# Stanford Webinar - Human-Centered AI: Designing Systems People Trust **Channel:** Stanford Online **Source:** https://www.youtube.com/watch?v=5UJ0uWqO1Ho **Transcript page:** https://www.withtranscript.ai/video/5UJ0uWqO1Ho ## Chapters - 0:10 — Webinar Introduction and Initial Poll - 3:50 — Foundations of Human-Centered AI Framework - 6:46 — Cultural Ontology Research and Audience Poll - 13:15 — Sovereign AI Meaning and Governance Levels - 18:50 — Human-Centered AI in Health and Education - 23:21 — Public Opinion on AI Benefits and Trust - 27:33 — Building Trust, Business Cases, Future Interfaces - 36:00 — Audience Q&A on Collaboration, Ethics, Education and Closing ## Transcript **[0:10] Speaker A:** Welcome everyone to our webinar on Human Centered AI Designing Systems People Trust, presented by the Stanford Online Generative AI Program. My name is Jess Nguyen and I'm a Senior Program Manager within the Stanford Engineering center for Global and Online Education. And I'm excited to introduce our two speakers. We have today Vanessa Parley, who will be our moderator, and professor, James Landay. Just a little bit about Vanessa and James. **[0:40] Speaker A:** Vanessa is the Managing Director of Programs and External Engagement at Stanford's Institute for Human Centered Artificial Intelligence, where she leads initiatives that connect interdisciplinary AI research with industry policy and real world impact. Before Stanford, Vanessa worked in management consulting, applying data science and machine learning across government, biotech and nonprofit organizations. James Landay is a professor of computer science at Stanford University and co founder and co Director of the Stanford Institute for Human Centered Artificial Intelligence. He's also a leading expert in human computer interaction and and he has held faculty and research leadership roles at the University of Washington, UC Berkeley, Cornell Tech, and Intel Labs. He is an ACM Fellow, a member of the ACM SIGCHI Academy, and a recipient of the ACM SIGCHI Lifetime Research Award. **[1:36] Speaker B:** Great. Thank you, Jess, and welcome everyone. Thank you for joining. We're so glad that you're here. Before getting started with James and I discussion, we thought we would ask you all the audience a question. **[1:50] Speaker B:** So, Jess, if you could go to the next slide or maybe the one before that. Yeah. So we are going to use the poll everywhere. For those of you who are familiar, you can either join on the web or through a text message. I'll give you a second for those of you who want to participate in the poll. **[2:10] Speaker B:** So to kick it off, we wanted to ask you all, when you hear Human Centered AI, what do you primarily think it means? A designing AI around user needs B making AI safer and more ethical. C involving humans in the design of AI systems and applications. D developing AI applications for strong social impact in areas such as health and education. Or E. All of the above. **[2:42] Speaker B:** And unfortunately due to the system some laggy, you won't see the results here. But I'll talk through what I am seeing on my screen. So yes, I see results coming in. We're at 57% all of the above. We've got 13%. **[2:58] Speaker B:** A, C and D are also in there. Looks like E is in the lead here, right? Yeah. **[3:16] Speaker B:** Oh, C is starting to get more traction. **[3:25] Speaker B:** All right. Yep. So most of you, about 47%, 50% say E, 33% say C. All right, so now I want to hand it over to James So, James, you are the co director of the Stanford Institute for Human Centered AI. **[3:50] Speaker B:** What does human centered AI mean to you? And how is that different than just plain AI? **[3:57] Speaker C:** Yeah. Thank you, Vanessa, and welcome everyone for joining us here today. And I look forward to your questions as we go along. So, yeah, what is human centered AI? When we founded the institute over seven years ago, we felt doing AI in a different way because of the societal impact this technology was going to have on everything from education to health care to finance to government was really important. **[4:22] Speaker C:** To do an interdisciplinary disciplinary way across all seven schools at Stanford, and in a human centered way was really key. I realized after a couple of years into the institute that that meaning meant different things to some people. To some people meant, hey, just that you were doing AI in an important application area like d there health and education. And to me that really wasn't enough. You couldn't just have good intentions and necessarily make AI work for people and make it human centered. So we really tried to think deeper about what that means and in some ways it's really answer e in this question. It's all of these things and in particular thinking about, well, what is the right way to involve humans in the design of AI systems to get those outcomes around being safer, more ethical, more productive, having a positive impact in those areas like education and health. **[5:12] Speaker C:** And for me, that's to go beyond the traditional, what we call user centered design, something we've been doing in the field for the last 30 to 50 years, where we involve users in our design and go beyond that to think about AI systems because they're different, they often have side effects or impacts on people beyond the direct user. So whether that's deciding whether you get a certain health care treatment versus someone else, you might not be the user of that system, it might be a doctor. Or whether certain safety resources are located near your home, let's say firefighting resources in Los Angeles. In what area, again, you may not be the user or whether it's just people who are labeling training data in another part of the world for a system that they might not use themselves. Again, we need to go beyond this user centered design and think at the next level of what does it mean to have community centered design to involve those different communities that are impacted. **[6:09] Speaker C:** And finally, if the system that you're building for AI is really successful, it can start to have what I call societal leveling impact. So think about systems like the newsfeed in Facebook changing what people think is real or true or not, or Instagram affecting young Women's idea of body image. Those are societal level effects. And so we need to start to think about how do we design with society in mind and how do we integrate across those levels of society. Community, end user. And so that's what human centered AI means to me. **[6:46] Speaker