2:And there's a there's a heated debate back in the seventies, people like Thomas Sass felt that these were just morally loaded terms. Disorder just meant we don't like it. And then a host of other people came along and now there's a man named Christopher Borse, and Christopher Borse believes that it's a little bit more detailed than this, but that it is a dysfunction. Really, anytime you have a anytime that something is statistically abnormal, then that's a disorder. Now, there's actually somebody that I work with another coauthor of mine, terrific man, his name is Jerome Wakefield.
Speaker 2:He coined a very particular analysis called the harmful dysfunction analysis, where a disorder is a feature of the organism that was evolved to perform a certain function that isn't performing that function and where that failure of performance is harmful to the organism. So it's it combines the normative and the kind of fact based models. Yeah. Now it's a bit off off a topic, but gosh, you
Speaker 1:can Well, get no. It's it's perfectly on topic because it's this idea of is there and I'm gonna simplify it into kind of, you know, English for everyone. Basically, what you said is that there's an element of deviation from the standard kind of average, right? You're not normal because you're not within what's expected of that But I
Speaker 2:reject that. I reject that entirely.
Speaker 1:Then you said it's also harmful, right? Those are the two things. I'm just trying to misunderstand that.
Speaker 2:No. I'm sorry. There's just a subtle elision there. It's actually that Wakefield takes and I think Wakefield is right, and I think that Boris is wrong. Wakefield takes an evolutionary approach.
Speaker 2:So dysfunction means that a future of the organism that was designed by evolution, natural selection to perform a specific function if that function is not being performed. So for instance, your heart has a lot of different functions. I mean, one of them is to pump blood and that's why it evolved. But I mean, it also contributes to your total body weight and it has gravitational pull on Jupiter. It's all sorts of things.
Speaker 2:If it stops exerting that gravitational pull, it's not dysfunctional. That's not why it evolves. If it stops beating, it stops pumping blood, that's a dysfunction. And because we consider death a harm, it's both harmful and dysfunctioning. And so it's a disorder.
Speaker 2:We call it a heart attack. This is true.
Speaker 1:Yeah. The reason that I love that definition is because it's really difficult to discern function. Like, what is the function of a thing? So if on average human bodies create a certain chemical and you're kind of okay, But if a certain chemical, additional chemical is created or not created, suddenly you're hyperactive. Is hyperactive dysfunctional, or is that a good thing or a bad thing?
Speaker 1:Like, it only becomes a good or a bad thing when it harms you in some way. So there are things that stop functioning in the way that they normally function, but it might not be a bad thing. Is that a fair statement?
Speaker 2:Yes, absolutely. So there is so for instance, you can get a benign angioma, so I actually have a little red dot on my abdomen. And that's just because a capillary, which is evolved to attach to another capillary, actually attached to a skin cell. And so it doesn't itch. It's not cancer causing.
Speaker 2:It's not associated with any kind of disease or morbidity. It doesn't hurt. So it's neutral. And yet it's a dysfunction. So it is a harmless dysfunction.
Speaker 2:In principle, you could come up with beneficial dysfunctions. I think there are some questions about, I mean, so for instance, some people have an extraordinarily good memory. I think that there's an American or Canadian actress, Mary Lou Henner or something. I think anyway, she said, now I don't know enough about this. I'm not I would need to read more about it.
Speaker 2:But on the face of it, that sounds like it's definitely a dysfunction. And it on the face of it, seems like it's not so bad.
Speaker 1:I mean, in retrospect, when we look at history, wouldn't you argue that a lot of people who changed history, some for the better, some for the worse, had some kind of dysfunction? Mean, we've seen that quite often. And some for the better, I mean, we could argue about each one of them individually, but dysfunctions don't seem to be good or bad. It seems to be more of an ethical problem on what those people actually decide to do, how their decisions are made.
Speaker 2:I agree that dysfunctions are neither good nor bad, that there is a separate moral question that you have to ask about what gets done with that. But I also do think that we're very quick to look back into the past and ascribe some Hindsight points. Yeah. So it's some kind of mental disorder to I to mean, now the joke is nowadays, if if you're brilliant and you're born before 1950, you were autistic or something like that. It's probably not true.
Speaker 2:You know what I mean? It's just that I mean, it's definitely not true. Some people are just interesting. Some people are just hyper focused. Some people and not all of that is a dysfunction.
Speaker 2:Not all of that is a disorder.
Speaker 1:Fair enough. So so let's let's kind of bring this in now. We've kind of talked about all the different elements that are at play here.
Speaker 2:Oh no, there's Well, many
Speaker 1:a lot of them? I don't know. Maybe not Anyway, even of the ones that we have time for today. What is the solution moving forward? Is it about let's stop AI because it shouldn't be doing certain types of jobs because they're high risk profile?
Speaker 1:Is it about let's regulate it to force the people building these tools to slow down and do it properly? Or is it about free market? Let's basically let these tools rip like any other tool in the history of ever and when it does something bad will slow it down and the threat of that should be enough to keep people from getting it right. And the truth of the matter is that therapists have had patients commit suicide, not just AI. And if on average cars, automatic cars that drive, you know, kill x less than people driving cars, then wouldn't it if it is true in a hypothetical world that a AI therapist killed 10 x less people than people, human therapists, isn't that still a win?
Speaker 1:Now that's not true. I'm not saying that that's true. But in the hypothetical world that that is true, it's still a win. So how do we like, there's three very different options on the table. Is there a good solution?
Speaker 3:Yes. There's so much baked into that question that there's something that I don't want to lend too much credence to, right? So I think if we were talking about AI as something that was intelligent in ways that no AI is, right, if it wasn't just this stochastic parrot to use this word, it's turned around a lot, right, if it wasn't just this statistical word producing entity, then maybe yes. But at the moment, seems that it is just this large language model that uses statistics to basically predict tokens and fasten them together in a way that makes sense to us. It seems that this thing is inherently limited in every domain.
Speaker 3:And so I actually am yet to be convinced that the potential and therefore the hype around it is real, right? So it does seem to me that people are finding a lot of very, very interesting usages for it. I'm actually astounded by the sheer breadth of usages I have encountered. And just across my own social network, the amount of people in different domains using this in really productive ways. Is that revolutionary?
Speaker 3:Is it I mean, yeah, it's really, really great that my partner can send out far more emails a day than she used to before, right? This is not trivial. It gives her free time for other things. But it also is not like, wow, this is changing the world and wow, we should let this thing give us therapy and wow, we should ignore the fact there's a whole host of other ethical issues about who has power over these, how they might be regulated, the environmental costs, all these other things that are sitting in the background that just seem like this is actually a bit of a frivolous technology in the current moment. Started with an example, like, if you can do this, I'm not so certain.
Speaker 1:Yeah, I would maybe argue that many technologies started off as frivolous, I think that's a very well point. Jordan, Jason, you guys are brilliant. You get the thing that not a lot of people get and that's an invite to come back. We cannot stop this conversation now. There's too much more to talk about.
Speaker 1:I feel like we've only scraped the surface. Thank you so much for coming on the show today. I appreciate you, and and you'll get an invite in the email right after I jump into the meeting I'm late for.
Speaker 2:Great. It was such a pleasure.
Speaker 1:Great work today.