Transcript
0:10 Good afternoon, everyone. I'm Chen. I'm giving a talk on a particular sweat of my lab's research focusing on designing AI for social good, and I will jump straight in. this line of work was initially motivated by so a lot of concerns we hear in popular media about how generative AI is going to impact our society, not just in individuals or specific user groups or communities, but impacting our society at large.
0:44 >> >> How many good jobs are there going to be? How How many people are talking to chatbots instead of people? how many people are having sort of inflated self-understanding because they're talking to sycophantic chatbots every day? Those influences, although quite small if you look at each individual user and each individual interaction, but over time and over sort of in population scale, these small impacts actually add up pretty quickly.
1:17 and as we've learned from social media's impact on users' mental well-being, we know those like small, slow-burning, incremental influences cannot be neglected. They are actually quite impactful. and that's why in this line of work, my students and I are mainly looking at these impacts that are not just impacting particular vulnerable groups or user populations, but really kind of how does how does generative AI impact society at large in this kind of pervasive way.
1:54 but then we also realized an inherent challenge in studying this kind of impact is this kind of impact is really hard to measure unless that impact has been entrenched deeply and widely in society, just like social media's impacts on user mental well-being. but we have to measure it. We have to try to take actions on it because again, if we learn anything from social media and mental well-being, if we wait until we can fully measure and understand those technologies impacts on well-being and society at large, by the time we can measure it, attribute it, it's probably too too late. The change has been so deeply entrenched that we probably have missed a real a really important opportunities to to course correct. So, we set out to try to work on these sort of really elusive problems from the beginning and our goal is to say, can we start acting on those societal impact of generative AI even though we don't fully understand them.
3:02 Can we start working on them, taking action, be proactive before we fully understand those impacts? And that's the overarching goal of this sort of set of work I'll be presenting today. here is the same idea represented in a sort of a graph form and I'll actually point back to this same graph repeatedly throughout this talk. we take as a starting point like we suspect there is some sort of societal harm that generative AI is causing, be it mental well-being, loneliness, social connection, be it impacts on people's creativity, job quality, and so on and so forth. Let's say we suspect there is some sort of harm that's happening but we cannot quite put our fingers on it. we cannot quite measure it, but can we go from that suspicion and figure out some way to arrive at an action that we can take today? That's the goal of this line of research.
4:08 There are of course a lot of researchers already started looking at this. and here is my brutally simple simplified summary of what they currently do. they look at they they have a suspicion that there's some sort of societal harm that's happening, so they would then build an alternative generative AI application that's free of that harm. So, for example, if I suspect Claude is automating people's thinking, people are not being as creative or as cognitively engaged as they are, I build an application that helps people think, pushes people's cognitive engagement, improves their creativity. And then, people also, I think scholars and designers of these systems also recognize, well, Claude and ChatGPT are probably not going to adopt my design cuz they need the the user engagement and the income that the the money that comes with it. So, they're probably not going to adopt my system proactively, and users are probably also kind of lazy. Maybe they wouldn't want you use my system either. So, in order to achieve to actually solve this societal wide harm, I should actually ask policy makers to adopt my system.
5:29 wouldn't it be nice if public policy, if lawmakers require everyone to use my better designed AI system so that no one is harmed by those commercial chatbots. It's in that even it's in a sense we're seeing in the past few years, a radically, increase in, the amount of papers that mention policy, in in HCI conferences. More and more people, after they propose a system, in the discussion they say, "I call for policy makers to to do something about it, to to to endorse my system or to adopt some of the system designs."
6:15 unfortunately, those calls are very rarely, acted upon. Not many policy makers, that is to say, very rarely, policy makers actually acted upon, those calls. so, we keep seeing more and more researchers calling policy makers calling for policy makers to adopt their design, but policy makers are not doing anything about it. We're not seeing organizations or, large companies like Cloud or or Open AI adopting the, Open AI or Microsoft to adopt these designs. So, we're kind of facing this we're we're kind of in this deadlock. We're like, "Okay, do we do I just keep building? I can't build more systems that help with people's creativity and mental well-being. Should I keep building?"