B:** Great, great. Thank you. **[6:50] Speaker B:** So, for the audience, before we move to the next question for James, we wanted to get your perspective. **[6:59] Speaker B:** How confident are you that today's AI reflects your values and culture? **[7:11] Speaker B:** All right, I see some responses coming in. We're at about 50, 50amb. Oh, a C was just logged. Oops. Oh, changing. Click here. B and C are in the lead. Like a race. Announce. **[7:35] Speaker B:** Okay. About 55% say C, 33 say B. More answers are coming. **[7:56] Speaker B:** All right. Yeah. So we're still about 60% of you feel. Feel not very confident that AI systems reflect your community and culture. All right. Yeah. So part of the reason we asked this, James, I knew you have some research on this, and this ties a little bit to the society level of the framework you just described. I find the society level a little bit difficult to conceptualize compared to the other levels. Can you tell us a bit about a research project that you're working on that is at the society level of your human centered framework? **[8:36] Speaker C:** Yeah. So I was just going to mention, what are some of the questions here? So Vanessa and I just came back this weekend from the World Economic Forum in Davos, Switzerland, where you're actually interacting with people from all over the world. And one of the concerns that we heard there, but also in other trips to other parts of the world, whether it's Korea or Singapore or the uae, is people concerned that these large models in some ways encompass cultural values that may not match cultural values from their own society or country or community. And so, you know, this is a question of both, kind of, what are these models made of? They're made of huge amounts of training data that often comes from North America or Western Europe, because there's way more data from those countries on the Internet. **[9:32] Speaker C:** Obviously, some of the models that are coming out of China may have much more data out of the Chinese Internet, but it's a concern about, you know, how is culture and other values embedded in the model. So one of my research projects with my PhD student Nava Haiki, really is looking at, is there almost even an embedded ontology, almost the philosophical word of what does it mean to be or to be human embedded in these models? And we actually find that it is true, although other ontologies from indigenous cultures or other less populous cultures may be embedded in the models if you probe them enough. Kind of the defaults that you get out of these models tend to be from dominant cultures in the West. And so there's questions about how might you train these systems differently, how might you collect data differently, how you might even build the applications on top of them differently. **[10:31] Speaker C:** If we want to kind of explore a broader space of design that this different types of ontologies might even allow different ways of being. So it's almost the assumptions built in almost limit the potential design that. That we bring out in the applications that we might even conceive of built on top of these models. **[10:54] Speaker B:** James, not to put you on the spot, if you don't have the details, that's fine, but when I was reading the work you did with Nava, there was an example about a tree to start off. Can you share that example? **[11:07] Speaker C:** Yeah. That's an example from Nava's own life experience. So she asks her audience, when she gives the tactics talks, to imagine a tree. **[11:15] Speaker C:** Imagine in your head visually what it looks like. And for most people, they might just imagine, you know, a trunk and some branches and some leaves. When I do it, I often imagine those leaves kind of blowing in the breeze because nobody tells me it has to be a static image. But again, it's really based on the ontology of your culture. You know, for Nava, she thinks of a tree as really part of a bigger living system. **[11:45] Speaker C:** So she imagines the roots and how they're connecting to other organisms under the ground. But again, the ontology that, you know, is default for you may change that imagination. So if you typically query a large model, you're going to just see that kind of Western conception of the tree as an individual thing versus this kind of more connected idea or even looking at how it's connected in the root system. Also for Nava, who grew up in Iran, how she might think of a tree from her culture, again, might look different. And in fact, when she asked the agent, give me a tree from my kind of perspective as an Iranian, it actually makes a very stereotypical idea of what is would mean to be Iranian and some of the visual design that it uses. And it doesn't really hit on how she thinks of it from her cultural perspective. **[12:44] Speaker C:** So even something as simple as that seems to have embedding based on kind of in the models, based on the kind of data that they were trained on. **[12:54] Speaker B:** Yeah, yeah, great, thank you. I think that's such a good example because it gives like a tangible, I guess, idea that you wouldn't usually think about when you're just interacting with the systems. **[13:09] Speaker B:** All right, so another poll question for the audience. **[13:15] Speaker B:** Have you heard of the term sovereign AI? A is yes, B is no. **[13:33] Speaker B:** All right, see about 50, 50 here. The answers are still coming in 70 for B, 30 for A. **[14:00] Speaker B:** Seems like we're kind of sticking with that. Seven. Okay. Yeah. 70% say no, and 30% say a. Yeah. So the Stanford Institute for Human Centered AI has a group focused on ensuring policymakers are informed about this technology. And we're starting to hear a lot through that area this term, sovereign AI. James, can you talk a little bit about what that means to you, what you're hearing? It means to different groups and the implications of that? **[14:42] Speaker C:** Yeah, so sovereign AI is really a term that I think almost nobody would have heard about, you know, a little over a year ago. But it was a term that we've been hearing more and more. And in fact, again, it was one of these terms that was coming up a lot at the World Economic Forum in Davos last week. So sovereign AI is really about countries feeling that they need to control AI to meet a variety of different goals. And so what I'd want to talk about is kind of what might those goals be and then what are the different ways that people are thinking about it? And I'd also note that our team at hai, we have a team focused on policy and society, and they're actually