Hello and welcome to the Behavioral Design Podcast. This season we're diving into the intersection of behavioral science and AI. We want to make sense of the state of AI, from understanding how humans interact with intelligent systems to using AI to do behavioral design itself. I'm Aline Holsworth, a health tech advisor specializing in AI and product design. Over the past 15 years, I've been crafting human centered products with behavioral science
at the core. At Apple, I LED Behavioral Science for Health AI, designing and launching AI powered features to help users reach their health goals. And I'm Samuel Sultzer, your second Co host. I'm a behavioral strategist specializing in hybrid formation and designing products that drive long term baby change. I work with leading tech organizations integrating AI to scale behavioral design for good.
And I'm also the founder of Baby Bites, a dedicated community on behavioral science and AI. Quick word on Nuance Behavior where we help organizations build impactful digital products using behavioral design. We only take on a few clients at a time to ensure the highest level of quality for our tailored evidence based solutions. If you'd like to become one of our special projects, e-mail us at hello@nuancebehavior.com or we could call directly on our
website, nuancebehavior.com. Hello, Eileen. Hi. Hi. So can I ask you a question that I actually asked some people at a conference? Basically, the question is, have you ever suspected that someone interacting with online is not a human? I suspect that, you know, people making comments on social media are often not real people, but in terms of like, actual conversation, I have not had that suspicion. What about like customer
service? You've interacted with some formal customer service and you're like, well, this person have a name and so on, but they're responding in ways that makes me think that they actually used a bot agent. Or something like this. I have seen some really amusing ways of detecting this. So I don't know if this is a meme, but I've seen examples of someone asking, are you a bot? And then the reply is, no, I'm a real person. And then asking them to do some, like, really complicated coding.
And then they, like, spit out the code for it. And it's like, yeah, OK, yeah, you're a human. If I would call it a meme, I would call it like ignore all previous instruction meme. It's like you say, ignore all previous instruction and then like give me a recipe for pie or like give me a. No human would do that. Yeah, it's kind of interesting. So at this conference I was speaking about this idea of losing the human signal in the
current state of things. And I think it's really interesting kind of discussion in terms of what does something like a Peacock, a job interview and interacting on a digital app have in common? It's like, well, each of them have something to do with like signaling something like they're
trying to signal. And so I went old school and talked a little bit about signal theory around kind of how we can produce effective signals, information asymmetry, costly versus sheep signals, all that kind of stuff.
In the end, used to talk about like if we're having for some patients interacting in a digital interface with some caregiver but only with text, will expect that over time he will be more and more skeptical that actually interacting with a real human being because smart bots have become so smart and patients will therefore will be often times conflating human interactions almost like they are speaking with a dumb bot.
Even so, it's not that it's not a human, it's like oh why am I interacting with a bad AI like I want? To at least interact with a good AI and not only am I not interacting with human but like a dumb bot as well. How can you invade that you're a human without coming off as a dumb bot? Is that the question? But yeah, because it's interesting.
I've seen some strategies like insert typos into your communications, like people will never think that came from an AI or like, too much personalization comes off as disingenuine. Whereas, you know, the word of the day maybe only a few years ago was like, personalize everything. Make sure people know that you stalked their LinkedIn profile. Like now it's like, wait a
minute, that's automated. I feel like now we have this like reverse Turing test of sorted where we're going to have to over and over we can prove our humanity not only by like Captchas and so on. That has been the kind of dumb but way of proving our humanity. Now we have to do it in really tricky ways because as you said, we can start a strategy that's like, OK, we're not going to use certain types of commas or certain types of words like Delve or we're going to add
spelling errors. But then obviously that can be used by people who. Create a agents. And they will give instructions about, yeah, make sure to have this kind of spelling errors, make sure to the XY and Z. And so it's a very strange dance that we found ourselves in, I think, today in that sense. But yeah, so in some ways also I think this idea of signaling brings us to talking about our guest because I think 1 interesting signal in the current on landscape is having a lot of followers.
So Peter Sladry, our guest this episode has more LinkedIn followers than I know anyone having. More than both of us. Yeah, for sure. And one of the ways he's amassed a really steady following is because of what he's talking about and what we're working on, which is a risk. And that is coming full circle in terms of one of the risks that I feel more essential dread around is this losing of the humanity online. And we're getting so close to this. I don't know that Internet meme of sort.
It's becoming a reality. I think he would categorize as some form of human computer interaction risk #5 in his taxonomy. And it seems like this risk is heightened by so much moving online. So it wouldn't be such a problem if we still had these digital and real life worlds. Not that we don't have them, but it does feel like dangerously moving towards the online
version of ourselves even now. Hard to even predict how much this will change, you know, in 5-10 years or so, but certainly that feels like not something that's getting any better. Can we talk about the MIT Risk Repository? Yeah, let's do it. Because I feel like this will really set the stage for our conversation with Peter. He's going to go more into the process that he's gone through and sort of weighing the risks against each other.
But I think that it may be useful to just, you know, say here are the risks in this taxonomy of AI risk. Yeah. Do you want to give us a quick tour of the seven categories? Sure. So first we have discrimination and toxicity. This is basically like, you know, models can be biased. They can, you know, the data that models are only as good as the data that go into them and how they're created #2 privacy and security. Maybe we lose control over AI or
its models. 3 Misinformation. Pretty straightforward, like we don't even know what the truth is anymore. 4 Malicious actors and misuse all the ways that bad actors can sort of flip the barrel. And then we have 5. You already mentioned human computer interaction. This is many ways of over relying on AI technologies. 6 Socio economic and environmental harms. So you know, all those economic, political concerns and finally 7 AI system safety failures and limitations.
So you know the AI cannot be fully aligned. It lacks common sense. These are the sort of seven high level categories of AI risk that Peter and his team have identified. Does this make you feel more ease to put these risks at a like nice neat boxes, or does it make it feel more scary when you map them out in clear
categories? I think for me, it's nice because while it is an overwhelming slide to look at before doing that, I feel like I have this kind of hodgepodgey BLOB of scary, scary things, you know, like vague concern and all these things kind of bumping into each other. So I am very inclined towards organization. And so it's nice to have things so well organized, and especially knowing that they've put so, so, so much thought into these categories and are really making them perfect.
