 Welcome everybody, this is Sartreau of LS NTAP. We have a wonderful presentation today going on over artificial intelligence and the law. We're about one minute before 10 here and we will get fully started here in about 60 seconds. I just wanna remind people that this webinar is going to be recorded and it's going to be archived and put online over at our YouTube channel. We've just did our first half of a redesign of our website. Our second half is gonna be coming here before June, but I'm dropping a link into the chat for the, where the video will be up, which is NTAP videos. We've got over 200 of our recent trainings there and available. We also have a talk that we recently did on data standards and using machine learning in order to improve how we categorize things in legal and then how we're updating the National Subject Matter Index as part of that project. And that was with Margaret Hagan out of Stanford's design lab. Turning it over at this point to Sean Robbins of Pioneer Square Labs and please feel free to ask questions at any time throughout this presentation. There are two ways to ask questions. One of those is to drop that into the question box. I can read those out. The other one is to use the raise your hand function and then I can unmute people. We have a smaller group today and we are more than willing to have an interactive conversation. I'm not gonna answer any questions that are out there but we do have content that will last over an hour here also. So thank you for coming in today Sean and working with us to give us some basics around AI and then also look at some of the challenges in the future. Thanks a bunch, sorry. Am I coming in clearly enough? Definitely, sounds good. All right, it's awesome to be here. Thanks everyone for coming and everyone who's watching in the future. Sean Robinson, I'm the principal data scientist at Pioneer Square Labs, technology incubator down here in South Seattle. And before the world of data science, I was in astrophysics and so some of that may come out in the talk today. And I think SART says please do ask questions or send them in text and I'll try and answer them as promptly as I can. But really do, the structure of this talk, I'm going to start by talking a little bit about what machine learning is, what modern machine learning is and what we, as opposed to what we might think about it and then go into how that drives sort of what we think of as modern AI. And then talk about some of the fallout of that and where I think we're going to get into legal issues in the future where things like standards and at least overarching understanding of the problem is valuable. So I'll just dive right in here. Oh, there we go. So there's a whole bunch of terms floating around and you've probably seen a bunch of these or maybe you work with them on a daily basis depending on who you are. And I want to try and demystify some of this. In this talk, I'm going to talk mostly about machine learning and AI and the stuff in between, how one leads into another. Basically the question of AI is like can machines learn like human beings? And as we'll talk about today, should they be allowed to? There's a bunch of sort of quotes from luminaries in the field in which people conceptualized even in really the early half and certainly the mid of the 20th century that computers would just be able to think like people and we'd think nothing of it. And eventually that if they could think like people that they might become stronger than us. And one important thing to realize is that a lot of these were written in a very early era. Maybe you're probably familiar with Isaac Asimov's Three Laws of Robotics, this idea that you can tell a robot explicitly what to do and that you can write rules into it and say, well, if you're in X situation, do Y, if you're in Z situation, do A, and so on. And that that was how artificial intelligence would work, that we'd have a big stack of rules that were clearly understandable and changeable and programmers would put those rules into some kind of robot and that robot would interact. And really in the birth of AI in kind of the 50s, this was the way in which we thought is what you would call procedural programming with well-defined terms is how we thought AI was going to work. And unfortunately it didn't work out so well. There was a bunch of progress made but it was sort of also inflated hype. There was funding for it but the funding kind of fell out in what we call the first AI winter which happened sort of in the 1980s. So that first picture of AI which a lot of people still have this, well, I'll just make a bunch of if-thens and that's how my robot will operate. I can draw them all on a wall, it'll be clear. Really didn't stretch to the complexity of modern thinking problems. It's certainly not how people seem to think even if we have sort of a cognitive wrap where we can explain things in that term but on a fundamental base level it doesn't seem to be the way the brain works. So then there was a fallout and then new hopes happened in the middle of the 80s where all of a sudden expert systems were built and an expert system is rather than come up with explicit rules of operation we decided that instead we could encode a large body of knowledge and then make a sort of parser to parse that knowledge and come up with an answer. So to which I would have a database that said something like, well, the ball is made of rubber and the ball is blue and rubber things bounce when you drop them. And then I could ask it, well, will this ball bounce when I drop it? And it would say yes, because ball is made of rubber rubber things bounce. So therefore I can answer that question. And obviously it got quite a bit more complex than that. This actually was also put forth as this idea of the answer to cognition, the way in which the brain works that would stretch to being large enough to span the entire space of knowledge and the entire space of human behavior. People said that by the late 90s we'd have machines walking around and cleaning the house and mowing the lawn or whatever you need them to do, driving you around. And that certainly didn't come to fruition, partially because we didn't have the computer complexity and Moore's law was sort of not moved far enough along but partially because the underlying mechanisms of doing that were really good for driving things that were simple like, well, not necessarily simple, but not conceptually complex. Things like doing your taxes or running a business system. And a lot of those expert systems still have their place that were sort of quietly implemented, but people stopped saying the words AI again in kind of the early 1990s. They became kind of a dirty word and nobody said, they used other things, expert systems and other kinds of business applications. But then in more or less, as you notice the early 2000s, really the late 90s, if you wanna call the shot heard around the world, it's probably the deep blue system beating Gary Kasparov, one of the great grandmasters of chess in 1997. And an interesting point was that in 1997, a lot of people, myself included, I think it was what in college at the time, looked at each other and said, well, that's quite a wonderful thing that they've just did, but we'll never beat the game of go. And the reason was that our thinking had to do with the way in which deep blue worked, that among other things, it was taking a board state that is all the things on the board in the way that the board could possibly be moved forward and saying, I will project forward from this board all the possible future games and attempting to look forward at least several moves. At least that was our understanding at the time. We said, well, that's nice and everything. Chess is a very big game, but then we have games like go and they're impossibly big. The number of go boards is larger than the number of molecules in the universe. So we're never gonna do a game like that. And in fact, people thought we couldn't or it would be 100 years from then, but lo and behold, whatever that is, about 20 years later, AlphaGo beats Lease at all and everyone loses their minds about it. And so how did we get from here to there? Why are we in this explosive growth? How in the world does the new stuff work? It's not just encoding information. It's not explicitly writing down rules. If it's not even trying to look at the future and figure out what all futures could be, how in the world in 20 years did we go from, this will never happen to, of course this happens and computers are gonna be better than humans at everything. So we're gonna talk about what AI can actually do and the way in which this operates. And in order to do that, I'm gonna need to step a little bit back and talk about modeling. So my hope is, and again, if people, if I bore you or whatever, I can speed this up or I can dig deeper into any particular piece. But I found that often people are exposed to machine learning who haven't necessarily seen it from the ground up. And it's not as complex as all that, but there are a handful of moving parts that are worthwhile mentioning. So