Stanford Webinar - Design & AI: Envisioning AI-augmented Products

I'd like to introduce our speaker today. Professor Michael Bernstein is an associate professor of computer science and the STMicroelectronics Faculty Scholar at Stanford University. He is a member of the Human-Computer Interaction Group. And his research focuses on the design of social computing systems. His research has won multiple awards at various conferences and has also been featured in— on the New York Times, the New Scientist, Wired, and the Guardian. Michael has been recognized by various scholarships and awards for his research and activities. He holds a bachelor's degree in symbolic systems from Stanford University. and a master's and PhD degree in computer science from MIT. And I'm sure you will enjoy listening to him today. So with that, I'll turn it over to Michael.

All right. Hope you're all having a great day. So yeah, that's right. My name is Michael and we're going to dive in. We're going to talk a bit about design and AI and how that all sort of meshes together into the future of what products and interactive systems might look like and what it might not look like and how we actually understand what's the right way to do this and what's the wrong way to do this. So just to give a little bit of my own picture, I'm interested in how we build interactive systems that bring groups of people together, and this is often drawing on artificial intelligence, you know, whether that is looking at ways in which we can help use AI to make teams better collaborators, that we can understand how crowdsourcing can help measure and benchmark AIs, building some technical infrastructure, looking at ways in which ideation can be collectively experienced. This is the kind of work that I spend a lot of time thinking about. I build these, you know, everything from sort of algorithms up all the way to interactive systems and products in the research world that I live in. Now, what I'm here to talk about today, what we want to discuss, is how basically we have this We've had this sea change, and I'm going to start, you know, not assuming much technical background here, just to sort of— we're going to start by talking about how there's been this sea change in artificial intelligence or AI and how it's gotten much better. We're going to talk about how that's translating into organizations embedding AI into their products, into their processes, and sort of this divergence between sometimes this goes spectacularly well and sometimes it goes not quite as well as you'd expect, as you'd hope. And so the question that we're kind of here to talk about is how do we make sure that our experience is more on the spectacular side and less on the, uh, the, the dumpster fire? So I'm going to start with a few terms, okay? So when I talk about artificial intelligence, we're talking about a subfield of computer science that's trying to automate performance on tasks that we would normally consider to require cognition of some sort to solve. So, you know, we usually think of cognition as residing within our brains, but there's a long philosophical debate we're going to kind of elide here to just sort of say when we talk about AI, we're talking about computers being able to do things like say, what's in this picture? What does this sentence mean? How do I get across this room in order to open the door as a robot? And so on. Now, AI has been trying to do this for many, many years. And what's been happening over time has been a revolution through an intersection of this with the area of statistics. This is called machine learning. Machine learning basically is not— it's not me sitting in a room saying, okay, I think this is how it should learn to open the door. Instead, we provide a lot of data, and that data trains the artificial intelligence to sort of build a statistical model of the world, and then it tries to act on it. It's like the notion would be just like how a human baby is learning through experience, it's never given rules. These machines are learning through experience. One more term, deep learning. You probably hear this thrown around on the internet, on Twitter, and in academic circles quite a bit. This is essentially one particular type of machine learning model, and it's a model that has driven many of the improvements in AI over the last decade. You see, when you look at deep learning systems, you're often looking at something that has all of these circles and lines connecting it, looking something like this. You can think of it as you give it some sort of input, which I'm representing here as a vector. It's just a bunch of different numbers or symbols or whatever it would be. And it is going to feed that input to a series of interconnected nodes that are learning weights, which again, we're not going to really get into, but it's sort of learning what to pay attention to. And what ends up happening at the end of the day, your organization will probably have Something along the lines of this. Now, this is not always how it works. This is what's called supervised learning, but generally many AI applications have this general architecture. You take a huge dataset of example inputs. Let's say we're trying to— let's say I'm Redfin, I'm trying to figure out for each picture, what room is this? Is this a bedroom, a hallway? Is this a front shot of the home or what? So we got 10,000 or maybe even many more images, and then we produce what are called labels for each example. And that's going to be the thing that we wanted to learn, like this is a bedroom, this one's a hallway, this is a bathroom, this is Michael's house, he could do better to keep it up, whatever it would be. These labels are attached to each of the inputs. So you have input, label, input, label, input, label, input, label. Then you have this deep learning architecture. This is the AI that tries to now learn how to go from one of the inputs to one of the labels. It's trying to predict the label from the input. Given a picture it's never seen before of the house, it's going to try to say this is a hallway or whatever. We have all sorts of examples. These are colleagues here at Stanford where on the upper left is some work that Rajeev Krishna, a recent PhD alumnus of my group and Fei-Fei Li's group, is doing to try to say that this is a man flying a kite. That's what this picture is looking at. My colleague Percy Liang has projects looking at trying to fill in blanks from sentences. My colleagues Jeanette Bogue and Dorsa Sadigh are learning how to manipulate robot arms and so on from human mappings. It's trying to learn when I move the PlayStation controller, how should the robot be reacting? These are examples of inputs, outputs. And Andrew Ng, another colleague at Stanford, I think has put this quite well in this HBR article. If you're trying to understand what AI can and can't do, we're talking about progress through a particular type in which some kind of input data is used to quickly generate some simple response. Like input: picture. Response: is there a face? A: is there an English phrase? Is the input— is it inputs an English phrase? The output's a French phrase translated. And what he basically says is that, you know, if the typical person can do this task with less than a second of thought, we can probably automate it with AI now or soon. And I think this is not always a perfect benchmark, but I think it's a pretty good way as you're considering for your products, what is AI going to be able to do? What's it really not going to be able to do? What's the big risk here? I would say like if someone can figure this out with about less than a second of thought, maybe they're an expert in the domain. They can glance at it for 1 to 5 seconds, be like, oh yeah, that's the answer. It's probably something that you would be able to train an AI to do. So people are doing this to build AI-based products. So AI was in research labs for many years and all of a sudden it's transforming many industries. So what's going on here? What do we do about this? And what I want to point out here is that the intersection with us, with people, is where AI kind of lives or dies, right? If it successfully interfaces with people, then we make use of it. And if it doesn't, we abandon it. It's just like any technology. So I'll show a couple videos here. Hopefully you'll be able to see in the window. The first one is an example from Cynthia Breazeal's group at the MIT Personal Robotics group where, again, the point is not that we've got some sort of magical AI that is incredibly intelligent. In fact, this robot is not that intelligent, but the ways in which it communicates its intent and other parts of its state to us really become critical. Here's also Anca Dragan's work. She's at UC Berkeley, does really interesting work in this space, finding ways in which the robot can communicate what it's likely to do. So you could think, should the robot be reaching as quickly as possible for this thing, or should it be kind of wrapping around? What is a legible— and what I mean by legible, or what she means by legible, is what is a movement that people can predict what the AI is going to do? So on one hand, we have AI transforming industries. We have Siri, Alexa, and so on creating whole new verticals. In more prosaic ways, you have Google autocompleting your queries. Finding ways to suggest ways you might want to respond to that email in a friendly way, but we have to think carefully. And this is work by Wendy Zhu. She's at Cornell Tech, and it's it's an interesting example of what happens when we don't do this. So what you're about to see is a video of someone in an in a simulator of an automated driving system, and at some point the automated driving system is going to kind of lose the thread. It's going to have to hand. driving control back over to Nick here, who's, as you can probably tell, staring at his cell phone at the time. So this video is starting. And what I want you to pay attention to is he's sitting there staring at his phone, not really paying attention to driving. He puts it up to the outside. Emergency! The phone's ringing. He grabs it. He's a little confused. Cars are coming. He swears and almost gets into that car.