in the process right now of doing a report on sovereign AI, trying to really define it better and then map out what is happening with it with different countries around the world, because it's actually being interpreted in many different ways, which makes it both interesting and confusing. **[15:39] Speaker C:** But at the high level, it's about how do we, you know, control AI in some ways. Now, what are the goals? The first goal that some countries have is about national security. You know, how do I make sure that some country doesn't cut off my AI in some way that hurts us from a security perspective or somehow even uses the AI against us? So that's one perspective. **[16:00] Speaker C:** Another goal is about economic security. So it could be both, again, about doing damage to economy by controlling the AI, but also about how do we use AI to be positive for the economy of the country. We're speaking about whether that's just that we're not paying someone else for their AI and we're actually making money on AI based products. Another one is about culture. So similar to what we just discussed in terms of the cultural issues in AI, but it can also mean cultural and language. Does the AI that we're using even properly support our language as well as the underlying culture or even the ontology ontologies that I spoke about? You know, and there's a few other goals maybe about workforce also as well. So those are the goals. Now we're also seeing how do people think about sovereign AI at different levels of the technology stack and how might they control it or make sure that they're secure in that way? **[17:01] Speaker C:** At the bottom level, people are talking about infrastructure. **[17:04] Speaker C:** So that might say, oh, we want to make sure we have enough GPUs or graphic processing units. Those are the key processors that people use for building AI models as well as running AI models, do we make sure that we have them secure in our country, running in data centers in our country and that we control them? So that's the infrastructure level. The next level above that is data. Do we control data from our country and not let it go to other places or make sure we have the right data to build AI models or use with AI models? That's the next level of concern that people think about sovereign AI. And then the level above that, which is, I think one that people have thought more about, is models. So do we use these models from North America like OpenAI's models, GPT5 or Gemini from Google, or even open source models like Llama from Meta, or do we build our own model and control it and have it something that can't be cut off for us? **[18:07] Speaker C:** That's the next level. And then we get to kind of more nebulous levels above that. Applications. Maybe we want to just control what applications are used and built on the models in our country? And then finally talent, can we make sure that talent is being produced that has understanding of how to use this technology and build it all the way to can our society effectively use this technology? So those are the different levels that people are attacking for control for sovereign AI. And again, our early research is finding that different countries might have different sets of those goals that I mentioned first, but also different pieces of that stack that they're focused on. **[18:50] Speaker B:** Great, great, thank you. So diving into your research a little bit more out of your lab, can you talk? So I know you do work at the intersection of AI and health and education. Can you talk a little bit about a project in that area? **[19:07] Speaker C:** Yeah. So you know, my lab, like Vanessa says, we're working on a lot in health and education. So in addition to trying to figure out how to design AI to be human centered, that's one project. Like what would the right design process be that integrates those levels. We're also looking at it in particular settings, whether it's education for young children that is more engaging and uses AI to help drive stories that engage them all the way to what does it mean to have AI in our health care system? And we're working really in health care in two different areas. One is aging in place. So how do we use ambient sensing and AI in a privacy preserving way that would allow older adults to live in their own homes longer with a higher quality of life? That's one large project and a lot of issues there are how do we think about privacy and how do we think about the care workers who come into our elders homes. **[20:03] Speaker C:** Another one that's more personal for folks is how do we use AI to give everyday people what a lot of more wealthy people have, which is a coach. And so we've trained AI models that use the best methodology in coaching, which is called motivational interviewing, as a way to ask a person about their fitness goals, what's worked for them in the past, what's not worked, what's their long term things they're trying to achieve and then together with them make a fitness program that's generated by the AI but that the person has real stake in because they kind of co generated it through that onboarding conversation. **[20:47] Speaker C:** And then we have an application called Bloom that runs on the background of someone's phone. And as somebody engages in activities in the week, a garden of flowers grows to represent how well they've been doing with respect to their fitness activities. But also that AI agent, that AI coach kind of chimes in when they notice you've done certain things. So in fact I had a goal back in March that I was going to do the Tour de Mont Blanc in Europe this past summer. And I was going to walk or hike 120 miles over 10 days and climb 3,000ft up and down every day. **[21:25] Speaker C:** It was a pretty audacious goal for me. So I put that in as my goal. And over the next few months as I used the Bloom system, when I did something like a long hike or a run, it might chime in, hey, you just did a great run today. That's going to help you on Mumpla. Now it wasn't really stereotypical mechanical in a way where it just always repeated that thing, but it would drop in some of these knowledges about what I was trying to achieve at the right moments. That felt like a human coach. And then every week you check in with the coach to See how your week has gone and replan for the following week. In terms of how you might want to change what you're doing, maybe you need to make it harder, maybe you may make it easier, maybe the weather's changed. **[22:07] Speaker C:** You want to do something different and similar during the