the classic science modeling way, when I was in physics, undergrad and graduate school, computer modeling or numerical modeling was when you would describe the world in some specific way. So let's say that I wanted to like measure inside the sun and I wanted to know how the sun works. Well, what I would do is I would get all the experts that I knew together, who knew about gravity, who knew about fusion, who knew about radiation pressure and so forth. And I would say, let us model the sun and we know all this physics. So let's assume that we start with a big ball of gas and I'll simulate this big ball of gas and pretend there's so many molecules. Let them fall in together due to gravity and they eventually make fusion happen and the fusion makes radiation pressure and it holds the star at a certain size and all of that stuff. We already knew that from the get go. We had a very deep subject matter expertise level knowledge of all the dynamics that went into there. And maybe there were just one or two things that we would call free parameters. So let's say I know the sun, I know basically how it works, but I'm not really sure the ratio of hydrogen to helium in the sun. So that becomes a thing called a free parameter in this kind of method. And then what I have there is the basis of what we're about to call tuning a model. So my model is gravity happens, fusion happens, there's all this high complexity and there's a couple of things I don't know and it's maybe the overall mass and how much hydrogen to helium there is in there. So after I built this model, great difficulty and with a lot of help from the experts, I then say, well, if the sun had this mass and it had this hydrogen to helium ratio, what would the radiation come off? I can't go and put a probe in the sun, it's too expensive and far away, but I can measure what comes off the edge. So I point a telescope up at the sun and with spectrometers and so forth. And I say, if the sun were this mass with this ratio, what would I expect to see off of it? And then I look at the sun and I see how different what I see is from what I would expect in that case. And then I go about tuning those parameters. So I change the mass around a bit and I change the hydrogen to helium ratio around a bit. And in each case, I continue making this hypothetical sun. So I have a big stack of hypothetical suns. And I basically point my telescope up and I say, well, which one of these hypothetical suns looks most like the real sun to me? And then I say, well, I guess those are my best guesses as to the parameters. My best guess is that the sun has this ratio of hydrogen to helium and it's this massive. So this is really great. And it's what we thought, I think mathematical modeling was going to be like for a very long time. It requires this deep expertise. It can be really, really complex. And most importantly, it doesn't catch any unknown unknowns. If there are dynamics that are not known to your experts that are in there, your model won't tell you about them. It might be a little bit worse than its fit in how the hypothetical sun looks like the real sun, but it doesn't tell you about the unknown unknown. And more to the point, especially in the early 2000s and beyond, when all of a sudden we have data pouring into us from every direction and the world is now full of all these data problems, we have a lot of cases where this is totally not the way that we look at the world because we have a bunch of data, but no model. Here we have this incredibly extensive model and incredibly extensive expertise. And we just want to turn that into a couple of parameters and figure out what they are. But in a lot of cases, there's a lot of problems that have data, but no model. I mean, if you are, how do you best turn the wheel when you're driving down the street in a car? Or what sort of movie will you like? Who are you likely to vote for? These things are complex enough that there is no governing simple principle that you can just ask experts and make a numerical model out of it. So we need another way. And in fact, now that we think about it, humans do this same thing. I don't necessarily say, well, you'll get a baby or something trying to climb upstairs. It doesn't necessarily go, well, I know all the specifics of locomotion and I know the specifics of these stairs and I know how gravity works. So let me just tune a couple of things. They basically try a lot of stuff and they get a lot of examples of what doesn't work and a few things that do. And they sort of just sort of refine in on best principles. So we wanna do the same thing. If I wanna recommend a movie to somebody, there's no governing body that will tell me exactly all the stuff that's in a movie and exactly what makes it good or bad and what your psychology is. At the moment, too complex, but we sure do have a lot of examples of people liking or not liking movies, especially if we're Netflix or Google. So the solution that we sort of came to was that, instead we were gonna use this thing called machine learning. Basically in machine learning, instead I have a whole bunch of parameters. So I have some kind of a generalized model rather than it being specific, like the model of the sun. It's super general. I just say, I'm gonna take all the things that I can observe. I'll throw them into this generalized model and I know some combination of them leads to my answer and I'll give a solid example of that in a slide here. But basically the thing to take home is when most people say machine learning, as is at least mostly true. Most people are referring to supervised machine learning. And by supervision, I mean, the thing that I just suggested where we have many, many examples of an outcome. I have a lot of people and I know all of their medical details and I wanna know who is likely to get cancer this year. Or I have a bunch of examples of people driving a car and then I wanna know what I do in this situation if I am a car driver. So the supervised and supervision refers to all of those examples. They will be used as I'll describe to take this super general model and turn it into a specific one that we can use. There are some other ones that I will go through a little bit in this. At the end, I'm gonna talk about reinforcement learning a bit and I'll dig into that further. But there's also unsupervised machine learning which is essentially clustering and other attempts to determine the structure of data from within itself. A quick example here, I'm not gonna dig into that too much in this talk unless people want. But a quick example is that if I were to have the positional value of everyone in the United States that I could tell you that there's such a thing as cities. Even though I don't know what cities are and I don't really have any labels and I don't have a map. If you just give me latitude and longitude for every human being in the United States right now I could tell you there's something going on where we all group up and I could maybe start describing those groups or clusters as centers. So even without having a lot of labels you can still do that. This is gonna be less relevant to this specific talk but it's a worthy topic of study too. And so most of the things I'm gonna be talking about for the next few slides are supervised machine learning. And let me get real crisp about how that works. So we talked about modeling and now I wanna make a model. Let's say this is gonna be a model where I don't really know how this works but I know I have a bunch of data. By the way, this is all completely made up and I think I picked all these numbers at random so don't infer too much. But the idea is let's say that I wanna predict cardiovascular disease. I wanna know if I were to give somebody a complex test. Let's say the test is somewhat expensive and maybe it's hard to administer in the field or something like that. But I do want an idea of who should get the test and who's at high risk. So I wanna do what I call making training data which is example data of what I have. Supervised machine learning basically all works this way. Let's say I have a bunch of examples of people. In this case, I have everybody's age and everybody's weight and implicitly a whole bunch of other stuff, all of their statistics. And then there's this data that I'd like to predict or classify in the future. In this case, well, if I tested you for cardiovascular disease, would you be positive or negative on that test? And so what I'd really like to do is come up with some kind of model. And again, this is where I wish I had experts but let's say there aren't any experts in this case. And all I know is that, well, your age might matter somehow, your weight might matter somehow, your height might matter somehow, all the other things. But I'm not really sure how. But I know that if I put them all together in some way, it would add up to your cardiovascular disease risk. Now let's say I don't have any experts to tell me that, well, maybe I should look at ratio of X to Y and I just have to give it a shot. But I do have a lot of examples. I've got, let's say I do this with a million people and