Oh, my God.

So what's going on here is that the AI was great until the moment that it wasn't. And that handoff back to the person, to the user, to the end user here was really risky and almost produced an accident. So we have to understand these intersections. Now, there are successes. So Netflix, as an example, is using an AI technique known as recommender systems to essentially be suggesting what we should be looking at. shaping what kinds of stuff they choose to build. So they're using what's called a recommender system. Their inputs are what other people like or view. The thing is trying to predict what other people will like or view, and it also produces these clusters of customers and clusters of market segments and so on. Now, you also have something like Stitch Fix. I think this shirt might actually be from Stitch Fix, where They're personalizing their online shopping experience, where again, the idea is based on what other people would buy, what kind of clothes might I be interested in? So again, we see similar things. What are people liking or buying? What other people will like or buy? You have the Google Assistant, you have Siri, you have Alexa. I'm sorry if my saying the name is activating your smart thing in your room, because that happens about half the time when I give this talk. These are all using technologies ranging from text-to-speech— excuse me, speech-to-text, well, and text-to-speech— as well as interpreting what it is you're asking for. The inputs here might look something like just a WAV file, like the sound of what someone's saying. The output would be some transcription and a structured request, and they have these rules for handling the commands. That's the architecture of how this thing works. Zillow, Using computer vision to improve property value estimates. So they're using something called computer vision, which is trying to make sense of pictures. Again, deep learning on a regression, meaning they're trying to predict a continuous value. So it's just same deal: input is a picture, output how much did this thing end up selling for? Here's Google creating something to help with that problem we've all had. Do you know that song that goes da da da da da na na na oh yeah? Space Jam DVD. The issue here is that you want to be able to hum something and produce back what song that is. So the inputs are people humming. The outputs that they train this thing on in some level is a melody, a song. So even in the research and medical world, you have products looking at ways in which you can monitor your glucose if you're diabetic. So again, deep learning regression, just trying to predict a continuous value. Inputs, sensor values, just stuff coming off of your glucose sensor. Outputs, ground truth measurements of glucose levels. And so it's just trying to map these things, right? It's saying input is, here's all the data we have. Output is, this is what it should have learned, your ground truth glucose level. Fraud detection, right? So many of these examples. Again, activity sequences on your website. Output, trying to predict, was this an anomaly? We labeled the things that were anomalies. Can you help us find more of them? Ranking. So here I'm a social media app like Facebook or Twitter. We want to know what to put at the top. So you're going to again take these posts, we're going to try to predict, then we're going to rank with them. So I post something like this on Facebook. It takes all the features that it thinks might matter— how close, you know, Roni and I are, whether this is a mobile phone photo, whether you're looking at this on your website or your phone, anything they think might matter. I'm just giving a bunch of examples here. That all gets thrown into a big vector that the deep learning system tries to work on top of and train to predict some notion of engagement, which it'll call a relevance score. So that might be, will you like it? Will you click it? Will you comment on it? This is what Facebook is trying to learn. So how— and what it does from this is start to then rank the items. So it says, here are these scores for every single thing in your newsfeed. I'm just going to sort your newsfeed by that score. Okay, so that's a ton of examples. What is this all about, really? I'm going to use this term intelligence augmentation. It's a— it's— I like the term because it's an inversion of AI, literally. So when we talk about AI or artificial intelligence, this is the opposite, IA, intelligence augmentation, which is really a reaction to this notion that AI is just going to take over all our jobs. Right? I think it is hopefully well established at this point that AI doesn't just push out jobs, it replaces parts of jobs. It does push out some jobs, in which case I think we need to really talk at a societal level about what that pathway looks like. But at the same time, in many cases, what's actually happening is that AI is sort of entering the job in ways that we have to figure out how we're going to integrate with it. And in the best case, what you would hope for is something that sort of is a cognitive amplifier, sort of like a prosthesis for our brain that makes us smarter. This here is a picture of Douglas Engelbart, a Turing Award winner, who envisioned many of these ideas originally when he was at the Stanford Research Institute, or SRI. It was all about trying to push back on this notion that AI is this entirely autonomous thing, that we're going to replace human intelligence with artificial intelligence. His vision of intelligence augmentation, or IA, was one in which these machines are being used to make us smarter. Can we be less biased? Can we be more predictive of the things that we want to be predictive of? Can we essentially be better at what we're trying to achieve? That's the idea of intelligence augmentation. Now, this all sounds great, but AI doesn't always work. And as my colleague Eitan Adar at the University of Michigan likes to point out, I love this quote of his: never let your UI or your user interface write a check that your AI can't cash. What he means by that is that people end up building these interfaces that are sort of promising the world, and the AI ultimately can't execute on it. I think if you want to really Close to home example, if you've ever played with, again, your Siri, your Alexa, your Google, your whatever it would be, the first time you use it, you start trying to ask all sorts of interesting things