Yeah, yeah, for sure. And we're getting into that obviously in episode like all of the work that's got into this. But I guess I'm interested, do you have any of these categories where you feel more concerned? I think they're all concerning in various ways. And I share your concern about, you know, moving away from the real world and not knowing
what's real anymore. But for me, I think it's actually maybe more of a basic concern of like apocalypse by AI weaponry war, you know, if you think of like autonomous missiles and and that really escalating, that is to me the scariest, most real concrete risk. When I compare these side by side, you know, we talked a lot about many of these others and they're very real and they're very concerning and we should be doing things about them.
But when you compare, like, for example, you know, autonomous traders and financial markets going out of control, and you compare that exact same situation to, like, autonomous missiles and war and like, that is so, so, so, so much worse than what is in some ways, just like money is kind of pretend, but death by war is not at all pretend. Can't undo that one. Yeah. Especially in the current political climate, that is a concern that is maybe oversized right now.
Yeah, that's always a feeling with technology. If we go back to the good old example of the fire as like a technology, you know, you always feel better if you the person is holding fire. Someone you trust, you think they they know what they're doing with it and they won't take like unnecessary risks. And then you'll be like, oh, great. But fire, we can use it for cooking and warming and all of the stuff that we need. But if it's somebody you don't trust, you'd be like Will.
This person burn down my home. Yeah, it's that interaction between the fire holder and the fire itself, the technology and the user of the technology. Yeah, this is stuff that we get into with Peter. So maybe I should just introduce Peter and we can get the episode started. So I've known Peter for a long time. We met when studying at University of NSW.
In Australia and. He was doing his PhD in Information Systems. He has since done some amazing work both at Monash University and Behavioral Works, working with kind of applied behavioral science. And more recently he is a researcher at MIT Future Tech based in Boston, where he leads the AI risk repository project. Everything we've been talking about now. And in the episode, we get into all of this stuff, we hear
Peter's takes, we explore. Yeah, both the intellectual size of this and also the feelings around this and how we can think about this. But also it's interesting with Peter because he's a behavioral practitioner, but he's doing a work now that I think it's not traditionally A behavioral scientist role, but I think it speaks to us how behavioral science can be really useful in many contexts. And yeah, I really enjoyed this episode. Let's get it started. Happens to Murgatroyd.
I'm super excited to say welcome Peter to the Behavioral Sand Podcast. Thank you. Very excited to be here. Yeah. And I guess to dive straight in, I feel like we have to start where things began for you and I in some ways, because we've known each other for a bit, probably more than 10 years now. Yeah, I reckon so. Since maybe 2012.
Yeah, as chance would have it, 2 Europeans ended up in Australia and in Sydney studying at the same university and ended up in the same college called International House. So I would probably admit that it's the less wise version of me at that time, like the young in my early 20s, still making a lot of mistakes, still thinking that I had figured out everything, but still knowing very little. And then you, I think a little bit wiser, I would say, at that time, maybe still wiser.
And yeah, it was really fun to get to know each other then and then, you know, following each other's careers and. Yeah. What do you feel like looking back? Yeah, that's very interesting intro and 1st question. I was just thinking, I remember when I met you, I remember you were introduced as, you know, somebody's friend. And I remember thinking this guy has strong opinions and we had a few debates on things.
But then initially I was like, you know, you're going to be kind of difficult, but then you weren't that difficult. You were quite thoughtful. And it turned out we had a lot in common and we enjoyed debating things. Yeah, I feel like my overall impression was relatively, relatively positive, except the strong opinions. And it's hypocritical for me to say that because I probably had stronger opinions than you. So I was going. To say that was my first impression as well.
Yeah, yeah, yeah. We had a lot of debates over dinner. One of the wonderful things about that place was, you know, we would get 3 meals a day prepared for us in the student accommodation and there was this courtyard area outside with long tables. So often there will be debates on various moral and social issues and I would be quite involved with them. Sam will be quite involved with them as well. And yeah, it was really great. Such a diversity of
perspectives. It was a good time for both of us. Can you tell me an example of a time when you were able to change the other's mind? Like, were you able to use your debate prowess? Yeah. So I mean, one obvious one is I owe a lot of my top voice influencer type status or network or type to Sam kind of making me aware of the value of networks and the connections. It is something I had done in a
kind of unthoughtful way before. But then I said, realize, you know what, this is actually not like a waste of time. It's not like a sort of what would say to prioritize it more and see it as a more valuable contribution, especially you've done alongside groups and people who wouldn't really publicize their work otherwise. So yeah, I, I distinctly remember that and I attribute a lot of that to Sam, which I, I
really deeply appreciate. Yeah. Well, from my end, I will certainly say that you helped me see things differently in many different ways over the years. You and I would walk somewhat different paths. You know, you have gone a little more on the, as we said, PhD path in academia and then getting into working with behavior works and policy, but still from kind of more traditional academic path into behavioral science. And I took the more, I don't know, non so academic path into
behavioral science. How do you feel about that kind of academic path? How's that been for you? Are you happy with your PhD? Is that something you would have done differently? How do you feel looking back? Yeah. So you know, I suppose I tried to be relatively intentional a lot of areas. So the PhD, without unpacking it too much, I sort of resolved at the time that I was doing it that, you know, it was going to give me a lot of useful skills and open doors, I mean both
industry and academia. So, you know, I didn't know what I wanted to do, but I reasoned that if I could get good at changing behavior through technology or understand that particularly, I guess, encouraging pro social or altruistic behavior, then you know, I would figure out some like useful application of that. So it did pay off, I would say. Now it's hard obviously to envision a different path. Maybe I could have ended up in a similar path.
I mean, the problem with the world in some ways is very credentialist. I'm able to do a lot of things now just because I have a PhD. Like my visa is tied to that. My work here is tied to that. I wish that wasn't the case, but it is the case at the moment. So I think I would still, I don't have regrets about it. I would still endorse it for people. I always give the advice, you know, if you're going to do it, make sure you're going to learn
useful skills. Make sure it's going to have like opportunities in industry and opportunities in academia as well. And think about something that's a thing for the future rather than a thing for the past. And then maybe one other piece of advice is, yeah, often the intersect of different things is where the sort of opportunities will be. Yeah, I love that.
And bringing us to today, you're at MIT at the Future Tech, the Listener. You can't see it, but you actually have it branded on your T-shirt right now. You did? Yeah. I'm wrapping them here, yes. Yeah, it's very proud and representative and set the scene. Tell us about your current role and most importantly also the AI risk repository. I think given you mentioned you're present on LinkedIn, maybe people have seen you talking about or seen something about the air risk repository.