week. If something's not going right, you could check in like, oh, it's raining today, I don't really want to run. **[22:14] Speaker C:** And instead of you just giving up, the coach might help you think about alternative activities that you might engage in as well. So in a study that we just did and just got accepted for publication last week, we showed that the people in the condition with the coach versus an identical application we built without the coach. The folks with the coach changed their mindset towards their health in a way that's going to lead to, we think in a longer term study, longer term behavior change, which is really what the goal of these systems are. So this is pretty exciting that you could use AI in this way. It won't work out of the box. If you just ask ChatGPT this kind of thing, it will just tell you what to do. And that won't cause most people to actually do it and change their lives. **[22:57] Speaker C:** So you have to actually think about how do we build on top of AI, the best practices that in this case we know for coaching. And so I think this methodology may be useful in nutrition, it could be useful in corporate change, all kinds of things, education, where you need somebody to kind of change behavior. Having a personal coach that has a good methodology like this could be really world changing. **[23:21] Speaker B:** Cool, Great. So you kind of both of those examples you talked about products that are AI enabled. At hei, we release a report every year on the state of AI and one of the items that we highlight is a survey from IPSOS that asks public opinion of AI. One of those questions I wanted to pose to the group before James, you and I talk about it. So for the group, to what extent do you agree with the following statement? **[23:57] Speaker B:** Products and services using AI. Hang on, let me switch my thing here. Okay. Yeah. Products and services using AI have more benefits than drawbacks. **[24:07] Speaker B:** Agree to or disagree. And I'll share what the actual survey came back with. We can compare and discuss. **[24:20] Speaker B:** All right, starting to see some responses. **[24:31] Speaker B:** We're at about 50ish percent agree, about 15% strongly disagree. Oh, disagree is starting to come up now. **[24:44] Speaker B:** B agree is still in the lead, 30% strongly agree. Let's split between strongly agree and agree. **[25:00] Speaker B:** All right. All right, so yeah, we seem to have paused here. So about 5, 40, 45% of you, the majority show B, agree that AI products and services have more benefits than drawbacks. So when we analyze this survey from its source for the AI index, we analyze it by country, and we do find that in more Western countries. North America, Europe, about 40 to 50% of the respondents agree that products and services using AI have more benefits than drawbacks. **[25:41] Speaker B:** Countries in the East, Malaysia, Indonesia, China. Respondents from there agree about 60 to 70% of the time. So it seems like they are more optimistic in those countries than North America and Europe. So this group seems to kind of be more on the Western side of the optimism. **[26:11] Speaker B:** So one last question for all of you. What do you think drives this optimism or distrust in AI? And some of these differences. **[26:26] Speaker B:** A is job displacement, B is privacy and surveillance, C is bias and unfair outcomes, D is lack of transparency, and E is other. **[26:45] Speaker B:** A and B are tied. **[26:49] Speaker B:** More responses are coming. **[26:56] Speaker B:** Let's see. All are about equal right now. **[27:12] Speaker B:** All right, so as of now, the one with the most votes is B, privacy and surveillance at 42%. Job displacement is at about 30%. Those are the two with the most votes. So, James, turning it back to you. **[27:33] Speaker B:** Why do you think there's. **[27:36] Speaker B:** What do you think drives the public distrust in AI and how can we build more trust in these systems? **[27:42] Speaker C:** Yeah, so I think it has to do with, in some ways, cultural as well as economic. So, one, on the economic side, you'll see a lot of the countries where there's more optimism are countries that tend to be per capita income lower. And so a lot of people are seeing the opportunities that AI gives for services that are less prevalent. So, for example, if we had an AI agent that could be 80% as accurate as your doctor, a lot of Western folks would say, well, that's not very good. I want it to be as good as my doctor or better. And in fact, sometimes is better than your doctor, depending on what we're doing. But the perception is, hey, that's not good enough. On the other hand, if you're in a country where there's no doctors per capita or a very low number or low number of mental health practitioners per capita, the idea that I could help out there is very strong. **[28:39] Speaker C:** So sometimes you'll see this difference based on per capita income, but that doesn't really explain it because you also see very wealthy countries or countries that are fairly well developed in Asia and other places that are also more optimistic. For example, South Korea, where the economy has been quite good over the last 30 years. It's gotten much wealthier, yet there's much more optimism or, or Japan, much more optimism relative to, let's say, us, uk, Canada and Australia. And I think part of that is also cultural. So many of you marked B privacy and surveillance as an issue. That's an issue that tends to be much more of concern in Western culture or much more suspicion about the government. **[29:29] Speaker C:** And you find that less to be true in these other cultures. Again. So part of this, I think, is income, but part of it also is cultural with respect to these things where the optimism can be really different. And so it says again that if we want AI to have a positive impact on humanity, which is part of the goal for HAI as an institute, that we need to think about a lot of these different cultural issues and important questions, whether it's job displacement or privacy and valence bias and unfair outcomes or transparency, we need to actually situate them with respect to the cultures that we're talking about, where these, these problems are going to have a different meaning to people in general. They're all important in my opinion, but it's going to be different in different cultures as to how important those things are with respect to them. **[30:18] Speaker C:** And one way to get at a lot of this when we talked about sovereign AI is not necessarily think about a country