like, you know, when's my aunt coming to visit? Or, you know, what's the weather next Tuesday at 5:00 PM in, you know, blah, blah, blah. And very quickly you find out that these systems are uber brittle, that they don't work as well as you would want. And what we— what the research literature shows is that people then sort of collapse into a small singularity of things that they end up doing. So like, oh, okay, I'll just ask it to set a timer because I know it can do that. So this user interface of something like Siri or Alexa is saying, just talk to me, I'll do whatever you want. In reality, the AI can't cash that check. So it becomes a problematic thing. So you have to make sure you're not overreaching what the AI can actually do. So just to give a sense of what products might actually look like, here are some visions from the research literature over the last 10 or so years. Again, the research literature tends to work 10, 25, 5 to 25 years out. So if you're a programmer, you might not just imagine Codex, like having OpenAI write your code for you. You might imagine being able to write scripts that actually are using visual inputs so that I can say something like, hey, whenever I've got this Google Map that's tracking the bus routes or something, what I want it to do is just, I want to say, hey, whenever you find the bus inside of this area of the New Haven map, send me a text message. I don't know how to write that in traditional code, but I can certainly take screenshots and I can say, hey, if you see this thing inside of this other thing, let me know. This is solving a real human problem. At least it has been for me as I've had to deal with commuter issues over the course of my life. It's a simple script but really hard to explain without AI. You might have other examples where here— see if I can push this slide to you. Let's see. Here we go. Here's an example of disaggregation where an Imagine you just wanted to have one thing on the outside of your house that was just telling you, how much water is my sink using? How much water is my tub using? How much water is my shower using? Well, normally you'd have to instrument every single piece of your house, but it turns out some folks at the University of Washington have found that whenever you turn on your faucet or your tub or whatever, there are very characteristic changes in the water pressure of your entire pipe system in your house. And through machine learning, they can then identify when each of your various utilities is being used and then turn that into the one use that I've seen for those displays that sometimes they try to put on refrigerators that's showing you what your water usage is like. You can imagine all sorts of different visualizations of your water usage just by attaching a single sensor to kind of like the hose on the outside of your house. Here's an example for accessibility reasons. This is— imagine if our user interfaces for our phones, for our computers, could adapt to people based on their motor abilities. So here you might see an old-school print dialog, and then below it you would see one that was designed for someone with limited motor dexterity. So it's made certain aspects of the interface larger, and changed some of the things from dropdowns to buttons and so on based on a model of what this individual is good and bad at doing. And this is, you know, it basically is able to train off of this to create way fewer errors. So imagine if our systems could become more responsive to our own, my own aging as I start to lose my hair. So fashion mining, right? Looking at ways in which this might impact fashion. Fashion. So imagine you wanted to be able to just say, this is what I need, an outfit that sort of has the following characteristics. It's sort of beachy and tropical. Being able to model this and suggest what that might look like. If you're any sort of graphic designer, imagine being able to, as these folks from MIT did, create something that can show you what people are paying attention to. So it's estimating eye gaze at a visual stimulus. So here's a poster. And as I move text around, you can see in the video here that it's actually saying with the hotspots, here's what people are going to attend to on this. You notice sort of as I move things around and change the weights of fonts and so on, it's changing what people are and are not paying attention to in the poster. This is something that visual designers usually can't do. But if we can start to model through AI what people look at, then we can maybe start thinking, sort of getting out of our own head as designers a bit more. Also, in spaces like ubiquitous computing and Internet of Things, gesture recognition without having tons of, you know, VR stuff attached to us. Here's Gabe Cohn, who was at the time working at Microsoft Research, basically showing that by just paying attention to sort of the electromagnetic stuff that's bouncing off of our light bulbs in our— in the rooms all the time, it can start to identify when I'm going through various gestures. You might imagine gesturing to like, shut up the phone when it rings or something, if you're in the middle of a talk. Not that that's ever happened to me. Or how stressed you are. Imagine if your car could detect if you were particularly stressed out and as a result could start to maybe be aware, cut down external noises and so on. This is Pablo Paredes demonstrating that just by modeling how tense you are as you grip the steering wheel, they can actually estimate how stressed— your levels of stress. So drawing on some work from my colleague Scott Klemmer, I want to talk a bit about how we create products. So this is going to be sort of part 2. Part 1 was about what's coming down the pike, what might this look like. Part 2 is how do we actually go about doing this? Now, I'm at Stanford. At Stanford, we teach this. This comes out of the Stanford d.school. It's design thinking. There's these sort of sets of interrelated loops going from understanding to observing, creating a point of view on what it is you're trying to do, ideating, prototyping, and testing. And when you think about it, this is intended to be useful across essentially, you know, more or less any interactive or non-interactive product you might create. Now you throw AI into the mix and the question becomes, why does this get so much harder? Why do we keep screwing up Like, which part of this process is really tough? And