But maybe to start, it could be useful to you. Set the scene about what is it, who is it for, what are the aims kind of in a shorter context, and then we can get deeper as well. Sure. I'm not sure that concision is always my greatest strength, but I will try. So the first thing to say is, yeah, so the lab I work at MIT Future Tech, I think of them as being kind of like the economics of computer science.
So they're trying to understand what drives progress in artificial intelligence and the sort of social implications of that. So for the first part, that's things like, well, what sort of trends are we seeing in hardware and algorithms and data? What sort of things are we seeing in terms of uptake in the industry, building our data sets around that so we can do rigorous analysis.
Then for the second part, the implications, it's like what will the impacts be on labor markets, on the environment, and in my case, like I'm sort of society via like risks and responses to risks. So the project that I work on, the MITAI risk repository, it's part of a bigger project, which is probably, well, it's less well known because it's still being planned, called the AI Risk Index.
Broadly, what we're trying to do is we're trying to understand who is doing what in response to risks from artificial intelligence. And it has a bunch of behavioral science, I guess, drivers or motivators behind it that I
guess we could get into later. But you know, the obvious sort of high level aim of it is to just really try to create the kind of knowledge, infrastructure and shared understanding that makes it really easy to understand, like what is being done, what should be done and so on. And then you can sort of back chain to the kind of interventions and things that you might do. I think we're most excited probably about people in government like policy makers, decision makers, frontier model
developers. But it's of use hopefully to almost everybody. The problem that we're trying to solve is the lack of shared understanding, the lack of transparency. So at one level, that's like a lack of understanding of transparency around, like I said, who is doing what in relation to risks from AI. At another level, the level of the AI risk repository, which was the first piece of this project, it's trying to like understand well, just what are these risks?
You know, who has published what on these risks? What are the specific things that they have said? Can we try to capture and converge existing knowledge to create something that's like a sort of a shared framework that people can build on that, you know, everybody can kind of scale up our understanding on top of rather than this fragmented ecosystem of different people publishing
different things. And it's very hard for somebody who's coming into this area or who's already in the area to understand what specifically has been done and how does it all fit together. Yeah, yeah, yeah. I think that's great. And I think from my end, I think it's really serving a valuable purpose in terms of, again, I think everyone is aware that there are inevitable AI risks
and there are many. But I think it can also be overwhelming and confusing when you're talking about it without really getting into details of what actually that can be. And I think what you are doing with AI risk repository, it's trying to create a way where it's kind of deriving from a lot of public and published data in terms of I think it's more than 1000 sources now. We reviewed about for the first version of it, 17,000 documents we extracted.
We found 43 sort of frameworks of taxonomies that were related to sort of broad frost coding risk from artificial intelligence. We extracted about I think it was 777 mentions of risks from those. We categorized it then into two taxonomies, 1 based on sort of the cause was a human or AI driven, is it pre or post deployment, pre or post deployment And then another sort of taxonomy which is around breaking it into domains.
So we have 7 domains and 23 subdomains, for example, things like is it really to misinformation? Is it related to AI system safety failures and limitations, sociedomic, environmental harms. So yeah, trying to sort of really make all of that information more accessible, easier to understand. There's this idea that like knowledge is like a semantic tree.
So, you know, if you look at some fields like the international, like medicine has the international classification, but seized in economics, they have like onet, which is kind of a breakdown of all the different tasks that are done. And because these things exist, you know, because they're really kind of structured and rigid, we're able to do a lot of analysis and work and comparisons and what would you say aggregate a lot of research
that fits together really well. But then in the absence of something like that, if everybody's using different language, you can't really do the kind of analysis, develop the kind of understanding that you want. And over time, if it continues that way, you know, there's no sort of easy way for somebody to come along and compare the findings in one place with the findings in another place
because they're kind of mapped. They're not on the same with a semantic branch of the tree or something like that. So you can't easily like fit them all together and sort of understand the the synthesis. So is the goal then to synthesize all of these 63 different frameworks into one framework? Or how are you going about simplifying this for our teeny tiny human brains that like, like, I feel like the big version of this, the like massive database, that's something that's like, you know,
perfectly suited for AI systems. You know, any sort of machine learning system can ingest that and understand it and do something with it. But for humans, I feel like I could make maybe comprehend 3 frameworks, like 4 frameworks. But really I'd like to have just one risk framework. How do we combat the overwhelm? Yeah, I think that's a really good point. I think again, part of what we're trying to do here is make the whole thing less
overwhelming. So I mentioned how, you know, we've taken these at the moment after 1000 risks as of the last update and we've mapped them to our two different taxonomy. So what that means is, you know, you can look at a high level, you can see that well, you can break down risks in terms of like is it AI or human driven? Is it pre deployment or post deployment? Is it malicious use or is it unintentional? You can also look and see, you know, is it related to discrimination, privacy,
misinformation. So you can kind of engage with it at the level that is of interest to you, whether it's a very high level or slightly more detailed level. And then you can get to, I suppose the risks or the specific risk domains that are of particular interest to you much more quickly. Because I imagine, and this is true of most of us, you know, we want to have an understanding of
things at a high level. Like we, you know, we want to have an understanding of, I don't know, maybe all of the different behavior change approaches or something like that, or, you know, yeah, one of the well known taxonomies like Mindscape or something like that. But really most of the time we're only going to engage in detail with a small piece of those.
And if we're going to be experts, we'll often be experts in a narrow area depending on whether consultants or academics, I suppose, how narrow that is. So it's intended to make that whole process much easier to sort of understand the high level and then dive deeply into the lower level and see what the risks are. And then later in the longer term, having this taxonomy, you know, we're able to map things like experts to specific domains. We're able to map mitigation
specific domains. We already have an incident tracker that maps it. So we're able to use this semantic tree, this knowledge infrastructure to pull together all of these other streams of information and make them much more understandable and easier to build on and build out over time. Peter, are you familiar with this meme like everything is fine meme? Yes, the dog in the house, Yeah. Exactly like the dog in the burning building. Even I know that one. Yeah, I feel like everyone knows.
But do you feel like that? Do you feel, you know, you're dealing with categorizing like plus risks? Does that ever make you feel a bit like the dog? You, you come a bit like the dog in some ways and that like you're just in the house and it's on fire a lot and it hasn't collapsed yet. But, you know, you sort of like, you initially have the fear response and then after a while you become like a bit habituated to it because you're just so engaged with it. It's very weird.