I give them the whatever expensive test and I've made what I call a training data set. So in the future, we wanna basically cover up that right column, have a new person come in and say, yep, I think you're probably positive or I think you're probably negative. So this is essentially the same thing as I just showed. And this is really under the hood. Basically this is what's going on when people say mostly when people say machine learning. So we're gonna look at the data occupying space. This is literally the same as this. I just put it on a chart. So now let's say that, I've said individual data one, individual data two, but let's say like, age is on one of these axes and weight is on the other one or whatever I have. Normally this is going to be a highly multi-dimensional space because we have more than two, what we call features, but it's easier to show on a two-dimensional plot. So basically this is just the data set re-envisioned. I say, okay, I'll take everybody and everybody has like a height and a weight, so or an age and a weight or whatever it is, I'll put them all on this chart and then I will color, let's say people who had a positive test as red and people have a negative test as blue. And now you can kind of start seeing the game of supervised machine learning. See, I don't really know a priori any kind of model, but I can certainly visualize it. I can certainly just put everybody out and go, yeah, well, what's the area where everybody's got a positive test? And if people come in in the future and their stats put them in that area, we're going to say it's likely they have a positive test and that's the idea. So some new person comes in and I go, yeah, would this person test positive or not? Should I give them the test? Should I make sure they get extra help? And the idea of machine learning is to more or less draw a boundary across this space. We want to try and make a boundary that classifies these new and sort of unknown data rows. So I go, with this person test positive, well, one model is a straight line lives across this space. And if you're on one side, it's positive if you're on their side, it's negative. That's like one generalizable model. I didn't know anything about it, but I can certainly draw a line and then I can make that line so that most of the blue is on one side, most of the red is on one side. And then once I do that, I can call it a classifier and I say, yep, yes indeed, that person is likely to have a positive test result. So you should give them the testing, should you give them extra help or whatever. There's a whole bunch of these. Essentially when you hear about all these complex sounding terms, tree methods, random forests, caniarest neighbors, support vector machines, most of those are all just different ways of cutting this space up if you want to look at it like that. So the one that I showed is often typified by this thing called a linear support vector machine where I go, in two dimensions, let's draw a line and you can sort of extrapolate. In three dimensions, you draw a plane and et cetera. There are also tree methods where you do exactly the same thing, only now you say, if I was only allowed to cut this space in one dimension first, where would I cut it? And then you, for each of the regions, thus may you then say, if I could cut this space in one dimension, where would I cut it? And so forth. There are rules for when you have to stop cutting and there's a thing called overfit that I'll get to, I believe in a second. Caniarest neighbors is one of the easiest ones. A term you may have heard is just, basically if I go to any point, I expand around that point and I see, I pick a number of neighbors and once I've made a circle big enough to include that number of neighbors, I take the average of them and I say, this is probably red or this is probably blue. So all of these things live and die on the quality and quantity of data available, which is going to become really important in a couple of slides. But for the moment, I just want to kind of, I don't know if I can look at this because this is the sort of full range in some sense of the way in which you do modern supervised machine learning. It really isn't much more complex than this. There, as you may imagine, there's some metric of success that you have to have to tune these boundaries. Like on the left most one, I'm trying to make a straight line across the data. I'm trying to say, well, people whose agent waiter in this region are probably positive people. Agent waiter in this region are probably negative. And where do I put the line? Well, there's a little bit of a trick with that. Basically what you're doing is starting a line in a random spot and then changing all of the, changing the sub parameters. We're coming back to parameters here. If I have a line, you may remember that you can draw a line as long as you have like a Y intercept in a slope or two points that live on it depending on what you want. And you can consider those numbers to be parameters. So if I give you a Y intercept in a slope. So Sean, I'm unmuting Claudia Johnson here who works with a lot of the Law Help Interactive forums over at Pro BonoNet. She had a quick question just to kind of get it grounding and make sure that people were understanding stuff because what we're getting into is a little more complicated here. Claudia, you should be unmuted. Thank you, sir. Sean, just a basic question. Are you using linear math or does this also apply to non-linear systems? So I'm using the example of linear math with this leftmost one. But really it sort of depends on how you ask. I want to take a step back so that I can try and answer that clearly. Basically what I'm saying is if I have data and it's scattered all over, I sort of pick a model to cut it apart and guess which thing is red or blue. I am on the leftmost thing suggesting I will put a line in there. But fundamentally the process of figuring out where that line goes is a non-linear optimization. But it's not so complicated. What all I'm saying is if I have parameters, if I give you like two numbers of Y intercept in a slope, I can define a line. And then I want to figure out which line is the best. So what I do is define some metric of how good that line is by like how much red is on one side and how much blue is on the other. And then I wiggle the line around in an optimization process until it gets the most red on one side and the most blue on another. And then after that I use it to decide. So I guess that my best answer to that is that there is a non-linear process in here but no one ever needs to really write it. This stuff has been really canned up in algorithms for a long time. So technically there's this implicit non-linear optimization problem, but it's like we don't, you know, even if you're using the stuff you never really need to think about it. I don't know if that helps at all. Well, a little bit, I'm just wondering if we're assuming the problem is linear or non-linear behavior, right? Because non-linear behavior would be a lot more complex to model. Oh, okay, yes. And use these tools on, right? So I think that what we're saying is that here we're assuming a well-understood linear problem, not a very complex, not understood non-linear behavior system. Okay, no, no, okay, I, okay, it makes a lot more sense now, sorry. I wasn't quite grabbing onto the question. Okay, so the data itself actually can be complex and non-linear. And this is part of the reason why machine learning is so valuable. So okay, in the actual dataset that I've shown, it looks like they probably have a linear relationship, right? It's like the upper right's all red, the bottom left is all blue. However, and I haven't drawn this, but it could be really complicated, right? I could have a little region of red, a little region of blue, and then, you know, there are a bunch of possibilities if I had a really complex, unknown non-linear distribution of stuff. Like let's say there was just a little region of red in the bottom left and we didn't know why. That would imply that the dynamics here are super complicated. Like, you know, we don't really understand why that happens. But this is the value. So in that case, if I were to have done that and say, well, the dynamics are really non-linear complex, we don't know how they work, some of these methods would still work well. The linear SVM wouldn't work well on it. And that's, again, a kind of inside baseball, like knowing what model to use. But that's probably beyond the scope of what we need to dig into in this talk. But the other stuff would still work. So like tree methods where I cut the space up would still allow us to represent this complex non-linear system because the models themselves are very general. I guess, does that- So you could approximate it, but it would be an approximation to a non-linear system. But it may be hiding singularities or other phenomenon that could maybe be modeled with chaos theory and stuff like that. Yeah, and this is exactly- I'm glad, this is a really good question, by the way. This is exactly why modern machine learning is so powerful, is