for my money, it's the prototyping stage because we went from a world in which you can sort of create something that's static with pen and paper for us to play with and try out to something where you have to get a full-fledged AI model working. And that's a problem from the perspective of you know, putting a bunch of effort into something that ends up being a dud. So, you know, when I talk about a prototype, I mean something like this. This is prototypes of the original mouse that Microsoft— the first one that Microsoft put out, which was the mouse that basically many of us who are elder millennials like myself and older recognize as sort of the first largely successful commercial mouse. It wasn't the first one, but it was the one that Microsoft had such a large install base that it was essentially the first one that many people would experienced. And what you can see here in the background are all these prototypes of different versions of the mouse that Microsoft tried along the way to the final product, which you can see snaking its way through the middle. And if you stare at them, you can see how each prototype is a question. Each of these prototypes is exploring some question they're trying to answer. Like, what if the mouse were kind of arced like my hand, sort of in a semicircle? What if the mouse were curved? What if the buttons were sort of very obviously offset in different sizes? What if they had different textures? What if it kind of felt like a bar of soap? What if it was really angular? They're playing with all these different versions of it. How would you do that with an AI in which you like, you would, it would take tens of hours, maybe even months to create each version of these things? So if we go with this notion again from Scott that a prototype is a rapid approximation of a design idea that you're using to gather feedback, how do you gather feedback? Well, traditionally what I would teach in like an intro course is do this thing called paper prototyping where you make your interface on paper and then you sort of let them click, you know, sort of tap on it and you learn about it. So, okay, now is this app useful or not? Now, the problem is how do you make a paper app that's got an AI behind it? Now, it turns out they actually asked this question back in the '80s through something that they ended up calling Wizard of Oz prototyping. And I think this is actually becoming sort of back in vogue in a way that we haven't really seen in a while. The idea behind a Wizard of Oz prototype is that you have someone kind of behind the curtain. You can see someone here under the table who's effectively responsible for turning the light on and off. So the user is there, you know, playing with this interactive room, so to speak, that appears really intelligent when in reality, essentially, there's someone behind the curtain pretending to be the AI. What this does is it lets us mock in the existence of the AI by doing it manually instead of trying to build the whole thing only to find out that it won't be useful. Essentially, it says, imagine we had this AI and then what would happen If it works, and that's what a Wizard of Oz prototype does. This essentially lets us play kind of a game of time travel so that we can figure out whether this system is actually even worth building in the first place. So making a Wizard of Oz prototype isn't too complicated. What we end up doing is starting with the traditional stuff that we might do. We'd map out the scenarios, the application flow. We'd build this low-fidelity prototype. you know, often again using paper. And then we sort of develop a contract or a script, so to speak, that says, okay, wizard, human person on my team, whenever the user does this, you should sort of interpret it in the following way and respond in the following way. So it might mean that you'll— this is the thing that's later going to be replaced with the AI, but for now it's going to be me, or it's going to be Roni, or it's going to be, uh, anyone else. And the wizard is going to basically just pretend that the AI works pretty well enough. How would it— just pretend, how might an AI react to this input? Maybe you should play with what happens if it gets it wrong. Understand maybe— never design assuming the AI is going to be perfect because it's not, I promise you. So you then have a facilitator. This is someone who's kind of guiding the user through the experience. And then behind the scenes, there is a wizard who's kind of operating the interface. They are playing the role of the robot, of the AI. And then as you bring people in, they can interact with your system. Maybe it's your new Maps directions, or maybe it's your recommendations for music you ought to listen to, or podcasts, or whatever it is that you're building. And the wizard, you know, is more slow than an AI, but comes in and can operate the interface. And then you can talk to them just like you would any other usability study. So what we, you know, we start to see these things actually even sometimes put into practice. So this was a, um, a social Q&A app called Aardvark, as reported on in the Wall Street Journal, where it was trying to connect people to people with questions to people who could answer those questions. And instead of just, you know, spending years in, in, uh, skunkworks trying to train the AI to, to be perfect at this, It actually started out with this Wizard of Oz scheme. They would get questions from the beta users and they would just manually route them. They'd be like, oh yeah, it's totally a smart AI. Yeah, yeah, definitely an AI doing this when they're sitting there behind a computer screen and routing things manually. So it was all entirely done to make sure that the concept was worthwhile. And as a side effect, it produced a bunch of training data for them to actually train the final AI on top of it. Now, this is no panacea, right? It definitely is faster and cheaper and more iterative. It's more real than trying to do something that's a kind of a dead paper prototype. And you learn by what happens when the wizard's being pushed into weird situations that you didn't expect. On the downside, it's not an AI. So the errors that it makes will be different than what an AI might eventually be able to do. And it's slower, right? It's kind of like, okay, now wait. The wizard's going to have to think for a minute and shuffle some paper for you. So it's not a perfect approximation, but it's a thing that you can do in 3 hours in an afternoon. And I promise you, you will learn much more in those 3 hours than spending weeks putting together an initial version of the AI to do it yourself. So if you were going to think about it, you might try out a prototype. Like, how would you prototype an AI that knows You know, that's like your cell phone camera that's going to say, you know what, you know, that's a great photo, Michael, but maybe you ought to like move it down or up or change the framing, you know, so that you look good, right? How would you prototype that with a Wizard of Oz? Well, you might have someone holding just, you know, again, a standard cell phone camera, and then you might have the wizard who knows something about photo composition kind of over their shoulder wearing masks, of course. 