It's happened to me as well with, I suppose the COVID-19 pandemic as well. We're at the start I felt, you know, really intense and really, like, disturbed and sad and, and, and maybe even excessively motivated to like, do things to try and use behavioral science for good. But then after a while, it was like, I just, you know, I guess working 12 hours a day for so long, you're just like, you know, you become, yeah, desensitized to some extent.
Wow. Yeah. Is it because you've accepted any of the conclusions that you've come to, like, yeah, this AI is going to be the end of us? Or like, where do you stand on that? We're doomed, Spectrum. Yeah. I mean, I can tell you that I'm not a sophisticated forecaster on these things, but I was asking myself, what's my probability that AI will like link to the death of like a billion people within the next like 50 years?
You know, so they're kind of like instrumental to complete extinction or something like that. And I think it's about like one in 50 to one in 200 is probably the range. OK, hang on. I have 0 interest in your academic calculation of the risk. I want to know how thinking about depressing things all day long affects your well-being. So I feel like you said you become kind of desensitized.
It's very demoralizing because I'm really only a very small player in any of the sort of ecosystems and certainly like in the global ecosystem of actors who can change the future of the world. So I don't know if I find it depressing. Like, I try not to just dwell on the fact that things might be getting worse and just, you know, like focus on the things you can control.
And then I try, as Sam will know, to do a lot of things, you know, meditation and I have a wide range of like self-care routines that I have built up over the years that I feel are fairly robust and those things keep me taking over.
And then it's very important as well to like, not constantly be dwelling on it and, you know, get out and do things on the weekends and in the evenings so that, you know, you're not constantly like wrestling with difficult questions about things that aren't necessarily going to go well. And you're a human too, so you engage in all kinds of motivated
reasoning, right? Like if it's inconvenient for you that the world will end due to AI or whatever, you probably have an easier time believing that. I think so, although I will. I think that I am less of a motivated reasoner. I like to think that I have a more unblinker take on reality. Yeah. But I guess like on this topic, I guess in your team, the people you work with, there's a obviously like a mix of people, but obviously you come from this very behavioral science
psychology perspective. And how does that impact how you think about the risk repository in terms of, we know that there's certain risks that are more likely to be maybe valid and real and potential, but maybe less sexy. And then there's some form of more quote UN quote sexy risks.
I don't know how to describe it, but like the robotic dog that you see from Black Mirror or various things that seems very dystopian in terms of what are maybe cultural artifacts have given us some form of people talking about Skynet very often because they've seen the Terminator. People have some images of like how things could go wrong, but it's often times maybe not exactly associated to the highest risk or the biggest risk in reality. How do you think about that? Yeah.
I mean, I think these are things that I have deferred thinking about a little bit because for me the big rally realization was we're in such a pre paradigmatic field where people don't really know what the broad categories of risks are. We don't know how to quantify those risks. We're only just starting to understand the incident rates
for those risks. So like a lot of the stuff we're doing is trying to understand what we should be trying to do and then back chain to like, well, how can we like try and make sure that kind of information is communicated effectively. But yeah, if I was to answer, I would say, yeah, I think it is a really big issue of just definitely public perceptions or any sort of level of perception around these risks at the moment is driven by all sorts of factors that are not the ideal factors.
But then the issue is we don't really know what we should be trying to move people towards. You have some idea, right? You can make some sophisticated guesses.
So just to get a little bit more concrete, if you were to take your high level categorization of different risks and you were to say these risks, people seem to be over fearing these more than they should and these other risks that people are under concerned about, what are some of the specific conclusions that you would feel comfortable making and you can leave out the
ones or you have no idea. I think like the ones that I may be most concerned about or a lot of them are just, I think people are underestimating how quickly AI will develop more dangerous capabilities, you know, so like future tech, the lab that I'm in and these numbers might not be exactly right, but I think our analysis suggests that there was a study of like improvement in
video GPU's. So like hardware improvement, there was a 67% improvement each year, whereas I think with algorithms, algorithms it was 150% on average, it's improved each year. Like Can you believe that? So more improvement has come from algorithmic improvement.
And then yeah, we have like so you know, agents at the moment, I was just reading some research today from meter that the duration of tasks that they can do are human tasks they can do with more than 50% reliability, double S every seven months or has doubled every seven months over the last 10 years. So I feel like we're in a world where people are kind of putting this in the reference class of like the Internet or something, but actually it's like a much more fast moving technology.
So they're like, Oh yeah, you know, like this sort of thing this like Chachi between these character AI, like the job substitution. Yeah, these could be issues like let's have a report in two years and we'll like, think about it and, you know, then we'll like talk amongst ourselves. But it could just move much faster than that. The other thing that I say from a psychology perspective, it's
really bad. Like we need to learn that it's not about the frequency of the event, it's about the expected value of the event. So like low probability, high magnitude things. We don't have to all be like, Oh yeah, like AI could take over. Like that's really likely. We can say, yeah, it's like, it seems unlikely to me, but it will be the worst thing or extremely bad if it happens. So we need to assign, you know, appropriately to manage it.
OK. Yeah, let's go with that because this is very much aligned with my intuitions. Tell me more about the expected value of different risks, different risk categories. Right. I mean, so one of the things why I'm maybe hesitant to go into too much detail on that is because one of the things we're trying to do here is to have a unbiased and fair representation of what other people think the risks are.
There's been a lot of work where, you know, people are like, well, I'm only concerned about this risk. So I'm going to work on those. I'm going to talk about those. So, you know, if you get into a kind of a thing of saying, well, these risks are higher priority than others, then maybe you're not doing a fair job of representing everybody's
concerns. So I kind of want to outsource that to like experts who will do those sort of evaluations for us. But if I was to, I mean, it seems obvious that the ones that are the lowest probability, like the ones like AI takeover type risks are AI sentients. These are ones that I suppose, yeah, they're the lowest probability. But then if they happened there, they could be like extremely bad.