controlling these things. That's a reaction to, for example, large companies, mainly in North America, like Google or Open Air, others kind of controlling AI. Another one is to think about how do we do open source AI? And I don't mean just open weights, I mean truly open models where everything, including the data is transparent and do it in an international coalition form so that many people are working together on this and therefore you're not controlled by one country. That's something we've been trying to work on. In fact, when we were in Davos, we announced a partnership with ETH Zurich and EPFL and the Swiss AI Institute to kind of think about how we collaborate with Stanford together to do this. And we're hoping to do this with some other folks in other parts of the world. **[31:12] Speaker C:** And I think that's the way forward so that people have trust in this because they see it's been done in an open way and done in a way that is focused on good for society rather than just next quarter's profits for a big company. **[31:27] Speaker B:** Yeah. So, James, speaking of profits, kind of a follow up question to that. **[31:32] Speaker B:** You know, there's lots of opportunity for AI in these different areas, but if people don't trust it and they don't use it we can't capitalize on that opportunity. Do you think this idea of trustworthy AI or responsible AI could be made into a business case so that we have more AI adoption? **[31:54] Speaker C:** Well, I certainly think that companies that think about these issues are trying to think about it as a way of making their products and services more desirable to their clients. So, you know, we have a very successful industry program at Stanford Hai where industry partners join our program. And you know, at some point I was a little skeptical of that at the beginning, thinking, oh, do they want to join with us just to of kind, kind of wash their reputations about being responsible AI. **[32:24] Speaker C:** But in my interactions with these executives and CTOs and CIOs of these companies, they are actually looking to how do we do this responsibly? Because they don't want their reputations ruined by systems that do bad things and hurt their reputation in the market. They actually would like to figure out how to do it in a way that doesn't lead to loss of jobs, doesn't lead to bias, doesn't lead to other harms and, and also helps with people being better workers and enjoying their work and leading to also productivity and efficiency for companies. So they actually have positive goals, I think. But they're looking for leadership from places like Stanford and other universities to tell them how do we do this in the right way? And this is a very new field. Even though a lot of people are uptight about, hey, this isn't working yet. **[33:14] Speaker C:** You know, really many people haven't been working on this for more than three or four years and it's going to take a while to find the best practices that really help these companies do it. Right. **[33:23] Speaker B:** Yeah. **[33:24] Speaker B:** Yeah, Great. Yeah. So again, speaking of, we're still kind of in the early days. We haven't been doing this for too long. What is one prediction for the future when it comes to AI from your perspective? **[33:37] Speaker C:** So I don't want to be all negative about AI. I think AI is going to change in a positive way. Our health care, our education, our finance, maybe even our government if we're lucky. And in fact, things like that coaching application that I talked about, we're only showing it works for fitness, but it can work in many other areas. Like Susan has asked, how about for chronic disease? **[33:58] Speaker C:** Yes, exactly. It can help people manage chronic disease who need further help to do that because it's hard similar for that aging in place. But I want people to also understand from a future perspective, the way we interact with these systems is really going to change so we've been in this model of what we call the graphical user interface. Your mouse, your pointing, your finger, tapping on the screen, on menus and buttons. That's been the interface really since popularized in 1984 on Apple Macintosh, but really invented in the early 70s up here on the hill above us at Xerox Park. **[34:36] Speaker C:** That's been the same interface paradigm for 50 years now. You have people typing into the little box to type to their AI agent, or if they're lucky, they're speaking to it on their computer, on their phone. I think this is actually a small blip in the evolution of the interfaces that we will interact with this technology in five to 10 years, probably more like 10. We're going to be looking at what I call multimodal agentive interface. So you are going to use speech and typing and mouse and pointing and even understanding your body signals, your stress, et cetera, and the context of what's happening around you. **[35:15] Speaker C:** All of that is going to drive the interface and you're going to use whatever devices are at hand for what you're trying to do. So if you're in a classroom, you're not going to be talking to your interface because you're listening to the lecture or the discussion happening. On the other hand, if you're in car, you're doing something else. If you're walking down the street, you're doing something else. If you're in the privacy of your office, like I am right now, I have a big screen over here, I have two screens here, I have two phones, I have a mouse. I might use all of these different things in conjunction and do it very fluidly with an agent. That's helping me. That's the future of interfaces. And we're not there yet. We have this little box that these AI companies think that is the future, it's not the future. **[35:57] Speaker C:** And that's going to change rapidly in the five to ten year period. **[36:00] Speaker B:** Cool, cool. Thank you. Yeah, I find that so interesting to think about, like what is it going to be like and how different it will be. All right, so those are all the questions I have for you. I know we have a lot of questions for the, from the audience, so I thought I would go to those. I will read them out to you and then, and then you can answer as you wish. **[36:26] Speaker B:** So the first one, from your perspective, where do you see the greatest unmet opportunity for long term collaboration between academia, community based education and industry to ensure human centered AI principles are not just researched, but practiced? **[36:45] Speaker C:** Yeah, I mean, that's a great question and I appreciate Helen asking that. I