because I can say, hey, a super complex system and nobody knows how it works. I don't know, it's got some weird little regions, we don't know where they come from. Maybe there's seven dimensions in it, there's no experts on it. And I wanna make a model that's like that. It's like, well, I happen to know that this is a non-linear dynamics region and if I draw the equation, then it'll allow me to make it and there's like fractal behavior and all this stuff. But I don't know any of that usually, I just have a bunch of examples. So these are a way to approximate a model that like an expert would make, even if we don't have an expert. This is exactly the heart of why the topic is so powerful and why it's being used everywhere because, yeah, I'll go, I don't know how disease works and there's like 17 different dimensions of what I can measure in a person. There's complex stuff and it's probably non-linear and it's probably a bunch of weird behaviors. But if I have enough examples, I can still use these methods to kind of approximate the model an expert would make. I guess, does that make sense? Yeah. Cool, okay, yeah, that's actually a really excellent, excellent segue into a lot of the stuff that's coming up again. So that's why this is so great. Like, if you have a bunch of examples, you don't need experts anymore. You just go, yeah, I don't know how it works, but boy, I sure have a lot of examples of it. So I put them all up in this multi-dimensional space and I take my generalized models and I go, hey, generalized models, see how well you can fit the data that I do see and then I don't really need to know how it works. I just need to know if I see another example, is that like a yes or a no or a dog or a cat or do we turn right or do we turn left or whatever you're asking the machine learning system to do. And this is why, that exact reason is why the topic has exploded so dramatically because it can answer questions like we have experts where there are no experts. And it's one of the only things that can do that. Wait a minute, just, there we go. So because of that, it runs a bunch of stuff nowadays. Like there's, you know, people use neural nets which we'll get into in a sec for like crop yield estimates, people figure out your credit score this way, there's a bunch of office software and you know, software as a service offering and so forth that are all run on this. But all I've really done so far is said, I have a system that can predict. You know, if you give me a bunch of examples of a system and then you say, well, this product costs $2 or this is a dog versus a cat or you know, this person is likely to have diabetes so give them the test. How does that add up to like a robot? You know, there seems to be a missing piece. You know, right now I have a very powerful thing and this can, you know, figure out what kind of product you wanna buy or anything as long as you have a bunch of examples. What movie to show you? But how does this add up to autonomous behavior? As I'll point out near the end, all already what I've described within machine learning can get us into some problems especially in the ethical space. But before that, I wanna kind of go to a little bit more of a fun space and talk about real autonomous behavior. You know, we think about self-driving cars but for the sake of discussion. So, you know, when I work around this professionally to me a robot is the AI agent underlying the robot. So there's a lot of stuff. I think, you know, on average most robots are not actual physical beings walking around or like drones flying themselves. Although, you know, nowadays we're starting to have cars that drive themselves and so forth and they work in the same principle but I'm gonna talk basically about the brain inside the robot. So for the sake of discussion, I have this big tool called machine learning and it can tell me, given a bunch of examples, what the answer is likely to be if a new example pops up. So I'll put that on the shelf for a second and now talk about a robot. So the second argument, a robot's gonna be a bunch of inputs and some outputs. You know, the inputs are like in this case, well, you know, all the cameras, all the feedback sensors that might have knowledge of where its body is or whatever. It might be, you know, reading the stock exchange and just have a knowledge of where all the stocks were. Then it's got a bunch of outputs. In this case, again, it's a physical robot so it's like there's motors, there might be some data outputs but that's not quite enough because if I just hook all the inputs to the outputs, nothing will happen. So I need something else, some kind of a brain in it and in order to drive that brain, I need something finally which is, it's gotta have some kind of goal. In this case, I'm suggesting that like robots trying to give a high five. That's probably not true but so how, you know, so what's that brain do? Well, I've just said it's gotta have a motivation. So we have inputs, those inputs in some ways are gonna tie to outputs to accomplish some kind of goal. And so how do we do this? Well, machine learning, which I'm now just making this small gray box here, hopefully we've gone over it enough, answers questions like this, you know, is this an A or a B? If I have this product and here's all the data about it, how much will it cost? And into this machine learning, we pour a bunch of training data which is examples of what we've seen in the past and we train this machine learning system and then it does its best to answer. Well, this is A rather than B and it costs $25. So what's about to happen is we're gonna use that exact same wonderful tool to just answer a slightly different question. And the question is what should we do here? And it still works, like all of the logic sort of scans. I just say, you know, here's machine learning, it can answer any question as long as there's a bunch of examples. So I'll ask it the question, what should we do here? And we can now feed it with a couple of things. We can still use training data which is examples of what we've seen in the past. So, you know, again, too wit if I'm a self-driving car, I just have a bunch of examples of, you know, what people have seen when driving a car and what they did. And then I make a machine learning system based around the idea of like, what would a person do? And we can go past that, which I'll talk about in a couple slides here and also use simulated data. If you need more, you can sometimes produce your own data by making a simulated environment and like seeing how the current system performs. I'll get to that in a minute. But basically, the important thing is that what people think of as AI, autonomous systems that are doing their own thing is basically just machine learning used to answer the question, well, you know, what's the best thing to do? And it's like, well, what would a person do or what's most likely to lead to the correct answer? And you still train it in like, roughly the same way and still roughly the same models are even used. So it's just kind of a hop skip and a jump from machine learning to what we think of as AI. AI is kind of a nebulous term anyway. Machine learning has a very standard, very structured definition of there being a technical or there being a metric that you wanna get closer and closer to basically an estimate you wanna make. And the more data you put into the model, the closer it gets to being the correct estimate at all times. But AI is a bit more nebulous. I sometimes flippantly say the definition of AI is whatever we said a few years ago that computers would never do. And that's in reference to how quickly the field is moving right now. But okay, well, you know, so I'd suggest we just use machine learning, but I'm gonna come up with a couple of examples here that hopefully will make it clear. What you often do to train an AI, and this is part of the magic of how they can learn is that your training data can come from a simulation. This is the other like really important piece that allows modern AI to be so good. And this is how, you know, this is how AlphaGo beats Lease at all. And everyone loses their minds about it because they were able to make lots and lots and lots and lots of examples by simulating what a current algorithm would do. So basically in this case, this is just an example. It's a bit opaque, so I'll go through it a little bit. The idea is if I have a game, one thing that I could do, like I could make a brain, and if I have a robot that's gonna play this game, and then the robot is going to be fed in what the game state is right now. Or maybe what it is in the past or whatever, but this is sort of a stateful game where I can look at where the dot is right now and that tells me everything I need to know about the game. So I make a robot to play this game, and there's however many possible game states, in this case corresponding to all the places that dot could be. And we know we want to get to the end. So essentially, I can make a matrix. It's sometimes called the Q matrix or quality matrix that maps the game state, everything I can see to the things that