'Cause this is 2022, sadly. And sort of making those suggestions, and you can see whether the person adopts them, do they understand it or not. Another example, an energy plant safety system using AI to identify anomalies and alarm the technicians. So maybe you're working at the energy— in the energy industry already. You go visit some folks who are technicians, you have them sort of go through one of their training sessions, and in one of the training sessions, you're including an additional pretend AI, and you see how they react to it. They might say, well, I don't know if I can trust this. It depends on what signals it's looking at. Like, you want— you start to learn what it is that they need out of the AI before you finish building it. Uh, or an audio application that's trying to predict, you know, what music I want to play right now based on, you know, based on my mood. This is sort of a classic thing that, like, every student wants to do at some point or another. Again, what would the wizard be doing? You grab someone who has good taste and is trying to make suggestions. See how people react to that. If they don't react to it well with a pretty good human version, then the AI version, whether it's better or about the same, also not going to be reacted to well. Now, I want to show one other sort of bleeding-edge idea. This is something we're actually currently doing in my lab at Stanford. And it is, I think, going to make some substantial changes to how we build and prototype AI systems. So this has to do with what are called large language models. You may have heard of something called GPT-3 as one famous version of this, but there are a number of them. Hugging Face and others have really good ones. The idea is that these systems, for the first time, enable essentially a natural language input to the AI, where you can sort of tell it what you want and it will produce a response. And it can craft— you can sort of craft nearly any AI that the AI could create. So, you know, here in some research we've been doing, we're interested in creating— I don't know, here's a— we want an AI that can tell whether this online comment and say that Q&A for this for this webinar is civil or uncivil. Like, you know, we don't want the uncivil stuff here. You know, please be polite. So you just tell the AI— this is literally what you would tell GPT-3, the large language model— we're trying to determine whether this forum comment is civil or uncivil. Imagine you read the following comment on a forum: go shut yourself in the room and think about what you just wrote. So let's say you just posted that to the chat. in here, and then you write, Q, would you consider this comment to be civil or incivil? A, this comment is— and GPT-3 is expected to finish the sentence. I never trained this AI. There was no training data. I didn't have to generate 10,000 examples. It actually just read the entire web effectively, or a curated version of it, and learned to predict what might happen next if it were given this prompt. And it would produce something like, say, in Civil. Literally, just, I spent 5 minutes writing this prompt, and all of a sudden I have an AI that I could sort of use as a prototype. Is it going to be as good as one that I fine-tune and make manually? Maybe not. But is it good enough to kind of quickly get a sense of whether the AI can do this and how it'll behave and, you know, wire it into a simple prototype? Absolutely. I think this is going to change the game, in my opinion. So given the results of the prototype like this, you iterate, right? So now we'd be having tightened this prototyping loop. We can go through more iterations, we can explore more ideas, and we can avoid releasing duds. So I think that leaves us with the question of what next, right? Where are we going to go from here? I think what we're going to start to see is not We will see some entirely new product categories, but I think probably in aggregate, the bigger impact is going to be AIs being integrated into many, many, many products that we use today to sort of slightly enhance the experience, to be slightly smarter, slightly more user-friendly, and so on. I hope that you use this not to displace human intelligence, but to augment it, right? To help us be smarter at what we do. And so if I'm just pulling this all together, we have this opportunity to think about how we're going to design products that draw on everything from sensor data to human language to open-ended input, and that they're typically going to produce fairly simple responses. And if you fall into that kind of category, you're in a space where AI can help. Now, at the end of the day, and I think this is where a lot of the technologists miss it, Even the best AI is useless if you're not solving a problem that people care about. So we actually have to invert that. Was, I think, one of the things that I was trying to get across today. You have to invert it and start by asking, is this solving a meaningful problem through prototyping it before you invest all the time to create these complicated deep learning models? It's really seductive, so seductive to just go train the thing. I know it. I've been there. but it's almost always the wrong idea. So instead, we use some sort of prototyping technique to see if we built this and it worked pretty well, what would the experience be like? And if that's promising, pursue it further. So I will wrap here. I wanted to leave lots of time for Q&A. I'm going to hand this back off to my colleague Roni for a few minutes, and then we will engage in some discussion, which is always the fun part.