think we're doing some of this already. I mentioned a little bit our industry program. The purpose of that is just to make sure that a lot of these ideas just don't stay in academia and stay in research papers. But actually we're working with companies that want to get this into their processes. We also do this through a lot of executive education that we do at Stanford where we're go out to these companies all over the world and give education or they come here or, or just like this. This Stanford online class is a way of scaling up that way of impacting people in industry who are going to be using these ideas. And so, you know, we are very particular that we're not just teaching AI in this, we're teaching human centered AI, which includes a lot of these lessons about how to design it. **[37:39] Speaker C:** What are the ethics, how to be responsible in addition to the technical content of how the algorithms work and how you might think about it in your own future applications. So those are all important to us. I think the other part is what I hinted at. Again trying to create this international consortium of other players that are all unified by this idea of doing AI that's human centered. That again allows us to scale up our impact around the world rather than just go, okay, we're limited by how many people we have at Stanford and how many people we can bring through our programs. We're joining together with like minded other institutions to again scale it up. And then very active policy team and policy program where we are trying to help policymakers first of all learn about AI so they understand what they're legislating about. **[38:33] Speaker C:** But we also write reports that try to take technical ideas from AI and put it into a form they can understand. And, and that's again trying to change what's happening out there in societies based on what kind of regulation and other things are coming down the line. So those are all different ways that we try to get at having that kind of impact around the world. **[38:55] Speaker B:** Great, thank you. **[38:57] Speaker B:** So the next one, when designing human centered AI systems intended to build trust, how do you think about the tension between transparency and cognitive overload, especially for non expert users? And where should design draw the line? **[39:14] Speaker C:** Yeah, so this is really important. Like we think about trust. You think, oh, I need to understand what's going on with this thing. I need to be transparent and it's not clear. That's what everyone needs. Right? It may need that. It needs to be transparent to somebody who audits the system. **[39:31] Speaker C:** But then maybe you are going to feel good about it and trust it because it has a seal of approval. So imagine there could be organizations that pop up like UL Labs for, you know, appliances that you trust, that it's got that UL stamp of approval versus expecting everyone to be able to interrogate a system and understand that it's trust, trustworthy. So there's trade offs there between yes, that information should somehow be transparent and accessible. The question is, should everyone need to be able to see it? Because that might overwhelm us, which I think is what your question gets at. **[40:07] Speaker C:** On the other hand, some of this is down to just how does the system behave. So for example, if I ask one of these AI agents to do a certain task and it does it wrong, it makes some error that I kind of notice and then I say, oh, you didn't do this right? You got to do it like this. If the system either lies to me and says, I did do it like that, many of us have seen that. It says, yeah, I did do it like that. And you're like, no, you didn't do it, or it doesn't say that, it says, oh, you're right, I didn't do it right. And then now it tries to do it again and then it messes up again, then you start to lose trust in the system. So some of this trust is simply in the user interface of how these systems react when given feedback. **[40:50] Speaker C:** If a person did that I told him to do something, he got it wrong, I would go, okay, maybe I didn't give them the right directions or maybe they made a mistake and then I instruct them how to do it right and they do it wrong. Again, I'm not going to trust that person as an employee or as a friend or collaborator. **[41:05] Speaker C:** It's the same with AI. So how those interfaces react to feedback also leads us to trust, let alone just knowing what's gone into the models and whether it's transparent. Right now we don't have a lot of trust because we don't know what's in those models, right? We don't know what data is in there. If I go to the grocery store and I buy something, it says what the ingredients are. We don't know what the ingredients of these models are. So one way of having more trust is having transparency about the data. But again, these things are based on trillions of bytes of data. None of us are going to be able to inspect it all. But that's where we might require on other organizations that are going to Tell us something about what's in there. **[41:42] Speaker C:** If we have more transparency. **[41:45] Speaker B:** Yeah. So you kind of mentioned this, this idea of hallucinations. Have you seen approaches or what's your perspective how to manage these hallucinations? Reduce negative impact on users, prevent misinformation, especially this, this person asking the question, Especially in the healthcare space. **[42:04] Speaker C:** Yeah. I mean, hallucinations, that is the idea that the model generates content or facts that are not true or, you know, the common one you hear is giving you citations to, let's say, academic work or legal work. That is totally made up. This is a problem that is inherent to the underlying architecture of these models, which is based on what we call deep neural networks. It's probabilistic, and so it doesn't have the right answer always. It's kind of making it up based on probabilities of words or other concepts. Now, can it be reduced? We've seen reductions in this. Some of the ways are one, what people did first, which is, you know, ask a model multiple times and ask it in different ways. The next way was use multiple models to check each other. **[42:52] Speaker C:** That's another way to kind of reduce it. And in fact, the models have started to kind of do that under the hood. They're kind of checking themselves before they tell you stuff to kind of make it go down. But it's an inherent flaw which is probably going to still be there in these models, even with a lot of these fixes because of the underlying