Over the past two years, generative AI (gen AI) has been a rapidly evolving trend that has touched the lives of many around the globe. Which is why the design of these formidable systems must include experts from diverse backgrounds, says James Landay, a professor of computer science at Stanford University. On this episode of the At the Edge podcast, Landay talks with McKinsey senior partner Lareina Yee about how to develop safe, inclusive, and effective AI.
The following transcript has been edited for clarity and length. For more conversations on cutting-edge technology, follow the series on your preferred podcast platform.
Defining human-centered AI
Lareina Yee: You have been a champion of a human-centered approach to AI development for many years. How do you define human-centered AI?
James Landay: To me, human-centered AI is not just about the applications of AI, which might provide social value, whether in health or education. It’s also about how we create and design those AI systems, who we involve in that development, and how we foster a process that’s more human centered as we create and evaluate AI systems.
Lareina Yee: As the cofounder and codirector of the Stanford Institute for Human-Centered Artificial Intelligence [HAI], you’re uniquely interdisciplinary. Can you tell us about how you bring the community together at Stanford to look at and develop the future of AI?
James Landay: Interdisciplinarity was key to us from the start. It was also key to why we felt Stanford was a special place for doing this kind of work because we have world-class technical folks in AI, computer science, and other engineering disciplines.
But we also have a top medical school, a top law school, a top business school, and top social sciences and humanities departments. And since AI is a society-changing technology that’s going to be everywhere, we feel it needs to include every field—along with the different values and outlooks inherent in those fields—to help shape it.
We give internal grants for research projects, and the sole funding criteria is, “Are you bringing together people from different schools or different departments across the campus?” So we encourage interdisciplinarity by how we fund projects.
We also encourage it by whom we highlight in our communications, as well as in our leadership. Our two original codirectors were Fei-Fei Li, a famous AI computer scientist, and John Etchemendy, a professor of philosophy.
Lareina Yee: Can you tell me how that interdisciplinary approach—combining philosophy, computer science, law, and ethics—gives us a window into how that shapes the questions and the types of research you’re doing at Stanford?
James Landay: For a start, it sometimes causes confusion, because people in different fields speak different languages, so the same words can mean different things to different people. For example, I’m working on a project with an English professor and someone from the medical school. And what they call a pilot study is not what I would call a pilot study.
So you’ll experience confusion, but sometimes that confusion leads to new ideas and new ways of looking at things. For example, we’ve had people working on large language models [LLMs] who are looking at natural language processing [NLP]. And then they run into an ethicist with a background in political science who questions some of the things they’re doing or how they’re releasing their software without particular safeguards.
A technology with a mind of its own
Lareina Yee: Can you tell us why you think AI represents such a massive, profound change?
James Landay: Think about where computing itself has become part of our daily lives, like when interacting with your doctor. Education is full of computer systems, and kids today could not imagine being in high school, college, or even junior high without using a laptop or tablet for a lot of their work.
AI is this general-purpose technology, and almost every application built in the future will probably include some AI in it. But it’s a different kind of technology, and it is not as reliable in some ways. AI systems aren’t deterministic, as we like to say in computer science, where the same input always gives you the same output.
We need to think about designing AI systems differently, since they’re going to become ubiquitous throughout our everyday lives, from health to education to government.
What’s different about AI systems is that they’re based on probabilistic models, these large neural networks trained on billions or trillions of bits of data. And you can feed data into them and receive different results, depending on how that data’s processed in that huge neural network. That means they’re harder to design and it’s harder to protect against what they might do when they do something wrong.
That’s why we need to think about designing AI systems differently, since they’re going to become ubiquitous throughout our everyday lives, from health to education to government. We want to understand them better than we do the existing computing systems.
Why it’s tough to build responsible AI
Lareina Yee: The training of AI is also so important, and it raises the problem of hallucinations. Can you tell us a little bit about the science behind hallucinations, which underscores how we think about responsible AI differently with these systems?
James Landay: Hallucinations occur when these probabilistic models essentially make up facts that aren’t true. That’s a problem with these models that may even represent a fundamental problem. We’re not even sure why they occur, and this is actually one of the bigger issues concerning just who is building these models.
Right now, these models are controlled by a few large corporations, and academics don’t even have the computing power to build models big enough to understand how they work. So we are going to build large infrastructures of our societal systems on top of models that are very useful but have properties that we don’t fully understand.
Responsible AI is a field that considers this situation and asks, “How do we try to make models that don’t do harm? How do we put guardrails around them?” So responsible AI is trying to do what it can, but it’s pretty hard without actually controlling the underlying data, the underlying model, or even knowing what’s in the data.