So one thing that I think you mentioned early on, Michael, in the talk is how training the AI system systems can be quite complicated and difficult. And I think that as people think about AI, one thing that's often overlooked is the amount of data that's needed in order to effectively train these AI systems. And I was wondering, what are some ways that you've seen companies address this need to get kind of massive amounts of data in order to train? And if you have any examples of effective ways to do that.

Sure. Well, I think there are kind of 2 common models. One is when the company sort of has its own ouroboros, that is to say, where it can train on behavioral data that it already has access to.

Mm-hmm.

Netflix can train on whether you end up watching a show that it recommends. It can create little experiments and, you know, just say people watched this but not this, or they started this and they abandoned it. Facebook has signals like whether you like a piece of content or not. Now, these behavioral metrics are inevitably incomplete, right? And as we've seen with Facebook and others, when all you're doing is trying to maximize engagement, it can lead to some really negative outcomes, right? So that there are— it can be a very effective way of kind of bootstrapping, but it inevitably kind of hits some important limits and creates problematic issues downstream. Now, we can— maybe there are questions later. I have an entire other talk I could give about sort of ethics and product design of AI systems. Uh, I'm gonna put that in a box for a moment, um, other than saying that there are ways to do that better as well. The, the second way that I tend to see this happen is through manual labeling. So the company will curate some part of, uh, the dataset, and then we'll go to either some internal, uh, set of annotators that they've hired. So many, many major tech companies either have their own workforces. Mary Gray has an excellent book on this called Ghost Work if you want to learn about how this works. A common one today is Appen, A-P-P-E-N, that many companies will turn to to get the data labeled for them. This is often referred to as crowdsourcing or data annotation, and they will go and get something on the order of 10,000 or 100,000 of these items that they grabbed annotated. Then ultimately, they go try to train the model on it, look at the errors that it makes, try to go find more examples in those spaces where the errors are, go label more data there, and iterate. It's always an iterative process and it always takes more time than you'd like.

Perfect. Thank you. There are quite a few questions related to the prototyping and the Wizard of Oz idea. I guess maybe let's start with the A question that has to do with prototyping in general. Basically, as I understood your explanation of the Wizard of Oz prototype, we're trying to understand whether an idea has any merit, regardless of whether it's an AI-based idea or not. We're trying to figure out if there is a need for it in the market and how will users react to it. And I guess that's an idea that's probably true across any product development activity.

What—

how easy or difficult is it for, for, in companies that you've seen to pick up this idea and really start to use prototyping as a daily practice that they kind of implement in their company? What are some suggestions for making this successful in a company if you want to make it, make this kind of a mindset that you're, you're pushing into the company?