7:03 I think a lot of researchers are probably still doing that and I'm one of the most also. >> >> we're on the same boat here. but there are also some designers, some researchers, including designers in the industry and say, "Hey, why do I keep designing these systems, when I know it won't make a difference in the real world? When I know the vast majority of the users are going to still use the most available, commercial, LM chatbots out there."
7:35 and we're seeing that, in some of the empirical work in HCI where people studied designers in the industry and they noticed there's this just increasing sense of hopelessness and identity crisis in the industry where people say, "Well, I recognize how complicated this problem is in order to reverse some of these to to to understand and reverse some of these generative AI societal impact or AI's impact society impact in general, I don't quite know what to do. There is money involved, politics involved, policy involved. What do I do?"
8:13 I I particularly like the this paper's title from a few years from Kai a few years ago. Although not talking about generative AI directly, but I feel like it summarizes that sentiment that I hear from society-minded designers very often. The The This paper's title reads, "My sense of morality leads to my suffering, battling, and arguing." Like I want to do something good for society, but I couldn't don't quite know how to do that. And that sense of not quite knowing what to do, the current dis- dissatisfaction with the our current approach to address generative AI's unfolding societal impact is the kind of a problem that my work my group's work tries to tries to address. So, the central question of this talk is really how do we break this deadlock? How can we come out of this impasse?
9:09 and I ask this question particularly from the perspective of a technology designer, as someone who cares about social good, who wants to build systems. Still, I understand there's a lot of value in turning myself into an activist or start start my activism group or start to lo- be a lobbyist maybe to influence policy, but I never do less want to hold on to my expertise as a system designer, as someone who's skills central skills are in bridging human thinking and system and making AI things.
9:48 So, when we ask this question, we very much focus on like what's the right thing to build there? I know there's value in activism and policy work, in outreach. are there things we can do as designers, as a people build AI things? it's in that context, I'll talk about a few pieces of work. The first part of this talk is entirely empirical work, where I ask this somewhat philosophical question of what is the right thing to build? If my goal is to be able to make a difference, be able to move the needle in terms of addressing generative AI societal impact, what should I even do before I even build anything? so, I'll present two empirical projects that looks into this question. And then afterwards, I will share a few snapshots of the design work we're currently doing to show you how that new understanding has changed what we design and how we design AI systems.
10:54 I'm going to give you the takeaway right away. throughout these projects, I want you the takeaway is this. I think we should stop Well, stop we should build fewer harm-free generative AI applications. instead, we should build what I call harm measurement and harm mitigation systems. So, systems that instead of trying to build a generative AI application that doesn't cause mental health harms or doesn't harm people's creativity, we actually should spend more of our energy on measuring that harm at a societal scale.
11:34 And on top of that, when we build these harm measurement and mitigation systems, we should also start to think about how those systems can seed changes in public opinion, seed changes in policy, and seed new business models. now, for those in the audience who perhaps come from kind of more system-building background, they look at this and say, "Well, I public opinion, like that's not what I do for my work, for my research." Like business model, policy impact, these seem like broader impacts. Like I'll figure those out after I build my system and publish my paper. I hope through this talk I can convince you that no, in fact, if you want to make a difference in addressing these gen AI societal impacts that's unfolding today, we should actually not only consider these factors early on in our system-building process, but treat them as some of our most important design goals and success criterias. Cuz unless we can seed larger changes in policy, pol- pol- politics, and business models, we can't our individual systems will face a hard time reaching kind of societal scale impact. So, it's in essence for the for folks in the audience who are perhaps looking for more of an STS or business or ethics talks, unfortunately, this talk is not really designed for you.
13:08 >> >> despite the fact that this talk will involve a lot of the business and societal kind of policy questions, the the I very much designed this talk to orient towards to towards people who build systems. All right. let's jump right in. the first project I'm going to present it's kind of a a simple one that just we just want to learn from history. Like, if the thing if the problem we're running running into now is we build all these amazing systems that the that don't cause any societal harm actually can make the world a better place.
13:48 and but it was why why can't we influence policy? So, can we learn from anything like in the history? Like, what kind of technologies actually changed policy, actually got adopted widely, got enforced at the societal scale? That's the question we set out to answer in this project. as with any project we started a lot of reading, literature review. I named a few examples here to show you the range of readings we did. We read a lot we read a lot of science and technology studies which are which is a field in the social in the social sciences that try to retrospectively analyze how technologies changed the society. So, we looked into those scholarships to understand what technologies did influence policy and influence society.