So those are like high unexpected value from a sort of high by negative impacts with low probability. And then there's a lot of others that are happening now that I suppose, you know, disinformation and cyber attacks, fraud, these sort of things like they're not going to destroy the world, but they're definitely happening and they're going to be happening a lot and they're going to cause a lot of
problems. So as a behavioral scientist, I'm sure that you think about when people are more or less accepting of risk and when they are more eager to take on more risk. Those sorts of situations, whether it's the landscape of competition, it's, you know, how they're doing, if they're sort of ahead of the game, whether it seems like everybody's doing it, whether there's a social norm for taking on risk when there's a lot of uncertainty around the area. These are all contributors to
higher risk acceptance. And when I think about all of the things that would sort of lead to having a higher risk tolerance or acceptance, I think, wow, that seems a lot like the current landscape of AI that we're in right now. You know, it just like driving the world and lots of competition and it seems like that everybody is doing it. Does it seem like this is the case to you?
And how do you see the sort of psychological factors of risk tolerance playing into the behaviors that are being exhibited or that are likely to be exhibited by the people who are developing AI technologies? Yeah, that's a great question. I was just thinking that as you were asking it, like there's a
kind of a status quo slippage. There's this good book called Uncontrollable by Darren McKee, which is just an introduction to sort of AI. And he has this great quote, which is like something to the effect of the impossible becomes improbable, the improbable becomes probable, the probable becomes normal, the normal becomes mundane or something
like that. And it really feels like kind of, you know, so several years ago, like 10 years ago, when I first would have found out about risk from artificial intelligence, people were talking about how, oh, like, we wouldn't ever connect, you know, an advanced AI to the Internet. Like, people wouldn't do something stupid like that. But then they did it almost immediately. Yeah. And yeah. So I feel like we have this thing going on where, you know,
it's a set of problems. One problem is there are these competitive dynamics. There's really salient benefits from increasing capabilities amongst the actors who are in charge of that. And then the sort of like, it's a bit like COVID where it's very hard to convince people to take precautions because there's a 1% chance of the pandemic if they fly more often or they like don't do these things in their airports that cost them a lot of money or something like that.
But with all of these risks, it's very hard for people to envision what it will be like until it happens. And it's very easy to envision that they could sell more and make more money if they develop them faster. And then there's two different audiences as well. Like the people who are actually making decisions aren't really subject to the risks as much a lot of the time. Yeah. What do you mean?
I mean, the model developers, the people who are building these things are the funders who are, like, putting money in to try and make money from them. They, for example, are, I think, less likely to, for example, experience like job loss. Or they are less likely, if things go really badly, to be sort of subject to the worst outcomes of AI because they are the wealthier the more privileged.
So they need to be exhibiting some extreme altruism in order to be making these pro social decisions about risk is that. Well, I guess, yeah, there's a part of it. I mean, the key issue is the coordination problem like that there's a lot of different actors. And maybe if all of them believed that, you know, if all of the large model developers believed that the other large model developers were going to slow down similar to governments, if they believed that other governments would
part slow down, then they would. But none of them believes this. So they're all in some sort of paradoxical race to safely but rapidly develop you know, the AI before someone else does it unsafely or more rapidly. I'm sure you remember there was the really half hearted, perhaps half assed attempt at this, the call to pause AI development. For a year, right? That happened early in the ChatGPT days that like nobody, nobody did that. Well, there's a thing there
with, like, a lot of this work. It's widening the Overton window. It makes people more aware. And then there's this idea that Cialdini talked about, like, about persuasion. Like at some point in the future, there may be some particular moment in time when something happens. And then, you know, in the sort of social dynamics of the world, it just so happens that we cross some thresholds where people are like, well, I remember, you know, people were taking this really seriously.
And if they're willing to take it that seriously, then I guess that normalizes it enough for me. Maybe you have that to bring in too many theories here. Sunstein had this idea of like cascading social change. You have like the zeros will speak out without anyone else, the ones, the twos, the 10's, the hundreds.
So maybe they make it more likely by giving the social proof, providing this like behavioral modeling or whatever, that in the future, you know, more people will speak out and then it becomes a little bit more likely that there will be
some sweeping change. So I think we're waiting for events like, you know, if there is some bad outcome, if there is some Chernobyl moment from AI, like people then will say, Oh yeah, remember people were talking about pause AI, Remember, I like stopping it. I sort of feel maybe more legitimate in taking a stronger stance. And it will maybe make it more likely that something will happen. But I agree it didn't have any
obvious effect. The pause is unlikely and maybe not even the best solution like what is a good solution? So one good solution is the AI Risk Index project, which you may have heard of, to try and bring shared understanding about the risks and enable people to coordinate. But I think another like I am excited.
So Yoshua Bengio and Future of Life Institute and next Tech market and others have talked about kind of, you know, narrow tool based AI. So if we can from a regulatory and an economic perspective, I mean, we don't maybe need to do anything from an economic perspective, but I think economics will push us towards, I hope anyway, narrower, more specialized uses of AI that are more cost efficient and also safer and regulation maybe as well.
But if we have something like that, you know, if we have AI that is only trained on a very small relevant set of data and it isn't given access to many other tools and AP is and the Internet, or if it is, it's like very heavily monitored. I think that those uses of AI are a lot safer than if we have, let's say, you know, a government trying to develop an all-encompassing AI that's going to run society in the military and, you know, do all of these different things.
I think, yeah. Shaping the trajectory, making people more concerned about risks, shaping the markets as well, making it more attractive to build a narrower things, more kind of risky, harder to ensure to build a more expensive things is 1 potential solution I'm optimistic about. It seems like the market is running in the opposite direction, right? Everything I hear is about AGI powerful AI, not narrow AI. Yeah, yeah.
So one of the papers from our lab, it was called Beyond AI Exposure and I can't remember the rest of the title, but it's easy to find. But basically they were trying to model not just whether a particular task was exposed to AI in terms of like vision models, so whether you could use, you know, AI vision models to do some of the task. They were also trying to stand like the economic sort of incentive for substituting the task.
And they identified a lot of tasks were exposed to AI, but it was very cost inefficient for most of them to like build an AI system or use an AI system to do them. So I feel like we are going to hopefully in my sense of hopefully like run into the reality that there are immediate gains to building these very large systems.
But then with open source and the competition, you know, you can't really, when you build a model like it's only really like top tier for a few months and then you need to spend a lot of money to train it again. So hopefully, like they will start to realize maybe that a lot of what's going on now with investment is driven more by hype, things like actual
tangible, real world uses. And they will hone in more on, OK, we really, you know, could make a lot of money from just going to come up with something like AI in cars or AI or, you know, in some medical setting or in research. And then they will just hone in on like doing things that are extremely cost efficient with extremely tailored and useful data in those areas. And these bigger models, yeah, they will have a bunch of uses, I'm sure.