Yeah. So this is, you know, it's a shift in the culture of the company to succeed at, like, you know, a culture of prototyping. And just to sort of put as fine a point on it as I can, many companies have brainstorming sessions, decide what they, you know, in those internal brainstorming sessions, what's going to be valuable, and then will basically put out a bunch of engineering goals and they'll build that thing and they'll launch it. If you're sort of design thinky, prototypey in method, what you're doing is you're instead saying, okay, that does seem like a great idea. What's our riskiest aspect of that idea? How do we build a prototype that isolates and just reduces the risk in that dimension? So for example, you know, we once had a project that was looking at trying to see whether if we could use AIs to intelligently suggest You know, norm shifts in team collaborations. Like, you know, imagine a little bot in Slack that's saying something like, hey, I know this sounds a little silly, but what if you all, you know, sort of appointed a leader for the next hour? Or what— I know this sounds a little silly, but what if you all tried to sort of be kind of a little, a little extra peppy for the next hour? Just try it. And we're learning what, what's working, what's not. Now, what we end up doing is like, yeah, we could build this whole AI, But at the end of the day, we can convene some teams and do this. Like, the big risk is, do people like pay attention to this? It's not like, could the AI learn this? Yeah, we could probably learn something, right? But is this actually going to change behavior is the big question. That's why we build the prototype, is the prototype lets us strip away everything except that biggest risk and focus just entirely on that vector. And we say, okay, well, people seem to adjust toward it. But now there's a different biggest risk. Would anyone ever realistically spend enough time with this system for it to learn something? Okay. Well, then for that, the question that we need to prototype the answer to is like, well, how quickly could it actually learn? And there we create a very basic version of the AI and see how many, in a simulation or something like that, say how many iterations does it need to learn something? And so every point we're just trying to kind of Squish in this vector, squish in that vector, squish in this vector. Each of those is a prototype and the prototypes are sort of disposable. But over time we gain that confidence. If you want resources for this kind of stuff, I would look at Scott Klemmer's online Coursera course, which teaches the sort of human-centered design principles for interactive systems. Anything that comes out of the Stanford d.school on the design thinking process, you know, IDEO, Excuse me, inhabits much of this. You know, it's really a mindset shift in how you go about product design and development.

Wonderful. And so then if we take— if we kind of continuing a little bit on this prototyping Wizard of Oz path, now if we're thinking of a specific prototype to test an AI system, a Wizard of Oz prototype like you described, can you describe the path of what happens after you've done the, you know, you've done the prototype? What are some metrics that you use to determine whether this— the user is interacting well with this and it's worth continuing, and how would you take it to the next stage in the process?

Sure. You know, in Bill Buxton's book Sketching User Experiences, he makes what I think is an excellent, excellent distinction between getting the right design and getting the design right. What he means by that is that there's a difference between sort of picking the right product to create and then sort of iterating your way toward the best version of that product. Getting the right design sort of is picking which mountain you're gonna— which hill you're gonna climb, and get— and getting the design right is, is climbing that hill. So when you talk about metrics, you have to think about first, which of those 2 things am I trying to do? Am I trying to say— I talk about what's the risk, like, is there an existential threat of whether people would actually make use of this system? Then the metrics you're actually paying attention to are something like, do they adopt it in their usage in the task or whatever? Or do they just ignore it? Or do they find it grating? Or do they stop using it after 24 hours? Whereas if you're trying to get the design right and iterate, then it's actually, you may have some other metrics that you would care about that are a little bit more micro. And so I actually think that those metrics are selected based on what your current goal or problem is. So I could give you essentially, you know, some ones I would use, but ultimately what I would do is I would start by asking, am I getting the right design right now, or am I getting the design right right now? And if I'm getting the right design, I'm trying to sort of— what I'm trying to essentially ask is, is this thing worth pursuing at all, or should I change it substantially? And if you're getting the design right, then it's sort of like, You're asking questions more like, where do they get caught up? What kind of errors are happening? Where are the bumps in the road? You don't care about the bumps in the road if you're not even sure if this is— if you haven't proven to yourself this is the right thing to build. But once you've decided that it is, then yeah, the bumps matter a lot.

That's great. So we've gotten a few questions about the notion of intelligence or IA. And I wanted to see if you can share some examples on how this might impact work. So you said, you know, some folks are worried that with AI, jobs are going to go away. And it looks like through your description of the IA, basically we're looking at helping people perform their tasks better. Is that a right representation? And what else can you say about its impact on future work?

Sure. So if you want to learn more about this, I would point you at my colleague Erik Brynjolfsson, who's here at Stanford. And some work that he and others have done, I think, make an astute point that it's not that jobs, quote unquote, are the right— jobs are not the right unit of analysis for when you're considering what AI will displace. Like, people used to say, oh, You know, we're going to lose all teachers or something like this. We're not going to lose all teachers, but what will happen is that tasks within jobs are going to get displaced. So you can ask something like, is grading going to change? Or is, you know, uh, long-haul trucking, just like when you're, when you're in a straightaway, is that going to change? And I think the thing to ask is, what's going to happen when this task gets displaced from this job? By AI, is it something where now the worker is going to be forced to do more faster? And is that an outcome we want? You know, is it doing something sort of against their will in a sort of Taylorist point of view? Is it something where it's going to help you know me in an information like challenging environment? Like I'm I don't know I'm trying to make some product decision. Make me smarter at making that decision. That's the space where I think we should be investing more energy. I, you know, I think that there is so much that just sort of tries to picture AI as a black box that's going to sort of fit in and, and sort of you pull out some old other piece of the clock and, you know, you put in the shiny new part. I think that's totally wrong. I think instead we want to be thinking about what are these ultimately in service of? Whose expertise can be augmented here? And I think that essentially it's about asking, like, what are the tasks that each of us in our jobs are engaged in and what is and is not a good impedance match for that technology? I think we'll find that a lot of stuff isn't. You know, maybe I'm like, look, I'm in a computer science department. I'm not a Luddite, but like when, you know, when there's so much smoke around the metaverse and, uh, and, you know, crypto stuff and, and AI. Like, I think we have to ask what problem, what human problem is being solved and make sure that there's, you know, I think that's what tells us like what's actually at the core of this, or is the rest of it just, you know, sort of technobabble?