I can do, which in this case are just move up down or left or right. And what will happen is that I can initialize that with random stuff or zeros. And that can be in some sense, that's my entire brain. I take that matrix and say, yeah, okay, I get a game state of this, and the matrix describes how much total eventual reward I expect, how much I expect to be winning the game if I take each one of those actions. And now you'll notice that it's initialized to all zeros, which means we don't really know anything. And you couple this with an algorithm that usually tries to do the thing with the highest payoff. And occasionally, sometimes it's called the epsilon greedy strategy, occasionally does a random thing. And the reason it does the random thing is that so it can eventually keep filling that matrix in with like the expected behavior. There's some, again, some deep problems in the combinatorial and stat space that don't bear going into here unless people really want to and have some time at the end. But basically this is the idea. The idea is the important thing that, well, if I have everything that I could possibly see in rows of a matrix on my left, and I have everything I could possibly do in the columns of the matrix, then I can make this matrix of like what I expect my reward to be at the end. And it's nice because as long as you have whatever game is being played and you can simulate it, you can play it an infinite number of times and try every which possibility. And you will eventually settle down to a matrix that very well describes what to do in any situation. And now you basically have a robot. You do this for a while and eventually a robot will go, well, you know, I'm in this position. Okay, this is what I do. I'm in this position, this is what I do. And it sounds somewhat deterministic and it kind of is. But there's an immediate problem doing this. Probably you folks may have heard of the modern AIs that have been trained to play like Atari games and now other games. And the issue immediately is you go, well, yeah, but you said, you know, if every game state is going to be a row in this matrix and I'm playing an Atari game, that's whatever, you know, 640 pixels by 480 pixels or something and each of them can have any color. And so therefore I have this monstrous amount of possible game states. It's impossibly big. So you can't really, you know, you can't really make one thing a row. And so instead, some clever people figured out that instead of making that whole matrix, we could just approximate it too with a machine learning stage. And this is essentially where all of the, you know, all the self-driving cars and all the autonomous behavior and a lot of the stuff starts. But if I say, well, I'm gonna run a thing like this diagram here. You know, here's a robot. The robot is essentially now a machine learning stage that I will feed with everything it can see. And out of it will come just a few choices as to what it can do. So inputs and outputs. Then I'll put it in this environment where it's allowed to use its current, you know, what would a person do question to play a game. And after a while, it is rewarded or punished based on how well it's winning that game. And it changes its internal parameters around until it wins the game the most. And then we use a lot of computation to do this, you know, trillions of times or whatever. So that eventually we settle the machine learning thing inside the robot down to the best guess as to what would be a winning move. I hope I'm doing that clearly, but it's gonna be needed to talk over this. So here's an example. And here's where all the problems are gonna start. This is the most powerful thing in the world, but it's also where all the problems are gonna start. Whoop, this is that. Let me get that. Okay, so this is a reinforcement learn trained. This is not mine, I borrowed this and you can look at the link. But this is a robot in a simulated environment. So all it has is two hubs. It's basically got three parts and it's got two hubs and it has the ability to add torque to each of its arms. So basically it's got two little arms and we train this with a neural net, which I can get into in a minute, but we train this with machine learning stage and we say basically you have a bunch of parameters, change those parameters around until you are touching this ball the least. So its inputs are just where the ball is and where it is right now. And its outputs are how much torque it puts on each of those arms. So what happens is we don't necessarily know what it's going to do, but you see as we continue to train this it gets a little bit better and it gets a little bit better and it starts to touch the ball less. It figures out that it can sort of stand on one thing. And then there you'll see it kind of figures out this sort of jumping behavior, where if it's internal machine learning parameters are set such that they will correspond to it kind of getting way out of the way of the ball, then it gets rewarded more. So we continue to follow down that path. And what's gonna happen, and again, so we're just iterating this game, we're choosing a reward for each outcome. Implicitly there's this exploration exploitation problem where some of the time this robot is doing what it thinks is going to be the best and occasionally it's doing a random thing so that it can fill out its space space. But there you see it around generation 71 where we continue training this. This thing gets the idea to anthropomorphize that it can jump, that it can sort of hurl itself up because gravity exists. Now the really important thing is that this thing learns to jump over the ball, but we didn't tell it about gravity. It doesn't know what jumping is. It doesn't have a concept of anything. It's literally just training a thing to get out of the way of the ball and it figures out, well, jumping is the right answer. If it had more segments, it might be able to do an inchworm thing where it sort of inched along and let the ball under. But the important thing to all this is not just that it kind of looks cool, it's that we didn't tell it to do this. And that's in my opinion at least where all the trouble can begin. We can train something to do behaviors that we don't know how to do. We've seen this already a couple of times. In fact, the latest AlphaGo that's even better than the original AlphaGoes are trained just with simulations and in the absence of first priming it with human behaviors. And as a result, they're winning at Go games even better than the old ones were, but they're doing things that seem somewhat inhuman. They're apparently, and I'm not an avid Go player, but they are taking moves that humans would not have expected and so therefore would not have put into the system. So this is sort of glorious in some sense, but also it's a little bit spooky. I'm gonna play this and talk over. This is another one where now we've made bipedal creatures and nothing is put into them except they get penalized if they fall down, if most of them touches the ground and they get penalized if they are the first one that that happens to. And so we're not doing anything except just training that and then putting them in this environment early stages they just fall down and they eventually learn to walk and you can do this with other kinds of, again, this is not mine and borrowing this from people who did some awesome renderings, but this has in some sense, this great power to it. These are spiders learning to do sumo in the same way. They're essentially penalized if they fall out of the ring first and they're rewarded if their opponent falls out of the ring. And over time they figure out, A, how a four-legged thing like that would walk and B, how to sort of flip their opponent over and get a leg under and then kick the other one off. We didn't tell them to do any of this. We just said you win if your opponent falls down first. So people think, this is probably the crux of this and why I spent so much time talking about AI and the way that it works is people think in terms of these Asimov robot laws. Well, tele-robot explicitly do not hurt humans, explicitly obey humans as long as you're not hurting humans and so on. But in fact, none of that is the case. In modern AI, we have fundamentally sacrificed our ability to understand what is going on for being highly effective at it. So generally, the relatively complex but highly general algorithms that you use to drive the machine learning stage in between input and output here. Each of these things knows roughly where it is and it knows where the ground is and possibly where the opponent is, but that's it. And we put no stipulation on it other than I'll make a machine learning thing that will guess what the correct move is and I will train it a million times and it will figure out stuff like this that actually starts after a bunch of training to look kind of like soccer where one of them is trying to block and the other one is trying to get a goal. But we didn't say anything about people. We didn't tell what soccer was. We just said you were rewarded for this kind of outcome. So we