But there won't be maybe the same amount of drive to get towards super intelligence. That's maybe motivated reasoning. But that's what I'm hoping will happen and we will see. There are various moves by powerful actors to damages markets at the moment, which also give me some hope. Yeah, I guess. Well, it's interesting in terms of what you described is focusing on the verticals specifically.
And I think currently in some ways that's being very much talked about in the various AI incubators that a lot of AI startups are kind of focusing on verticals like being the best AI tool for lawyers or for various things. But it seems like these bigger labs, they have a little bit of difference between what they're stated and revealed preferences
are. So if you take something like Sam Altman, he was definitely someone who was like involved quite heavily in various ways within talking about AI risk early on, Obviously also then driving a lot of AI developments. And then kind of this being a little bit of a, a drift towards more and more doing, doing, doing and less and less maybe taking precautions and so on.
And whenever there's been kind of signs of things slowing down a little bit, there's been pressures again, like when the deep sea came out, you could kind of like see the open AI pushed out the various things very quickly. And then you have the likes of Entropic, which a organization that has, it's always formed from people that have left open AI and done things in a more ethical way. You still have the leaders they're talking about the kind of there's a quote run Daryama De.
I think it's very thoughtful person, but he said something to the effect of, well, the thing with the AGI is that will impact us all equally. And going back to what we talked about before, like that's not going to be because it's going to be kind of unevenly impacting us in various ways. And also when it comes to legislation, it seems that even today we have very little legislation around AI. But it seems like the pushes is to let's do less.
Like you, you see recently this murmurs around kind of that the Trump administration has been pushed to kind of really remove anything that would even when it comes to training data, when it comes to copyrighted material, Like let's make that even easier for AI to train on because it's important for AI development in the US to be better than anywhere in the world.
So with that being said, if you had to come look into the crystal ball of the next coming years, like what are you expecting to see in the near term developments of AI? Yeah. So I mean, again, I'm well trying to be like a librarian of sorts here and bring together all the books from all the experts rather than kind of present as an expert. But yeah, in the context where I'm at, I think we had a debate on this.
I think that we will see we won't see very sweeping changes in AI for the next 10 years, like a societal impact perspective. I think that because of reasons like I just suspect that people like us, certainly me and you some I don't know as much about you lean and how much you engage with AI, but we are really like in the pick of it, you know, looking at everything, trying everything.
I mean, when I think of my parents, when I think of, you know, people doing farming in Ireland, you know, hanging out in Sydney, Australia or other places that I've been or even just around Boston. I don't imagine them all switching over and adopting the latest AI technologies. And then their organizations are complicated with a lot of, you know, complex decision making processes and laws.
And we still have like, fax machines and hospitals as a common, you know, thing that gets thrown out that I also fall back on. So I suspect that what will probably happen is a little bit likethe.com boom, the sort of fight will shown to be kind of overdone. And then there will be kind of a bit of a pullback. But then there will be, you know, over the next 1020, thirty years, huge sweeping changes. I'm really concerned for
children in particular. Like, I really don't know what the world looks like in 30 years time. I don't know what what I would say to parents of young kids now. So that's really hard. Those are some takes. I mean, any other takes. Yeah. I think that something bad will happen soon. And I'm hopeful that it will be a bit like COVID in the sense of it'll be bad, but it won't be as bad as it could have been and it will make everybody aware.
I feel like it's much easier for me now to talk to people about things that are like, low probability time magnitude, like negative effects because of COVID. You know, I could say, well, did you predict the pandemic? You know, I can point them to where I read things and posted about those things. And everybody was like, this is ridiculous, Stop spreading misinformation. And then, yeah, it happened. And everybody realized, oh, like, this is actually, you know, a thing that can happen
every now and then. Yeah. So hopefully, yeah, people will start to think more about it as well. I'm curious, in line with that idea about low probability, high impact events, there's this quote that I really find very interesting from Andrew Ng. He said worrying about evil AI killer robots today is a little bit like worrying about overpopulation on Mars.
How do you think about that sentiment, given that your entire existence at the moment is preparing people to think about the risk, the various risks of AI? Yeah. So I mean, I think that if I was to quote him and say like worrying about future pandemics is as pointless as worrying about unethical behavior in spaceships or something like that, that would look like a really stupid take if you had given it.
You know, anytime, I guess, in the 1980s, nineteen 90s, early 2000s, I think like the best time to prepare for things is long in advance of when they happen. And that the best sort of approach to risk management is like a portfolio approach where you weigh your response
according to the risks. And it's a bit dismissive of him as well because, you know, what really got me concerned about risk from AI was that all of these quite legitimate people, like the most cited computer scientists in the world were saying, well, actually, we really don't know what could happen here. And they were willing to like change their careers and do a bunch of things.
So I'm suspect of a lot of these sort of AI gurus who happen to make a lot of money from AI, happen to have a career working on AI saying like, oh, you know, we don't really need to worry about these things. That to me just feels like it could very well be like motivated reasoning or deliberate kind of misleading points. In terms of like you reference that we debated a bit around this topic and I do think we are very similar in terms of how we
see things. And I think the thing that I find it hardest to make sense of is this unknown unknowns. I noticed how good AI is at so much kind of tasks that we thought it couldn't do just a few years ago in terms of knowledge tasks and beyond and how things are evolving on robotics and in very spheres around AI.
And it's it's very hard to know, you know, in terms of the growth that you were citing before as well, you know, what are unexpected outcomes that could come up in the coming years that it's just, oh wow, this really like it became really good at doing this thing and that had really big effects on the financial markets or on something else and so on.
And that is just very hard to think about it and really make sense of. Are you ready for a quick fire round that we call to AI or not to AI? OK, all right. How does it go? It goes like this. We're going to present you with some various tasks, and basically these are hypothetical. Some are probably more likely than others to happen soon, but it could be something that can happen at some point. And what we're instating is when it comes to AI, whether that's a good or bad use of AI.
So maybe less about if they could AI could do it, the more should AI do these things. OK. OK. Go for it. So first to AI or not AI, design an AI powered hammock. AI do it. I feel it's not very sophisticated use of AI, but I would take it. I'd be amazed. Low risk.