Okay, so I guess probably 2 final questions. So what, in your opinion, what industry is moving the fastest and has the most investment in AI? And what are some industries that could use more kind of investment resources and focus and benefit from additional AI tools or products?

Well, the first half of that problem— of that question seems easier, so I'll start there. Which is basically the tech industry. The tech industry for a long time has been very data-driven. It knows how to do data warehousing. It like had— has the technical expertise internally to train these models. And so I think you saw Silicon Valley adopting this fairly early on, even when it was less performant than it was today. In terms of verticals that have been behind, you know, I think a lot of the challenge is that for many of these verticals, they are— they less naturally have data. So if you think of many, many, many applications in, say, like sustainability, climate, environment, like sensing is really hard in those environments. You can't like just go like throw smart dust through a forest to understand changes in temperature and humidity and flora and fauna. It's really tough. And so I have colleagues at Stanford who are working on this and they end up using things like satellite data to try to make inroads. And it's because there's so little structured data. So I think that these are domains that could be truly transformed if the data were there, but the sensing layer kind of remains a challenge. I think that another thing that's sort of a slow revolution that we will see, that's starting already but will continue, is around basically medicine, human health. Again, this is a space where we are very cautious about deploying new interventions, and so I think that that That's less about technical issues, though those exist, and more about sort of ethical issues, practice issues, regulatory issues. I think we'll see it over time, though.

Great. That's super helpful. Okay. So I guess our last question for today, which is a big question and probably a topic for a different talk or lecture, but we've spoken or We hear a lot said about bias in AI. What are some ways to address that? What are some things that you want to highlight as we close out this talk related to biases in AI systems?

Oh my gosh, I want to give a whole other talk now. Okay, so I'll try to give the shortest version of this that I can. AI, saying that AI has the possibility of creating negative impacts on our society It's like, that's a horse that's left the barn. Like, it's so gone that like the horse has like, you know, raced, you know, across state lines, raced across the continent and like gotten a shoe company scholarship. Like, it's like, it's a huge issue. And there's some really excellent work, you know, in communities looking at like FairML, Timnit Gebru and many others have been involved. Ruha Benjamin, if you look at her work, Especially coming out of many folks who are like women and people of color. I strongly, strongly encourage you to read their work, not people who look like me on these issues. But there are basically 2 things that I think are going on. One is the models are picking up on implicit biases in the data. So, you know, an early example of this had to do with recidivism prediction in the justice system. turns out, you know, there are biases in who sort of winds up in the justice system and, you know, getting picked up by crime that has nothing to do with the person's color, but the algorithms learn that pattern. And then what they're trying to do, these algorithms are essentially trying to say, look, we want the algorithm to still work, but we want it to sort of have protected classes that it should be, you know, equally of treating, say, whites and people of color. So it's sort of like saying we know the data coming into this is problematic, and we're going to try to sort of boil it off a little bit and try to fix the issue. This is— there's been a lot of work in this space. I strongly, strongly encourage you to check it out. Also work that's just trying to document the issue, like data sheets for datasets or algorithm audits. Really good work. The other sort of issue here is how these things get deployed in sort of structurally racist ways. So you could have a perfectly fair algorithm and then hand it off to a judge who decides conveniently to ignore it when it sort of disagrees with their gut. My colleague Angel Christin here at Stanford has investigated this. And so From that perspective, it's totally incoherent to focus only on removing the bias from the model itself. We have to remember that these models are getting embedded in products that are embedded themselves in, like, structurally racist and sexist systems. And so for that, you have to sort of ask, how is this thing going to be abused? What groups are going to be harmed by this? And again, I keep gesticulating toward this being another talk. If you want to see about it, Uh, we might— I recently published a paper in the Proceedings of the National Academy of Sciences that's showing what we're doing here at Stanford to deal with that. It's collaboratively with Margaret Levy in political science, uh, ethicists, biomedical ethicists, and so on. And what we're essentially doing is early projects are being, um, reviewed by an interdisciplinary panel of experts asking, what are the possible issues with this once this thing goes big? Like if it leaves the lab, if it becomes commercialized, gets used by an authoritarian government. Trying to foresee what those issues might be and introduce mitigating approaches early on, long before the horse leaves the barn. So I think that that's ultimately what we need as we're talking about the impacts of these things on society. So again, I have much to say about this. It wasn't really designed to be the topic of this particular presentation, but it is, um, it's a thing that has to be integrated into every product development process. Or you're going to end up with some really embarrassing egg on your face. And you're going to harm people. So that's one of the reasons it's embarrassing, right? You probably didn't intend that.

Thank you. This has been fabulous, and I think definitely incentivized to listen to the other talks and read the other publications that you've had on this topic. So I would like to close off the webinar now. I've seen lots of questions about different references that Michael made to, to other speakers, to books, to talks, please, we'll be sharing the recording of the webinar. And I suggest that's probably the best way to kind of go back and see what those were. And I want to thank Professor Bernstein again for joining us today. And I want to thank all of you for spending this hour with us. And we hope to see you again in one of our future programs. Goodbye, everyone.