14:41 during that literature review process we realized that actually HCI did a lot of great work in influencing policy like privacy, like accessibility, social media, content moderation. but however, the HCI researchers don't always publish those like, how did you change policy? What did you do? we don't we actually don't know that. So, in 2023 at CHI we organized a workshop that specifically focused on asking everyone to share their story, their engagement with policy world. And both success stories and you know, frustrations and failures. we got 80 plus participants, researchers from across five different continents. Heard a lot of amazing stories and their stories informed the findings of this project greatly.
15:35 we learned all these lessons. We know, we now know what happened in the history. next we looked into do these historical lesson lessons apply to generative AI? Like, can we actually use those lessons to inform what we do today to address generative AI's still unfolding societal impacts? so, last year we organized a thought summit which was like a small week-long workshop at an Innisbrook, New York where Cornell is. we invited experts in mental health, mental well-being, tech policy, health policy, LLMs and probably oh, VCs, investors, and startup founders. So, all kind of academics, so all kinds of people interested in this question of LLM chatbots impact on populational mental well-being.
16:27 I would gather them together and sort of start to look into how how these lessons may or may not apply to generative AI. In addition to learning from experts, we also did a lot of hands-on exploration trying to apply the lessons we learned from past technology that did influence policy. Trying to apply those lessons to generative AI's various impacts, ourselves. not all of them led to publications as you could imagine, but we learned a lot in the process sort of delving our into various areas like creativity, mental well-being, learning, and so on.
17:08 so, that's the sort of somewhat messy and well-lensing research process. and I want to share where our learning sort of have brought us so far. before I talk through the findings, just recall that this is the diagram that I pointed to you in the beginning, the pathway that many technology designers today envision for as the pathway to address generative AI's societal impact. So, if they're concerned about the potential harm, they take two steps. First, I design a less problematic application or more beneficial application, and then I call for policy changes to enforce that design.
17:46 our findings suggest that, however, historically, yes, these two steps did help, but they actually play a much smaller role in the much more much longer and much more complex set of sort of pathways. first of all, let's look at the harmful like harm-free system design. if you recognize, for example, a teenager is addicted to social media, our instinct might be, let me give the let me give them a social media platform that's less addictive. However, if you look at those history and also how service currently treats these addictive symptoms, they would tell you the first thing you want to do is to actually make sure that person, that teenager, has the ability to live and thrive without that pleasurable but harmful technology in the first place. And we see similar evidence from LLM chatbots as well. there are studies show that in some countries that where where AI companions were banned, after those bans, people just people still didn't talk to their friends. People just were even lonelier. so, the first thing we need to do before we take away those harmful technologies is actually give people the ability to be able to live without it. so, it's in a sense the technologies that don't don't cause harm plays only a part of it is only part of an equation.
19:29 now, let's look at the policy change side. I would say policy change as much as I champion those workshops and called for design policy collaboration. I actually our studies actually suggests policy change also plays a smaller role than we initially thought. And the reason being in a democracy, at least in theory, policy actually cannot be too out of step with public opinion. So, unless you have a public that recognizes, "Hey, we need to use less of those chatbots cuz they harm my creativity, they harm my well-being." Unless you have that public awareness and civic understanding, policy would it's unlikely that the policy would charge ahead and and propose some saying propose some radical change. It's in that sense that I put the policy change block in on the side here rather than in the center cuz I think these two, public opinion and policy change, really come hand in hand.
20:39 let me punctuate the value of public opinion even more. In a way, if we have a public, if we have citizens who really all recognize the value of limiting their use of chatbots to for the benefits of their mental well-being, for the benefits of their learning, for the benefits of their creativity, they could have just adopted those better designs on their own. They don't even need a to demand them to require them to do that. So, public opinion or civic understanding plays a huge role in addressing societal harms in that context.