Tread cautiously and test thoroughly
Lareina Yee: A lot of businesses are leveraging their data by combining it with a base LLM. The proprietary piece is largely their data. How do you think about the right testing and understanding in that context?
James Landay: There is a little more control because you’re feeding your data into the model to fine-tune it or even just to look something up. So while there is a little more control, again, the underlying model doing a lot of the work is using unknown data.
Companies are going to have to be very careful and really test things very thoroughly. That’s the best bet they have right now, putting guardrails around things, essentially like blacklists that look for certain words or phrases to never mention.
I also think we’re going to see a new business model selling those services, lists, or underlying base LLMs that implement those kinds of things, depending on what a client wants.
Societal questions requiring answers
Lareina Yee: With the level of excitement over AI and kind of a call to action, what are some of the questions that you believe need to be tackled at a societal level?
James Landay: One, we may need better design processes to include a broader set of communities impacted by AI systems. That may help us get at some of the problems earlier on that we can fix before they are released with negative consequences.
Two is education, making sure students going into computing and AI have more of an ethical basis to think about their decisions. At Stanford, we’ve implemented something called “embedded ethics.” So instead of requiring only one capstone ethics course, we embed ethical lessons in different courses along the way. This is something we unabashedly borrowed from Harvard.
But finally, there are going to be some things that happen that cause harm, because somebody either had bad intentions or simply made a major mistake. And in that case, that’s where law and policy come into play. We need to make sure that if you do something bad with AI, it carries a cost.
Hopefully, that stops people with bad intentions in the first place. It will also cause companies to make sure they’re being careful to avoid the downsides from legal risk and then also reputational risk.
You can have good intentions and say, ‘I’m going to do AI for healthcare or education.’ But if you don’t do it in a human-centered way, if you just do it in a technology-centered way, then you’re less likely to succeed in achieving that good you set out to do in the first place.
Good intentions aren’t enough
Lareina Yee: This is very much in line with something you once said that drew quite a lot of attention, which is “‘AI for good’ isn’t good enough.”1 Can you expand on that?
James Landay: You can have good intentions and say, “I’m going to do AI for healthcare or education.” But if you don’t do it in a human-centered way, if you just do it in a technology-centered way, then you’re less likely to succeed in achieving that good you set out to do in the first place.
So that is really the introduction to a design process that goes beyond designing for just users because AI systems are different in that they have impacts beyond the immediate user. They can impact a broader community around the user, so the design process should consider how to bring those folks into the conversation around designing an AI system. And we might find that some of those people should be our users as well.
Finally, if an AI system is really successful, it becomes fairly ubiquitous and may start to have societal impacts. So designers of these popular systems might want to ask themselves, “If the system I’m building is successful, are there any negative impacts it might have? How might I mitigate them?”
And they should think about that in advance so they’re prepared to deal with any issues. It’s much less expensive to fix some of these problems early in the design process than after you’ve released a product.
The benefits of diverse and interdisciplinary teams
Lareina Yee: I think it’s neither here nor there. It’s actually more about asking people to change the way they’ve done things, to redesign what product development looks like in today’s digital economy, yes?
James Landay: Yes. It’s going to require changing processes and actually changing people as well. Right now, we mainly have sets of engineers, like responsible AI groups or safety teams, who are meant to check products before they’re released.
Unfortunately, there’s a lot of incentive to just push something out the door. And these teams don’t quite have the social capital to stop it. A different way of doing this is to embed a lot of that expertise in the original team.
So we need teams with these different disciplines—the social scientists, the humanists, the ethicists—because then some of those problems will be found earlier. And as team members, those people will have the social capital to make that change happen.
For example, we saw a lot of examples where computer vision systems could not recognize Black women or people of color in general. Those problems weren’t that hard to fix in the end, but they weren’t found until those companies released them and were publicly shamed.
And different companies dealt with it differently. Some immediately went and fixed it, while some fought it. So part of this is changing the process, and part of it is changing the teams. They need to be more diverse and interdisciplinary, and that will help solve a lot of these problems.
AI and the future of education
Lareina Yee: There is a lot to think about, but this is just a portion of your research. I was also watching some pretty amazing work you and your PhD teams are doing around the future of education. Are you optimistic or pessimistic about the impact of generative AI on education?
James Landay: I’m very optimistic. I think AI in education is going to be huge. Now, I don’t envy anyone with young children right now, because I do think the next five years are going to be a really rough time in education at all levels as the system tries to understand how to deal with this technology. Educators are asking themselves, “Do we ban it, do we allow it, how do we change how we teach, and how do we change how we evaluate?”