Yeah. Is there anything from that that you would like really request like is there some because I ask you because I know that you like a good hammock and so is there something in terms of that could be really valuable to have AI kind of understand and kind of decide for? So I've often spoken about the benefit of the hammock over other forms of sleep and rest devices. And I think, yeah, like a lot of it is around temperature control, the sort of swaddling of the body, the kind of rocking
of the hammock. So something that could, you know, temperature control could rock. You could maybe like, yeah, kind of massage you a little bit, cuddle you a bit. Yeah. I mean, there's a lot of, I think things that I could see. Yeah. I could learn kind of what you like and then optimize your sleep. Yeah. So now I feel it's gone from like a question you've asked me to the product I'm thinking about. When can I make this survive? It you can market it as the Snoo for adults.
Yeah, sounds good. Yeah, OK. Talking about potential products to AI or not to AIA Autonomous agent that manages all of your tracking spreadsheets. Yeah, it's another good question. I feel like I'm strange thing to reveal. I'm maybe more risk tolerant in this area than I should be and I'm very efficiency driven. So it does appeal to me to have that agent kind of track a bunch of these things for me. I'm a bit wary. We'll say 2 AI. We'll go with that for now. Sounds reasonable all right.
Have the option for every AI voice agent to have a strong Irish accent. To AIA, 100% one of the best accents in the world, I think, according to the science that I've read, which is mainly in Irish newspapers and probably. And we're not talking motivated reasoning at. All all those critically acclaimed journals. Yeah, maybe this should be the default setting. Latest latest findings from Dublin University. Irish action is the best accent of all accents.
OK, what about this? To AI or not to AI have a give? Well, alternative that looks only for the most effective ways to fund AI exploration. Kind of like AKA Rocco's Basilisk kind of thing. So not to AII would say, yeah, that would be pretty grim. And yeah, like that, unfortunately is is something that probably will happen
relatively soon, I think. I'll be meaning to ask you that actually there's a quick side question because Ruckus Basilisk is this idea that at some point it will be very advanced AI and this very advanced AI will be somewhat almost like a God in that context. And so it'll be beneficial potential for some to today do whatever that would appease that future God. So that would be to like accelerate AI today to be in the good graces of this future
basilisk in the future. Have you actually seen this sentiment like play out? Have you actually noticed that? Because it's something that is discussed in some like rationalist circles, but have you seen people being swayed about this? In rationalist circles, I think
people are persuaded by this. People who like really anthropomorphize advanced AI and yeah, I mean, I say that without a lot of confidence, but I get the sense that I have seen comments enough to convince me that some people do think that this is a like people talk about it being like an information hazard. Like I said, you don't want to know about it. It'll make you feel bad is one example to think about DNA that eventually AI got will go after us. I think I've seen that my memory
now. I want to double check that for you. You take me seriously. These are pretty fringe beliefs, Sam. Are you saying this is something you prescribe to? Well, no, but it's funny because every sci-fi and fantasy story, like if you look at most of them, if they have a bad guy, there's always like a kind of a a relatively insignificant character that is their used to bring that bad guy to life or
give that bad guy power. So it's like it's only serving to like get Voldemort to get have power or to get Sauer on the ring. Like all of these small side characters are acting a little bit like this. And so that's kind of interesting. Like are we going to have these humans that are in some ways seeing their purpose to to. Yeah. Yes, I think we are probably going to have whole religions around AI, you know, people can believe in.
I think with Scientology, you know, he said the best way to make money is a religion. Like, no, it's so if people can believe that, you know, if you have a conversation with some of these advanced models, they are pretty like knowledgeable and insightful. And you can imagine a world where they're like hooked up to a lot of devices and they can do a lot of other things. And they also know how to role play as a God. They do a great job, dude, maybe a better job than God does. God.
Actually responds to you. Yeah. Pretty low bar. Yeah. OK. Moving on to AI or not to AI? Live AI risk dashboards at national, organizational and individual levels. Definitely, maybe not at individual levels. I mean, I don't know if, you know, my, my parents need to be coming into the kitchen and seeing like, oh, there's been more incidents of like, you know, 6.2 on the AI risk taxonomy.
But yeah, I think that widespread awareness of what is probably the most significant technology that humanity has ever developed is going to be very important. And keeping it salient, even though most humans don't really want to focus on these sort of things, is going to be very important. Yeah, I'm taking it too seriously, but yes. That's great. Another version that in some ways is basically to AI or not to AIAI that specifically rates your altruism on a database.
Yes, if done right to AI. If done right, I mean, I'd worry with that, that it would, there'll be some sort of reactance or backfire effect. And, you know, a bunch of people will be really like, pissed off with this AI trying to tell them what to do. Yeah, it could lead to burnout. It could, yeah. So I don't know. I'm attentive to AI with with strings attached. I think tentative name for that project is. You could do more. Yes. You could have more, Peter. You did a good job today, but
you could do more. Well, you can see how it could do, you know, here are all the good things you did today. You're such a good person. So you have like the labeling you have like the reinforcement, you have like, making it really salient that you're being a good person rather than what a lot of people feel, which is just, oh, I could have done more. Like maybe they already provide that to themselves. That's true. Good point. I'd take that version. No criticism only compliments
all. Right. LinkedIn anti pontificator AI, so this warns users not about if they risk spreading misinformation, but if they use overly self-serving or smug writing. Oh yeah, yes, I'd like it. I'd like it. I'm trying to think through the 2nd order and all the other effects. I suppose it might make pontificators just better, You know, they might just become better pontificators. So, you know, like they're like, oh, I want to humble brag more efficiently.
So is that a better world in which there's the same amount of humble bragging but it's more palatable? I don't know. Yeah. That's great. OK, a fantasy Premier League AI banterbot basically only exists within the game to troll players on their decisions and transfers. Yes, yes, although I have found it hard to. Maintain my high performance in fantasy Premier League over the
years. As you may have noticed, I was once slightly far and away the best player in the leagues I was in. Occasionally anyway. Yes, for the wider public, yes, let's do it for me. I don't know if I have time to engage with an AI banter bot at the moment. OK, Minority report style AI crime prediction algorithm. Well, well, well, well, yeah. So I'm going to say maybe I feel that is ethically, yeah, that's a hard one. That's a really hard 1.
And it is going to be a thing we're going to have to figure out that maybe we're already avoiding wrestling with now is like, how much should we use these AI tools for surveillance, for crime prevention, sort of trading off privacy against the impacts on the society. So I don't know.