21:22 That's the bottom of the diagram. Something that I discovered in this research and and people I think I think don't talk about this enough is the importance of humble technologies that measure harm. let me give you a concrete example. let's say smartphone overuse. In fact, HCI researchers were some of the earliest researchers that recognized, "Hey, people are using their smartphones too much." Way ahead of any political momentum, way ahead of this popular discourse around smartphone overuse.
22:04 And when those researchers recognized this potential harm, the first thing they did is to start to run studies at trying to measure smartphone overuse. This might sound like a simple measurement, but it's actually not as straightforward as it might seem. How much use of smartphone counts as overuse? Like, do you wait until someone goes to a therapist to count as overuse? That's it would be hard to to to ever reach that bar. So, the researchers actually spend a lot of time sort of studying what's the best way to measure that, and they they came up with a various measurements like do people feel the time they spent on smartphones are time well spent? Turns out proxies like this are much more useful than absolute numbers people of minutes people spend on the phone.
22:53 Things like that as the simple measurements like that can help in a number of ways. First of all, these measurements can first help people who recognize a need to self-improve, to to to limit their smartphone use to start taking actions even without government intervention or causal evidence at a population scale. >> >> this evidence over time, by the time the right politician came into play came into power, by the time there was enough political will what when there was enough public awareness of this harm of smartphone over use, these researchers have already accumulated a a bunch of evidence like population-level evidence to show like, "Hey, here is our evidence why this is indeed bad." And that evidence helps sort of catalyze the policy changes that people have been hoping for all along. A little bit delayed, but it's sort of really good cuz you need actually to start accumulating those evidence early.
23:57 It takes time, but if you start collecting those evidence when the when the right politician is in place like by the time you finish collecting that evidence, that politician's probably gone. lastly, it also changes public opinion. So, it's evidence like this, publications like this, science communications like this that sort of catalyze public opinion change over time. And again, that has sort of an interactive effect with policy change as well.
24:28 So, in hindsight, oh, and we see similar patterns not just in smartphone over use, but in a lot of other technological history as well like civic monitoring gadgets. they're they played an important role in catalyzing certain air pollution intervention policy interventions. So, it's in essence in hindsight, this these technically quite simple measurement systems actually were able to make these otherwise slippery elusive societal harms can measurable trackable actionable. They played a huge role in addressing those societal changes that are otherwise difficult to act on.
25:16 So, that's the picture and I want you to draw to attention to the time frame of this whole process. there are really two time frame two time scales at play here. One is designing technologies for measurement. You can run your usual HCI research study a few months maybe. you can get that done. you can design prototype and evaluate your harm mitigation system like oh I'm going to ask people to reflect on their smartphone over use. Those are relatively quick. However, to accumulate enough of these evidence these measurements and evidence into something that can create public opinion change at large and wait until the right political moment the political will to accumulate those take time.
26:12 They can be really slow but I don't I don't think those are reasons that we don't start acting on those harms today because again when the political momentum comes if you miss it you miss it for a long long time. It would not come again anytime soon. So, it's in that sense in my our findings would suggest that this process this whole workflow actually should happen in two stages. You first build these technologies and you actively waiting for the right moment to come to to create a societal impact and that's why in the title of this talk I say we design technologies to seed policy and political changes. Those seeds do not blossom immediately, but they will pay off at some point.
27:02 so that's finding number one. This is sort of a nice playbook in terms of how historically technologies HCI researchers have impacted policy and made a made a difference in in society. next question, does does this playbook apply to GenAI, apply to GenAI's impact on mental well-being, creativity, and so on? Mostly, with one important caveat. if we need generative AI to detect the harm, then this detection itself, this measurement itself takes money. And this can be like if I want to run log analysis and analyze how much people's conversation with LM chatbots impact their mental well-being. It can also include things like computing how much social media impacted or accelerated misinformation spread. Because these types of harms are in natural language and rich media, you kind of need AI models to do detection.
28:04 You cannot just use your off offline GitHub repository to to to run those detections. And it's in essence you need AI to run those measurements, to run those detections. However, because you want to reach population scale, recall that social media's impact on mental well-being, ChatGPT's impact on mental well-being are really small if you only look at a small population. It really relies on large-scale tracking to show the aggregated effect. So that money that cost adds up quite a bit.