have to start being careful about accomplishing this because we're no longer in a space where we can say, well, but make sure that we take this action, this understandable cognitive action. We instead can be very good at whatever we want but everything now revolves around, everything now revolves around what we call the loss function or the merit metric or whatever you want, which is basically what you have rewarded. So we can make very complex systems. We can no longer tell them anything. We can just reward outcomes. So now all of a sudden what outcome to reward is vitally important. So with things like self-driving cars, this is what gets us into trouble kind of immediately. And I'll go on to talk about that. So there's gonna be, I'm gonna come up with some scenarios. These are gonna seem a little bit silly. And frankly, they are a little bit silly and hopefully they continue to be a little bit silly. But these sort of typify the issue that people have with AI. Okay, so this is funny. This is again, somebody else's. This is I believe the Grand Theft Auto 5 engine, which is one of the ways in which self-driving cars have been trained since it's so fully featured. I think this is just rendering, but since it's so fully featured, you can train a self-driving car by having virtual cameras on it and putting it into one of the modern, very broad, very fully featured game environments. And then doing things like rewarding, getting component at point B, but penalizing, hitting anything. And you can train actually a reasonable algorithm for driving that way, just in a virtual environment. And then on the real road, there are sort of emergent behaviors that make the real world a bit more difficult. But let's say that I take myself to a certain level of self-driving car and this is what I figure out to reward. Remember, it's like all we can really do is structure a very general algorithm and then train it to optimize some things. And those define its behavior. Not our choices about its behavior. We just have to say this is what we... So the first thing that you would wanna do is say, all right, car, I don't want you to, I don't want you to hit anything. So if you hit anything, then your negative penalty will be like infinite. And the first thing that will happen then is that car won't move at all because the ratio of infinity to whatever you were giving it to get from point A to point B is such that even an infinitesimal fraction of the possibility of hitting someone still outweighs it. So now we have a problem because now in order to train my car, I have to come up with a numerical difference between the value of getting there in time and the value of not hitting a person. And intuitively, and I guess from an intuitive ethical perspective, we want that to be an infinite ratio. I wanna say, no matter what, never hit a person. And that's the exact sort of thing that you would want to say to sort of a 1950s style AI. You go, never hit a person no matter what, never make that choice. But in the modern era, we can't do that because if we say never, we put a zero or an infinity in the thing and the method no longer trains because the only thing that we're basically training it to is its reward function. So now that we, this is by the way, the first big legal thing that I think needs to be is somewhat above my pay grade and it needs to be discussed because you can't tell the vehicle, just don't hit anybody. You have to tell it, this is how bad it is to hit somebody and that has to be a non-infinite number. So you basically tune it and tell it at least tries to go from point A to point B. And then you might tell it, well, get to your destination on time. That's important but not as important as not hitting anybody. So it's a badness of a thousand if you hit somebody, it's goodness of 10 if you get to your destination on time and then maybe it's a goodness of one or two if you live as long as you can. So you optimize your route and don't drive way out of your way and all the other thing. But let's say that we just did that and didn't add anything else to the fitness function. We then get a car that could figure out that basically it lives as long as it can, as long as it's not being driven. So for example, one thing that it could figure out and this is, again, it seems somewhat unreasonable but this is very much the sort of thing that could happen is that if you had a very full-fledged simulation it might figure out that well, if I drive my people off a cliff but it's a cliff that I do not die from hitting the bottom in, then functionally no one ever drives me again and I live forever. I know, it sounds silly but this is exactly the kind of strange edge case and merchant behavior that we have to watch out for. I put this forward exactly so that it doesn't happen. So that people are thinking along this direction. So if I were to use these very simple, very reasonable rules there's a possibility that in addition with a very fully fledged simulation I would get cars that figure out if they just go off a medium-sized cliff and kill everybody in them that they quote unquote live forever because I've defined living forever as not having a transmission go out for example. So, first problem, equally silly is this kind of thing. I could imagine taking a robot and saying, here's a lawn mower, here's a bunch of blades, here's a burner thing to burn weeds and your job is to make sure that you pull out all the weeds in this neighborhood or something. However, in the modern AI sense I would have to say I can't really control what you do very well but you get rewarded the least number of weeds that are out there. As the number of weeds drops, that's your goal. Which normally would mean go around and pull out all the weeds but it could also mean go and threaten people or attack people because they will figure out that they need to pull out the weeds or else you will do so. So now I have a system with two emergent behaviors and I kind of hope I get the right one. This is the second thing that I think the law is going to have to handle and probably by careful test cases which we'll get to in a minute. But again, this is a silly scenario but consider this as kind of an examination of what could happen. I take an autonomous robot and say, you have the potential to do whatever you want but optimize this outcome. And there's two solid outcomes. One, every time it sees a weed it picks it up and it gets rewarded in its internal mechanism. Two, whenever it sees a weed it goes and attacks people and then people would figure out like, oh my gosh we have to make sure there's no weeds or else the robot will come and attack us. And both of those are sort of emergent behavior spaces that live in this kind of highly complex latent space inside the machine learning algorithm inside the robot. So we can't really demand easily unless we're very careful that the robot trains to one or other of those behaviors but we might be able to train a bunch of algorithms and see which one seems to behave the way that we want. And so that's the kind of thing that we're gonna have to handle via standards. Right now as I'll get to in the end it's kind of the Wild West. There isn't a lot of oversight over this. So we rely on hopefully the expertise of people who think these things through if they're gonna do something like an autonomous system. But again, there's emergent behaviors and it's very hard to know that you've seen every edge case when you're just watching an autonomous system though. This one is not mine but it's from somebody named Nick Foster who was talking about ethical issues in 2003 and came up with this idea of the paperclip maximizer. So suppose I have an AI and I say, hey, you wanna make a lot of paperclips because I run a paperclip factory. But it has in some sense, unlimited resources. It can go ahead and pull more things and then do more things I just programmed it to make all the paperclips it can rather than to make paperclips to a certain amount or to make paperclips with a certain efficiency or to make paperclips and et cetera. And hypothetically, if this were the case then this could run to infinity and just say, I don't care about anything I don't have any concept of the outside world I'm just going to make paperclips for infinity and kill people to do it and tear down buildings and so forth. It's a little bit of an ambulance concept but this is kind of the third legal issue is that there needs to be some limitation on the scope of operations. And I think that that's going to be a strong point that needs to be hammered out in the near future. So if you have an autonomous system in some sense it should live in some kind of a bounding box and we should establish by either end-to-end tests or by other sort of structural analysis of the output that that bounding boxes is solid to it. Kind of sort of work a way around it. And normally that's done with very careful application of the fitness function, the thing that you're trying to maximize. But in this case, if you just maximize paperclips with no other regard then the bad things can happen. Finally, and this is going to