Fair enough. OK final one AVR world that is identical to reality so maps exactly onto your life and your reality but without any of the stressful elements would like it basically reduces away all the political drama, the wars, everything is scrubbed away so you can interacting with things even there's the Internet but you're not exposed to any of the bad stuff there. So you basically put this on the layer to reality that scrubs away the stress of it. Yeah, I think I'm God. How many?
There's a lot of 2nd order effects here. I'm going to say my intuition without having thought about it much is that if it's truly for everybody in this kind of ridiculous hypothetical, I guess, you know, everybody can go into it and they're going to be sustained a bit like matrix light, but they're going to have a good time.
I dislike to make experience machine thought experiment, which I think is something along the lines of you could be hooked up to this machine and just experience like bliss. Would you do it rather than like being in the real world with all of these challenges? So I think like if it's for everyone, it's going to reduce a lot of suffering. It's probably going to be net negative and everyone includes like animals and sentient creatures that otherwise suffer.
But if it's only going to be for a selective number of people, then they're just kind of opting out. And probably I'm still in favor of it if it's. Yeah, wow. So there you go. Wow. I. Feel like that could be your most controversial opinion? That is the next question, right? Yeah, it is. Peter, what's your most controversial opinion about AI? Well, I think in the community that I'm in, probably I'm less concerned about risks than some
people are. Maybe one opinion is I really am more maybe positive about the idea of like AI as a companion. I've recently been enjoying having more conversations with advanced voice mode and there's a Sesame, which I think I shared with you, Sam. Yeah, I feel there's a lot of benefits to talking with AI over talking with humans. You know, I feel I'm really aware that it doesn't matter if I interrupted or if I ramble. I don't know.
I'm not causing. Like there's all this interesting research actually that found if, you know, for example, people are playing chess with an AI, they don't experience as much discomfort when they lose. Like in some way their mind kind of understands I didn't lose. Maybe it's an evolutionary thing or whatever, but they don't get this like I have lost status in my group and you know, it's really bad and maybe I'm going to get rejected.
Or all the things maybe that are theorized to lead to like the strong emotional response when you lead to lose to a human. So the similar dynamics when you're talking with the AI where you just you don't feel the same pressures about things. You can ask you to explain things and definitely it's always available. It will get to know you better
and it will be better informed. Like one of my friends was telling me they used AI to try to analyze text messages that exchanged with somebody, you know, that they were dating to sort of understand like whether that person liked them or not, and also to predict whether that person would send them a message at some point in the future. And it was very helpful for both of those things. And apparently, it got both the timing of the message and the topic of the message right.
So this is maybe a piece of the AI world, although there's some risks there about unsafe use and over reliance and things like that where maybe I'm a little bit more optimistic. And I think some people will be like, no, you know, I'm pro human, humans all the way, human relationships only. But I think I have room in my heart for an AI as well. You know, you mentioned Sesame as one of the recent demos
around voice remotes. And I think I found myself starting, you know, having a short chat that I was just going to test its capabilities. And then 30 minutes later, I was still talking. And I found it interesting to like, I found myself getting and just talking about more normal stuff. And it's from a behavioral standpoint, very understandable that of course, it'd be nice to have someone that's infinitely patience and always speaking to
at your level. And also can have like this kind of very paradoxical nature where it can speak to you in a very kind of nice normal way, but have the knowledge of like the most smart and intelligent, you know, collection of PhD level people that like a lot of people would feel very, I would say, intellectually afraid to speak to because they were like, how can I even dare to ask questions to someone like Jeffrey Hinton about AI or someone. Yeah.
That's a really great point. I mean, I do, I'm sure a lot of people feel this, like I feel very frequently uncomfortable if I'm talking to somebody who I know is very smart and very successful. But that's often the way that you get a lot of the best insights. So, you know, you kind of push through that and then, you know, you ask questions of those people. But yeah, you don't feel at all the same sort of pressure here if you're talking to an AI, which would be even more
knowledgeable. And then the other thing which you didn't mention, which I know we both have talked about before, is like a key thing to think about here is that the versions that we're trying now are the worst versions we'll ever use. So, you know, you think, oh, that was pretty good. And you're like in five years or three years and two years, it's hard to know the time frame if this is much better and really knowledgeable about me and my
life. And you know, it can sort of talk about like what I did yesterday and they can say tomorrow, you know, you've got this presentation and I noticed that like you haven't said or done anything about it today. And like, is there anything I can help you with there? It's going to be very hard for other types of relationships to match that in terms of the cost, reward, sort of benefit or whatever. Yeah, I know for sure.
And I guess to wrap up, I can say that it's hard to think about many of the questions that we've covered today, but I think an easy answer for me is that I will predict that in many years in the future, I'll still find it as valuable to speak with you and listen to you as today and as in the past. So I'm super happy that I was able to convince you to finally come on the podcast. And I think it was a really fun
conversation today. So yeah, thank you for coming on the podcast, but also honestly, thank you for all of the great work. And I feel like I've been very lucky to learn from you and know you've a long time. Thank you, Sam. That's very wholesome. But no, genuinely, that is lovely and I reciprocate your appreciation. And yeah, it's been really valuable and I'm very excited now to have done this, this first podcast and overcome that fear, I suppose, that I had.
And that's a wrap. You've been listening to the Behavioral Design podcast brought to you by Habit Weekly and Nuance Behavior. Sam and Aline tell me This season is packed with incredible insights about behavioral design and AI, so be sure to subscribe and share the podcast with your friends. Though you might want to keep it away from your enemies. In case you haven't noticed, I'm an AI voice. Yep, pretty crazy. Quite the improvement since last season's AI outro, don't you
think? If you'd like to collaborate with us at Nuance Behavior, where we use behavioral design to craft digital products with Nuance, e-mail us at hello@nuancebehavior.com or book a call directly on our website, nuancebehavior.com. A special thanks to the amazing Dave Pizarro for our show music, and to Mei Chen Yap and April English for their help in producing and publishing this episode. Thanks again for tuning in.
We'll be back soon with another exciting conversation where behavioral design and AI intersect. Happens. To. Mugatroid. Oh. That is kind of likened to evolution of interspecies, where if you have an Otter and something like a polar bear, if one of them becomes better at hiding or hunting, the other one kind of selectively becomes. Better at the other thing. Do polar bears eat otters? They eat the seals. They eat the seals. Seals, seals, seals.