28:41 why does that matter? it matters in this particular part of the workflow, and let's take a closer look to that. let's say I want to monitor cloud and track GPT same pattern people's mental mental health and I know there are certain costs attached to running that measurement. Cloud and open and third party and open air probably are not going to pay for it. it would be somewhat strange to ask third party to pay for it. So, most likely you're going to ask for government funding, public funding to fund that basically public mental health monitoring program. So, you need some sort of policy change or public budget change to fund this work. but oops, but recall that in a democracy, you need political public opinion to enable that kind of policy change.
29:40 but also recall that you need your tool to show that there is harm such that you can change public opinion. But then I need the money to run the to run the measurement to change the public opinion. So, that's where this this seemingly simple question of budget actually creates a deadlock in the process of starting the monitoring of technology's impact on society. Which perhaps also explains why historically we have been just so slow, so reactive in acting on technology societal impact. Cuz there is we haven't found a way to come out of this deadlock. So, the only way is to wait and wait until there is public outrage towards this harm that's already entrenched. And now finally there is political will to allocate budget for the monitoring.
30:39 >> Yeah. Australia This work is funded by the Australian >> into into like any or policy because I think that there's probably a long period where people maybe like disagree and then like there needs to be some kind of consensus, right? So, like we try to like a narrative like how does that kind of look? >> I don't think there is a uniform answer. That's a great question. There is I don't think there is a uniform answer.
31:10 partially because policy public policy isn't entirely based on science. in some areas they are. In health care for example, rough at large there is more of a culture of believing in evidence, you know, randomized trials, but in tech policy, consumer tech policy for example, they're not that evidence-based. So, consensus plays a role but not always. And in essence I would think consensus is relatively relatively loosely related to this process.
31:44 >> So, can I can I actually ask for more follow-up questions? Maybe like a bit of a spicier question but like >> >> how do you square that? But like a lot of this is about messaging, right? And then a lot of it's like needs like kind of push the public opinion like in a certain way such that like good is ultimately done through policy. but sometimes like the most accurate messaging is not necessarily the messaging that will actually achieve the most change. Like you might want something that's mildly inaccurate or like very that's very clear to the public. Like how do you square this like for yourself?
32:19 >> I'm on the next slide. I'll come back to that. >> >> And are they related or I can wait till the end. >> Yeah, let's wait until the end. Sorry. okay. where are we? Oh, there's this notch. we need money to get things going, especially in triggering the public monitoring that we need such that we can trigger the range of actions that we need for societal change. So, it's in essence I modify this in the context of AI monitoring of societal harms. I would put money as So, the first trigger, the first step.
32:59 Now, to the audience question if I were allowed to make this diagram even more complex I would draw additional arrows from that money icon to policy change. That's lobbying. I would draw additional arrows from money to public opinion change. That's political campaign. So, money The key point here is money is actually sitting pretty top pretty upstream in this whole workflow. It plays an incredibly important role.
33:31 Hopefully that addresses the question, too. Okay. So, it's in essence after this review of history, review of HDR research this is where we're at. Turns out we're pretty good at designing. We already know how to design technologies that would be able to trigger societal actions against pending harms or pending impacts of new technology. We just need to make the money work. Then, that's the question.
34:03 Our second project looks into And in particular we turn our eyes to entrepreneurs and intrapreneurs. Entrepreneurs are people who build a technology system into a product to make sure someone buys it, pays for it. Intrapreneurs are people who do that within an existing company. We looked at their workflow because for people who are actually interested in who actually want to start a company for social good, they actually also face the same conundrum we face, which is to make someone pay for the good deeds that benefits everyone rather than just themselves. In that context, we looked into particularly a group of entrepreneurs and investors who call themselves I think startup for humans, entrepreneurs for humans. a lot of them work in digital health, so they're building technologies that benefit for people's health and in that way benefiting society.
35:05 and they work in this particular system that's called the US healthcare system. meaning many good deeds are not paid for. They there are many technologies that would benefit society but there isn't, according to the people we interviewed, there isn't a known business model to make it financially viable. This includes technologies for public health and if you consider chatbot LM chatbots' impact on mental well-being as part of public mental health, this this would that problem would fall into this category because US health insurance tend to be relatively short-term, employer-based.