blend into the talk about ethics and bias in AI is that I could just take a robot for example and say, have as many people as possible like you or get on Twitter, which is going to be very similar to what I'll talk about in a moment, get on Twitter and sound like everybody else and get as many followers as you can. And that we don't really know the full ramifications of what would happen. We know that people are very influenceable by social media and social structures around them. So if I were to make a robot that could go and talk to people and be online and be an online presence, there's every possibility that gradually optimizing the number of people who cared about it or did what it said or responded to it or whatever, it would actually end up in a place that was very harmful rather than being innocuous. So how does this affect business, present and future? Basically to cover this, a lot of tasks that can be described as like given it puts a do action or make assessment be can be done whole by algorithms. I'm only touching this a little bit, but I'll come back to it because of this, a lot of jobs are sort of on the block. A lot of people who are drivers, people who are whatever accountancy and there's a lot of things that are being in principle replaced. A lot of them are quietly happening right now. Software as a service and so forth end up improving human productivity, especially at repetitive tasks. But physical and general robotics as well as those sorts of online systems might be our near future. So one thing I wanna talk about, I'll come back to this is but what will we do if productivity explodes but employment shrinks? I don't honestly know, this is a big problem we're trying to work at it in a professional environment, we're trying to make sure that we also add jobs that we also ensure that there are additional things that people can do or that we move forward and we sort of create more powers for humanity and thus more industries instead of just reducing the ones that we have. But what about ethical actions taken by AI systems? Well, if it's an autonomous system, this is the take home message. Instead of, if X then do Y, we just have to program them such as, well maximize things Z and possibly things Z one, two and three and that can lead to emergent behaviors that are really weird. And some of them are great, they play games better than we do even very, very complex games. And some of them are potentially very difficult. If I wanna make a car that drives itself, it's impossible to put it in front of literally every scenario you could see. And so we're wrestling with the problem of like how good is good and what because no one, there are no simple set of rules. We've just trained a thing that mostly does the right thing according to us. So we're gonna need some kind of framework to think about ethical outcomes with these systems to think about how we limit them and so forth. And I'm gonna swing back to that at the end, but also, did I get a question? Sorry, I was unsure. Nope, we're good. So now, unfortunately it was a little bit fun to think about the AI apocalypse because it seems distant, it seems like a thing that's gonna be 20 years off and we can work on defining the problem. I think that might be true, but there's some way less fun problems that are happening right now. So the sort of more depressing thing that needs to be addressed quite right away is the real world bias that already exists. So moving away from autonomous systems and back to machine learning, machine learning by itself drives a lot of things. Who do I give a loan to? Is this person the right person to come in this building? Who's likely to, I don't know, who's likely to commit crime or who's likely to pay off their house or get a disease or anything. And in principle, those systems are very powerful and they can and should be used to help a lot of folks and to give us a lot of flexibility and an ability to do things that would otherwise require a lot of human labor or might not even be able to do. But in that, we have a bunch of risk. So here are a couple of headlines and structures that there are things that have happened basically in the contemporary world that have really given people a lot of shock and this is, I'll probably go into more examples here by saying, well, you know, wait a minute, how in the world can an algorithm be biased? Now, you guys are probably laughing because you're very familiar with this, that an algorithm can be biased depending on how it's written and depending on what the outcome is. But I wanna say maybe five years ago, a lot of people would have laughed at this. It's just, it's an algorithm, you put into it data and you ask it to do a thing and that thing is sort of ethically neutral and it comes out with the best answer and the best answer is the best answer. Now, people have sort of become much more savvy but I think partially because of these stories that have hit the news. So in order to try and look around this problem, I'm gonna say, talk about the ways that an algorithm can be biased. As people talk about the output, like the thing that happened, the negative event, but it's often a little bit more opaque to talk about why they happened and what we can do about them because some of them are easily fixable. Some of them are quite a bit more difficult and I'm gonna lay out four here. The first one is, first and probably most obvious one is that your example behavioral data or your example data or whatever can just have a strong bias in it. Remember that one of the ways to put supervised machine learning if we're talking about taking actions or guessing a thing is, well, what would a person say in this case or what would a person do in this case? And so very obviously, well, what if people are biased as well and you can imagine in scenarios like this where n is a very large number, I can just say, oh, well, who should I arrest? Oh, I'll just go look at the police and I'll figure out who they did arrest and then I'll do that because that's right. Well, is it, it's not necessary that your data is free of bias and if a population has bias, then your data is certainly not free. It will reflect that very crisply. An example of this, a little humorous but there are way less humorous ones is that in 2016, Microsoft attempted to launch a Twitter bot called Tay.ai and Tay was an algorithm designed to listen to people on Twitter and sort of scoop up all the things that they say to her and repeat them and sort of remix them and so forth. It was within 24 hours. Now, partially you can blame the internet troll community because people were doing some of this on purpose, sending messages like this but she was saying all of these awful things and they had to shut it down and I think in retrospect that that should have been obvious but if I have, if I put something out there and I say, hey, act like a human and then the humans that that thing is exposed to or the data is exposed to are awful, then it will be similarly awful and one possible solution here is to change the success metrics to basically be smarter about what we say, how we say to behave to not just say, hey, look at all the data and figure out what a human would do in this situation or figure out what was done in the situation or figure out what will win in the situation but instead add other terms to that reward function and say, yeah, you do wanna sound like other people but we also want you to either avoid these things or to use these other metrics. Basically the same thing that you would imagine, I guess I suppose teaching a kid, you say, well, all right, you're basically going to learn to speak like everyone around you but don't say these things or these topics will hurt someone, that kind of thing. It's an open-ended answer because it's kind of an open-ended question but this is one possible solution and this is one of the big sources of problem. You get biased because your data is biased and if you don't know that then your algorithm looks exactly like your data and it does in some sense the exact correct thing which is to give you a biased answer. So data, sort of garbage and garbage out. That's one of the things that happened but there's more. One of the sort of famously happened and I wanna say this is, trying to remember whose algorithm this was. A couple of them I think, several people had this problem where they made facial recognition algorithms and then facial recognition algorithms did not do equally well across groups and to wit, this is one of I think the easier ones to deal with because it's easily detectable and in some sense it's not too difficult to fix. Imagine that I make a camera algorithm and this camera algorithm is designed to figure out the closest face, the largest face in the image and re-center the picture on that assuming that that's what I want. I'm taking a picture of people and someone's kind of out front, it's probably me or whatever so I'm gonna make this algorithm to do it and in order to do that I have to train the algorithm to figure out what a face is and what a face is not. So in order to do that I take all the pictures