AI is going to force those questions, and some schools, teachers, and administrators are going to be dragged kicking and screaming all the way, but some are going to embrace it and do something smart with it from the beginning. So it’s going to take a while to figure it out, but in the long run, it’s going to change the educational system in a lot of very positive ways.
Lareina Yee: Can you tell us about those positive ways and why you’re not one of the ones kicking and screaming?
James Landay: AI is going to provide people with a personalized tutor that understands where people are having difficulties and how best to motivate them. Because both kids and adults respond better to different motivational strategies tailored to meet their needs. Imagine a tutor who understands all that and helps you learn, as an addition to regular schooling.
In my research, I’ve found it’s also useful to target folks who maybe don’t fit in the traditional educational system, who just don’t think that’s what they’re good at. How do we motivate those folks, take advantage of their capabilities, and allow them to learn and eventually contribute to society?
We’ve looked at how to make narratives and stories as a way to draw kids into learning, and my “Smart Primer” project is based on that concept. We’ve written different stories where you have to engage in learning activities as you read the story. And by engaging in the learning activity, you get the story to move forward.
We use AI in many different ways, whether it’s using augmented reality to recognize objects in the real world or even using AI to get a kid to write more.
From index cards to AI flash cards
Lareina Yee: One of your teams took a look at Quizlet’s online flash cards and made it a richer experience. Flash cards are how I grew up, writing things out on index cards and sitting with my friends and testing each other for a science exam the next day. How is the concept of an AI flash card different from my good old index cards?
James Landay: We did a trial in China where we were trying to teach expats Chinese. And one of the ideas we tried was using different flash cards tailored to the context of your location. So if you’re in a taxi, they teach you how to talk to the driver. And if you’re in a restaurant, they teach you how to order food. So we used AI technology to take advantage of the context and location to drive the flash cards.
Lareina Yee: What’s interesting about those cases is that it’s starting to define a different relationship children have with machines.
James Landay: Yes. And we should think about that. What does that mean? What kind of relationships do we want? Do we want a kid’s AI agent to be their teacher, or do we want it to be a tutor? Do we want it to be a companion? Do we want it to be a pet? None of the above?
That has to be thought about and designed, but we have to decide what we desire. I think we’re still far off from really having all of that, but those are the kinds of research questions we need to consider now. Because the technology will be there in a few years to allow these kinds of things.
Subscribe to the At the Edge podcast
Upheaval ahead for universities
Lareina Yee: James, let me ask you the contrarian question. The academic institutions you’ve been a part of—UC Berkeley, Stanford, and Cornell—are all more than a century old, with a rich tradition of traditional education excellence. To put it plainly, is the juice worth the squeeze, considering how challenging AI is going to be?
James Landay: Higher education has worked, and these institutions have been successful. But they’re not perfect, and they’ve changed in the past. The American university system was in some ways a modified copy of the German system, which was a different version of the British system.
So these institutions have transformed over time due to major societal and technological changes. And I think AI is going to change the educational system because it can’t continue to exist the way it does today, which is largely based on rote learning and certain ways of evaluation, which is hard to do with the AI tools out there.
So that change due to AI is actually going to lead to other changes in the educational system. And in the next five years, people are going to see a lot of upheaval. But in ten years, we will look back and think, “Wow, we’re really educating people better than we were ten years ago.”
Lareina Yee: With all of that change, let me ask you some fun questions. I know that you enjoy skiing, scuba diving, and many adventure sports. If you had your dream gen AI application related to adventure sports, what would it help you do?
James Landay: Many of us have tried using gen AI to plan a vacation with some amount of success and failure. But one of my students has a smart ski-boot insert that helps her make better turns, with an agent that’s speaking into her ear as she’s skiing. So even while skiing, you can have a personal coach who’s watching every turn and telling you what you need to do better. I think for helping us get better at things we like to do, AI is going to be great.
Lareina Yee: Going way back, I read that your dissertation was one of the first to demonstrate the use of sketching in user-interface-design tools. If you were to be a PhD student again, what would you focus on?
James Landay: That’s interesting. Although I’m not a technical-AI expert, in that I don’t create algorithms, going all the way back to that PhD dissertation to my research today, I’ve used AI in the systems I’ve built probably 75 to 80 percent of the time. And with the AI capabilities we have today, I could build them all way better.
A lot of computer science, at least PhD research, is time travel in the opposite direction. You’re trying to imagine what something might look like in the future, simulating it with the technology we have today but imagining it’s going to be faster, better, and cheaper.
And sometimes, we’re just too far ahead of ourselves. So in some ways, I was imagining something in 1995 that I thought would only take five years. But it took 20 to 30 years for that technology to become good enough to do what I was imagining at the time.