35:43 there isn't a great business model for technologies that benefit people's long-term well-being. there isn't a lot of money to be made by making everything easier, but you can make a lot of money by every by making everything harder and making the patient going through more jumping through more hoops. things like that. so we actually found this group of startup founders, investors who are actively trying to solve these problems. We're trying to build companies out of this these types of technologies, trying to make someone trying to make the money work for technology for social good. and here's our finding, which is the the the the headline of this talk, they really saw their design process as planting a seed. They don't expect immediate returns per se, but they see themselves as planting seeds for new policy, new politics, and new business models.
36:43 Here's what I mean. when these people try to design make their their tech digital health products work financially, they every single one of them really understands the business landscape, the regulatory landscape, and the political landscape working in that space. They know exactly how often they will get paid by Medicare or Medicaid if they provide certain kinds of health benefits. and it's based on that knowledge, they can then design their technologies to in that are financially viable, that can survive on their own.
37:27 a lot of the participants we interviewed they call them they they say, "Well, you're like an architect. most of the time you follow the building the zoning laws. Only once in a long, long time you build a building trying to change the zoning law, trying to change the entire neighborhood. Most of the time people follow the rule books, and those rule books are the existing business models like selling selling to Medicare or Medicaid and things like that.
38:00 but once upon a time you do intentionally try to create a new product that tries to break that pattern, tries to find out a new way to make some social good product financially viable. here's one example one participant talks through. They're trying to cre- find a find find a buyer for their digital apps that already has some evidence of effectiveness in improving people's mental well-being.
38:33 >> >> They first I know the arrows are a little bit messy, but that's the messiness is very much to the point. number one, they first try to sell it to hospitals. Well, but the hospitals needs randomized trial evidence, which is really expensive to run. So, they thought, "Okay, what if what if I just put this app on the on the app store so that I can get some users and I can get some data to start to show some level of efficacy."
39:00 so, they did and they did collect some data to show some level of efficacy of their of their digital health intervention. but that's not enough to it's not strong enough an evidence to sell to insurers or to hospitals. So, they went to try to sell to employers, the tech companies that try to provide mental well-being benefits benefit packages to their employers. So, they try to sell this product to those employers and those employers say, "Well, I don't really care about this person's long-term mental well-being, but I care about how many sick days did that person take off. Like if you can show this digital intervention made my employers happier, work more productively, take fewer sick days off, I would buy that." So, these startup these funders so tweaked their design still trying to improve mental long-term mental well-being of the users, but also try to simultaneously optimize for those goals that the buyers want.
39:58 And they did do that and they did manage to sell the product to some employers. However, it's still not enough money. Doesn't quite balance the doesn't quite make the the money work entirely. So, next thing they started looking into insurers. Kind of what kind of premium pri private insurance plans can can potentially pay for that. So, this process goes on and on for several years and still were happening when we were interviewing them. By the time we were interviewing them, they were also exploring up the potential of building a private insurance company themselves while also so that they can fund the technology develop themselves. So, they're really exploring novel ways of making the money work. Make a technology that provides social good to survive and even thrive in a otherwise pretty capitalistic market.
40:56 >> >> here's the interesting part. Every person we interviewed, every person in that community knows what are the experimentations that are going on. All of them say, "Well, that's probably going to fail." but that's the that's the game you're playing. You intentionally You're intentionally trying to break the mold of what this this kind of market is allowing for. Therefore, it's worthwhile to do it these experimentations even if most likely the technologies are going to fail.
41:30 here oops. Here is one participant summarizing this idea which I think really powerful summary. They say, "Well, yeah, most of them are going to fail cuz you're intentionally trying to build a technology that's not following the capitalistic kind of market rules, but if just one of these companies works, then every future company can follow the playbook. Then forever on, everyone would know how to build a company that provides, you know, public mental health that can provide social good." everyone else can follow that rule book. You just need one success stories.
42:11 I found this like such a powerful message. You think a lot of times we don't don't hear that hopeful message as much anymore, but very powerful here. so, to summarize where we are at, we started this project thinking, well, conventional wisdom is if I just build a technology that doesn't cause harm and like ask and then as long as policy makers adopt them, enforce them, problem solved. Turns out things are a lot messier. a lot more complicated. but, I think there are still a lot of positive sort of takeaways from this. for people who are like me who build systems, who care about societal good, look at how central technologies how central a role technologies play in this whole workflow.