architecture. Until we have a different kind of AI, this is part of what's going to happen. Now, people are flawed to. People make mistakes too. So it's not, I think, the end of the world, but it is a problem. And it requires that people using models still have critical thinking to interrogate the answer they're getting and, and really use some judgment to decide whether it's correct or not. And the problems really occur when people just take what comes out of the model as, you know, commandments on stone tablets, that is truth and have no questioning of it. **[43:46] Speaker C:** That's when people are going to get led into trouble. So you have to keep some skepticism like you would anything written on the Internet, because again, a lot of this data is coming from the Internet and a lot of it is not truthful. So you have to actually keep that skepticism and critical thinking if you want to use these models in an effective way. **[44:03] Speaker B:** Yeah, yeah. So speaking of critical thinking, there's a lot of concern about how generative AI is going to impact learning and education, especially critical thinking skills based on Your perspective, how is AI going to impact the system and the next generation? **[44:24] Speaker B:** And what do you think the impact will be for small children versus college versus lifelong learners? **[44:31] Speaker C:** Well, the good news is people are paying attention to this question now at the start of this revolution rather than 10 or 15 years later like we saw for social media where we were like, oh, look what we just did to our kids. As somebody who has kids who went through that, I can tell you a super negative. And we people weren't paying attention. So the good news is people are paying attention here near the start. **[44:56] Speaker C:** The bad news is all of learning and teaching and education is going to have to change. K12 education, college education, life learning, learning, education, all is going to change. Because the whole model is often based on, yeah, it was based on you learning skills and you being able to have critical thinking, but a lot of it's also based on people being able to regurgitate information. And that simply doesn't work anymore when you have this thing that can regurgitate it quite well or write quite well. And so we're going to have to really rethink how we do it now. **[45:31] Speaker C:** I have confidence that in 10 years we will look back and we'll say, wow, we are really teaching kids much better than we were 10 years ago. But the next five years, I think, are in a tumultuous period. As one schools try to keep it out. Look, oh no, we can't have that AI in our school. It's going to ruin everything because people aren't going to think. **[45:54] Speaker C:** Or two people do allow it in, but they're not really changing how they're teaching and learning. And so people are just relying on it. And, and some of these critical thinking skills that you alluded to are declining because of it. We really have to rethink how we do things. And that's going to take time. **[46:11] Speaker C:** Some of it will come from places like Stanford, but some of it will come from anywhere. And people are going to see what works, do science around it to verify works and then replicate it at scale. That's what's going to happen in many educational institutions at every level over the next five years. But it's also an opportunity because the education system tends to be one of these systems that's very hard to change. And AI is almost a Trojan horse. **[46:40] Speaker C:** It is going in the gates and people are realizing we have no choice but, but to bring this in and think about how we're going to reform our education system. And there's lots of educational innovation that's been locked up in schools of education all over the world that that Trojan horse is going to also allow to get in there that are actually even independent of AI. So there's going to be a lot of curricular and other reform over the next five years that is going to happen. That has not happened because the system is, is really hard to change. And so the good news is again, in 10 years you'll look back and be positive. **[47:16] Speaker C:** But if you're a parent with young kids right now, I think the next five years are going to be kind of crazy as you see how it's happening. And it's going to be different, I think in certain public schools versus private schools. From what I've seen, a lot of private schools are way more aware of this change and trying new innovations. And we see some cool schools around the world that are doing fairly innovative stuff like Alpha School in Austin, that I think are the kind of models that we're going to see. Whereas a lot of public schools, again, they're very resistant to change. They're much bigger systems, it's harder. And so they might come along a little more slowly. **[47:54] Speaker B:** Yeah, yeah, that's great. In my opinion, I think education is one of the most exciting areas because the system we know hasn't been working as well as it should for so long. And this is a great opportunity to, to fix it as long as we do it right. **[48:12] Speaker B:** All right, let's see. We have time for maybe one or two more questions. So do you believe that in the future there is a potential for clashes between human centered AI models, applications and the hardcore technical performance, profit oriented AI? And how do we manage this? **[48:32] Speaker C:** I mean, I think there's definitely a concern between trying to do things right and a trade off of those just trying to make money. **[48:40] Speaker C:** But I actually think we're already seeing that those who just go in with an idea of replacing workers simply, hey, I'm using this agent, going to replace the worker. They don't actually seem to get as much return on investment or efficiencies as they were expecting. And because the real returns on this are when you deeply think about what are you trying to achieve, how do we change work processes with AI and the people we have and do it in a good way and work with the workers to kind of design that well, those are the people, I think who are going to unlock more of the profits from this. And again, that was a topic I saw coming up in Davos on many of these panels that I was with ctos and other leaders of these companies, they found, yeah, people have done the first cut, hey, we replaced this one person in this process. **[49:32] Speaker C:** And then they find it doesn't work that well. **[49:35] Speaker C:** And I think it's because people's work is much more complex than economists