43:02 Technologies have great potential to benefit society in that sense. I also think HCI researchers, people like us, are particularly well positioned to sort of solve this puzzle. for example, we're really good at being suspicious, noticing noticing potential societal harms before there are, you know, large-scale causal evidence. We are also quite good at juggling different stakeholders' needs and wants and everyone wants a slightly different thing. in this context, in order to raise money to build a technology that benefits not just the people who paid for it, there is a lot of sort of need negotiation and service design that can that that needs to happen here and our expertise can be really valuable.
43:53 that being said, I also think in order to make a difference and move the needle, we do need to work differently from how we work today. I I I wouldn't mind if Kai has fewer papers where someone prompted element and make people more more creative. in a way we know how to do that. We also know the bottleneck is not how to design a technology that makes people feel better and more creative and learn more. The challenges lie outside of the technology on it itself. It's in essence I think there's great need for us as technology designers to embrace these perhaps less familiar topics like policy, like public opinion, and like business models.
44:42 there are not every evidence can prompt policy change, right? a lot of times if you think of like cigarettes kind of regulation, a lot of times evidence of how people use technology that harm themselves ends up being kind of become a story of individual responsibility that the user is irresponsible rather than here's something, here's a collective harm. Everyone in society hurts is hurt by it. Let's do something. So there are a lot of sort of skills and challenges in designing technology that I can actually seed policy change and political opinion change.
45:20 >> >> that's the diagram. That's the where we are in terms of here's a road map I think we that might get us closer to moving the needle in changing improving gen AI's societal impact. I'll I'll use the next few minutes flip through some of our ongoing projects real real quick. just to give you a sense of what how this larger line of research changed to what we do in the first place, not exactly how we executed each system.
45:54 Happy to talk more if folks are interested after the talk. so in our own project we specifically looked at chatbots impact on populational well-being, particularly how it changes people's intention to people's ability to understand who you are. There are studies There are studies showing that even talking with a sick friend dick chatbot for a short moment inflates people's self-assessment, think they're better than their friends, therefore less likely to reach out to their friends.
46:30 We also know from controlled experiments that chatbot use directly is directly associated with less social intentions. So, these are all things that would cause some concern about people's long-term mental well-being. So, the first thing we did was to build systems that measure harm. In fact, I mean, we're going to show I'm going to show you a super complicated diagram, but it's actually technically pretty simple. in very much in the same spirit of monitoring people's smartphone use, like we try we actually borrowed proxy measures from cognitive psychology to try to measure those human AI conversations impact on users' mental well-being.
47:12 Diagrams look like this. And This is one participant's 2-year-long worth of conversation with ChatGPT. And when we look closer into it, we can actually see some patterns through which people start some patterns that are recurring across different conversations, and we can show that, "Hey, these are some concerning interaction patterns. Those conversations are likely to lead to less resilience against future stressors and difficulties in life." And we can see that across multiple time points of the same participant. And then we can run the same analysis on a wide range of cognitive psychological measures just to monitor which measure might go wrong.
48:01 Which measure might be worth more studying. So, that's what we're doing trying to a measure harms and also potentially inform people take actions protecting themselves and also potentially in the future informing policy. We also did a number of studies trying to improve people's ability to live out live without chatbots. Turns out people are pretty bad at talking to other people about who they are, about their identity crisis. And turns out people are pretty bad at offering other people help in terms of other people's identity crisis.
48:37 usually things go one of the two ways. The people talk about their identity crisis in a really sort of indirect way, vague way. The other person will just utterly confused. Or they got they had this conversation so directly that it turned into a fight because the topic was so sensitive. So, our studies really show people usually seek help from friends first and then they were frustrated by that process so much that they turned to chatbots. So, a lot of the work we're doing now are just trying to increase people's ability to talk to each other before we even try to take chatbots away.
49:16 We also did some policy related work in the interest of time I'm going to jump a little bit. >> Thank you. >> Oh. We also did some research and thinking what does the exact what does it mean to build a chatbot that doesn't cause mental well-being problems. Turns out it's pretty complicated as well. Like to what extent at what point you've crossed the line from being just benign friendly to harmful sick fancy. It's a pretty hard line to draw. So, we did a lot of policy kind of related legal framework review trying to understand what exactly content is bad.
50:01 I'll jump ahead here. And in this area, turns out making money work is also one of the pain points. There are actually companies that try to build chatbots that are not scientific, actually effective clinically, and try to get FDA approval for it. All of them so far went out of business as far as I am aware. So, another really important area of really important question that I don't think enough academics study at the moment.
50:32 All in law this line of research really at least changed how we ourselves approach designing tech for social good. We sort of took a pause from designing chatbots that make people more creative, mentally more well-to-do, and so our learn better in instead of really trying to understand harm at a societal scale, and then do things that hopefully will seed political changes, policy changes, and business model changes in the future.
51:04 With that, I thank you for your attention and have a good time for for Q&A. >> In an ideal world, what do you think public policy would even look like for AI? As in, how do you think that we could effectively regulate a sector that A is moving so quickly, but B will become very central to geopolitics, and to labor, and to the economy?
51:45 so, in an ideal world, what do you think that would look like? >> I'll give you an answer. It's my answer. Whatever that future is, it's not going to be satisfying. Because policy is supposed to public policy and law are supposed to set the minimum floor of what's allowed. It's likely not going to meet the our standard of what's being ethical is what benefiting society is.
52:20 tackle tech policy or healthcare policy and so on. These laws set the floor and in essence I don't think policy changes up will will, you know, make us entirely happy in essence. >> thank you for the talk. one of my questions was related to in your diagram you have this relationship between the gen AI harm measurements and public opinion. And you've described that as a unidirectional relationship, right?
52:57 but it seems to me like there is a bidirectionality there because often times what we choose to measure or how we choose to measure it has to do with public opinion, right? So for example, in a school environment you're probably going to have a lot of parents who raise concerns in the town hall about overuse of technology in the curriculum. And so maybe opinion is not the right word to use for that, but some kind of latent concern that sits underneath before that becomes what we might call an opinion.
53:32 But that becomes a trigger for what we might choose to measure, right? Or even how we choose to measure it. Like we're going to measure it this phenomena particularly in middle schoolers. so I was wondering how you think about that? And then also, what do we do about harms that we cannot anticipate or harms that we don't have any kind of latent indicator of? so, we're not seeing any indication that this could be a concern, but it is a concern and then might become a concern in the future.
54:07 >> Yeah, that's a great Well, it's a great question. So, for your first question about the bidirectionality of this arrow which is here, I don't think so, actually. I think the Here, I guess though this workflow is more trying to say let's do this. Let's build the Let's make sure we have a way to measure because like we said, a lot of those harms don't have indicators or don't have great proven proxy measures. Even the mental well-being and chatbot when I showed the fancy diagram of, there's no proven metrics.
54:46 They're like we're adapting a lot of the measures from elsewhere to treat this and the group. It takes a lot of resource and energy to validate it. So, here I guess I was mainly trying to make a point about workflow. First, make sure you have a measure, which is your second question. And then, we can start to think about how to actually use that to change your opinion, which comes with its own challenges.
55:19 >> thank you for the road map. I think it's actually interesting because I think we as I guess the research community have the So, basically, I like the thing you pointed out that we start picking up the evidence early on and like different researchers, different teams pick up It's almost like different tiles around the same effect almost, but I feel like as the research community, maybe we're not good enough at like, you know, like some version of like unifying or amplifying that a lot of these things have the same effect. And that then can maybe the question earlier of like influence public opinion better when you can show maybe the scale cuz like maybe someone did a small study different group or something like that. I guess >> That's a great That's a great comment.
56:00 So, a lot of the think tanks in the sort of policy making world, for example, aggregate scientific evidence from different scholars. There are some calls in HCI asking like more HCI researchers to also do, you know, evidence review rather than just keep turning out small-scale studies. So, that and evidence synthesised work synthesised work is is really important. I I totally agree. It it plays a huge role in informing policy and public opinion as well.