or other people who try to analyze it in a more simplistic way as just a bag of tasks that you do. And if we automate enough tasks, your job's gone. There's a lot of messy glue between those tasks that are often you communicating with other people, understanding their needs, their emotions, doing them something for their job so they help you. And that kind of social skill is not what AI is good at. And so these things sometimes fail when they've done in this way where they're not thinking about the real role or how the work process should change and how the role should change. **[50:12] Speaker C:** That takes more effort. And I think the ones who do that though are going to unlock much more income, but also happier workforces, all of the positives. And so there is a trade off. But I think the ones who have this longer term thinking and thinking about a rider are going to win in the long run. But it does require more work and we need to see what happens for these trade offs. **[50:36] Speaker B:** Great. So that was more at the organization level. At the individual level, do you have advice on how we prepare ourselves for an AI driven future? **[50:50] Speaker C:** Well, some people I think are worried about AI. They talk about AI and they kind of just push it off because of that fear. I think actually people need to be jumping into it and using it in their work, in their home life. And what you start to learn is you learn what it's good for and what it's not good for and you'll find, oh yeah, this thing, it does it really well. Any of us who have to write a lot have learned, hey, I got to shorten this thing to a shorter paragraph. It helps you summarize something rather quickly. You need to check it to make sure it didn't change the meaning in a way you didn't want or lose something you thought was important, but it's pretty good at that. **[51:27] Speaker C:** On the other hand, there's tasks where you're asking it to decide on something or schedule something, do something that takes some logical reasoning and it messes up. And the key is these models keep evolving and getting better and changing. So you actually have to be doing these kind of tasks in your work, your everyday life, and seeing how it works over time. And that will also help you figure out, hey, how can we do something Bigger in my company or my organization that might be transformational. So you have to kind of jump in and just use it so that you have a better sense of it. **[52:00] Speaker C:** And this whole idea of education, I think is really better and important for this, is having that critical thinking, understanding that these are not God, they make mistakes. And so how do you critically check it, how do you test it, and how do you be skeptical of it while also leaning in and trying to use it, I think are really important. So you want to really use this stuff. I'm surprised at the number of companies where they're saying, oh, we're not ready for it yet, or we ban our employees from using it because we're worried about it leaking information. I think that's a mistake. I think you have to actually be using this in a lot of things you do, and then you start to get a real sense of where it works or where it doesn't work. And that's going to be important for everyone at every stage of our society. **[52:43] Speaker B:** Yeah. Yeah. I'll tell you a story of my favorite personal use case. **[52:47] Speaker B:** My brother got married in May, and I had it write my. My wedding speech. I gave it some, like, memories and told it the tone that I wanted it to write in, and it was great. Everyone loved my speech, so I could not have wrote in it as well. **[53:02] Speaker C:** You had to make sure that it had your tone. **[53:05] Speaker B:** Yes. **[53:06] Speaker C:** Like, oh, you know, some robot wrote your speech. **[53:09] Speaker B:** Yes, exactly. And that it didn't make up stories. The stories were true. **[53:15] Speaker B:** Okay, so last question to close us out. How can the audience stay up to date and learn the latest about AI? Generative AI, human centered AI. Every day there's a new model application. It's difficult to know what's important to pay attention to. **[53:35] Speaker C:** Yeah. So I have three answers to this. One is AI is actually pretty good at creating like a daily or weekly digest of recent interesting things happening in AI that are relevant to your interest. But you could actually query these models and say, hey, I want you to create once a day or once a week, a little capsule of articles or other information about what's going on that's relevant to my interest. And you could give your interest, but it can also learn your interests over time from use AI. So that's one way. I don't do that, but I know people who do. Another way is our institute, Stanford Institute for Human Centered AI. You can go to our webpage, go to the about link, and you can find a link that says sign up for our mailing list, which I know is there because I made sure they added a couple of links last week. **[54:26] Speaker C:** Sign up. **[54:26] Speaker C:** Everything we do, we put out a lot of email about seminars, conferences, all of this content that we put out through that is for free. And you can watch any of our seminars or conferences on YouTube afterwards. We put most of the stuff online, so there's a lot there that will be from leading edge people at Stanford and other people who work with us, and so you can stay up to date. And then finally, education. So courses like this one we're talking about, Stanford Online is something that we will update on a regular basis because AI moves fast. **[54:59] Speaker C:** You know, we're already updating what we did two years ago to be updated because it's moving. And so that's another way to get kind of a crash course and what's going on and understand the issues, which then allows you to have the background to even understand the things that you're reading out there in the media or in academic press or online in a better way because you have the basic context to. To be able. Able to do so. **[55:22] Speaker B:** Great. **[55:23] Speaker B:** All right, well, yeah, thank you everybody for joining. Thank you, James, for taking the time. Jess, anything else to close out? **[55:32] Speaker A:** Yeah, thank you so much. Vanessa, James, very engaging discussions, wonderful audience questions as well. So really appreciate both of your time and thank you everyone for joining us. I will close out the session, but hope everyone has, depending where you are, a good rest of the day, good rest of your evening. Thank you so much.