that I know have people in them and make a little training set where I say here's a face and here's a face and here's a face and here's not a face. And if in doing so, but basically what happens in the guts of the algorithm is there ends up being sort of a limited what you'd call latent feature space where all faces differ but they don't differ by that much, almost everybody has two eyes and a mouth and all of that. So when you train an algorithm to figure out where faces are you basically end up encoding all of those things. However, we all have different facial spacing slightly. Some of us have eyes or nose or a mouth or hair that's shaped somewhat differently. So if in making my algorithm I have 700 people who maybe look just like me and then 70 people who don't look like me. I can train an algorithm and what will happen to that algorithm is it'll come back and say, yeah, we do really well. Most of those people we found and I may not realize until that is literally deployed that what's going on is it spent all of its sort of latent variable space on the 700 people and it does very terrible job on the 70 people but when I went and checked it was still doing really well on quote the majority of people. Now the nice thing about this one is that there's a simple solution. You just take more data. You take more data of the minority population and if you need to you retweak the way the algorithm is done and it's easy to check but we need to be checking and this is sort of the underlying thing that I think blends into the legal question is if you're going to do something like this and release it to the wild you need to figure out beforehand so this doesn't happen. So you don't release your camera and have it centering on somebody in the background because it can't figure out that the person in the foreground is a face because it wasn't trained very well in their sort of face but it's fixable. You go and figure out, you know from the get go that could be a problem, you check and your algorithm isn't done until that's not the case but there have been several very unfortunate things that basically this has happened but in some sense it's nice because on a technical level this one's fixable. You just got to fix it and we got to find it. Very similar to that even if I'm not talking about something with this complex latent space is the thing that I call the 99% accurate cancer test. So this is a classic rookie mistake by the way like every data scientist makes this at least once in the course of their career is trying to take a data set where one population is dramatically smaller than the other where you have basically a minority group and hearkened back to my earlier point about people who might have cardiovascular disease or something. That's something that like most people don't have on most days like cancer pretty bad and you want to catch it. So I take everybody's stats and the results of tests and everything else and I train a machine learning algorithm to guess do people have cancer and it comes back and it says, yep, this algorithm is 99% accurate and I go 99% accurate. Well, that's great but the problem is that if only 1% of people have cancer and I really want to find it and I continue to train my algorithm until it gets the most answers right and it can do whatever it wants then it will quickly just realize that if it guesses no every single time that that's usually right. And now I have an algorithm that doesn't do anything but I can claim that it's 99% accurate and worse if I'm a bad data scientist and sort of asleep at my desk and not thinking about this, I might not even realize it. I could literally deploy this and kill people because people are now being told that they don't have cancer when in fact they do because I released this crystal ball that doesn't do anything. There's a solution to this but right now again, we are at the whim of people who are hopefully good at their profession in the data science world and the machine learning and artificial intelligence to understand that these are problems and to try and catch them and if they don't bad things can happen. I'm not really certain what the proper sort of legal answer to that is except that we need standards and we need best practices and so on. So the technical solution to this is you just resample the data. Basically, well, there's several of them. The simple one is you just resample it. You say, well, I'm gonna take data and synthesize some that's very like the minority class or I'm gonna under sample the majority class or I'm gonna tweak with the thing that I'm trying to maximize. There's a few ways to do it but it's a well understood answer. But so that's another thing that can happen to you. You have a minority class and your algorithm just freaking ignores it. Like, you know, there's just, but you can fix it. And then finally, and this is one of the hardest ones, the fourth thing that I would put a finger on is that sometimes bias happens because there's underlying population differences. So like say I wanna issue car loans and I'm gonna make an AI system that just gets the best returns. And I have a bunch of data on people. You know, I've got people on Zip Code, I have average income. One or two quick things here, Claudia dropped into the questions which I redistributed out to everybody. The ProPublica study that shows bias in sentencing algorithms. And then the Amazon example for failure for facial recognition for people who wanna see more about those two examples. Thank you so much for sharing those. Oh yeah, absolutely. And those are really good examples. I mean, if you, difference in sending, and I need to go back and look at ProPublica, I think the differences in sentencing might be because the example data is biased. And so it just repeated that. And the facial recognition is very much this thing. It's, you know, well, you know, you can, we could fix it, but we didn't catch it. They just didn't catch it. And so they didn't fix it. But there was a way to catch it and a way to fix it. So I have hope for the future in that regard. I think basically with proper standards, we're gonna be able to catch these things. It's me being an optimist, but it's not just me having to be good at my job. We also definitely need some kind of legal environment in which people know to do this. So, but say I wanna, you know, I wanna issue a car loan, I wanna get the best return. So I have all this data on people and I know everything about them. And let's say that I even know stuff that I wouldn't wanna use. Let's say I know like, you know, race and gender and all the other things that I, you know, you wouldn't wanna put in an algorithm because you'd end up with this obvious bias. But if I do this, let's say that I just optimize returns. And I'll show an example of this in a second. This strongly risks producing a tool with like really implicit bias along features you don't wanna use. You're gonna find that all the people you issued loans to on average are much more of one source because they live in a neighborhood and not of another sort because they live in another neighborhood and people are sort of clustering among neighborhoods. So I could, you know, make a model that just excludes those hot button things and like, you know, no race, no gender, no age maybe. But the power of machine learning is that it's very good at inferring those things. So you end up just forcing that variable into again, what we call the latent space. You know, if I say, I have a really good algorithm and it knows your, all your demographics and where you live and how much you make, and et cetera. And I go, well, well, pull out the demographics. We wanna make it fair. Then what happens is you usually end up with an algorithm that can still have implicit bias because basically it's first inferring those demographics and then still using them. And it's very tricky to figure out what to do. I don't have a full answer to this, but I do have a more of an exploration of this. One possible solution to this is you've got to explicitly evaluate and correct for the bias and how are we gonna do that? So this is, please go to this link by the way. This is a really good from one of the Google research blogs exploration is sort of an interactive exploration of the ways in which we can attempt to make this kind of thing fair. There's, the field is thinking about this issue a lot. I know that yours is as well. There's, so there's several possibilities. Let's say in that same example, I wanna maximize my return for loans. Now I give loans to people, but I know that this is a limited example. They will be extrapolated into more complex data space. But let's say you just have one axis and it's like a whatever credit score or something like that on the X axis there. And then there's two populations and one is blue and one is orange in this example. And some of each of those populations would have paid back alone and some would not. But you don't wanna, it was something that's very biased. So now the fun begins, right? Because, and I've over time heard people that have espoused basically each of these. One, well, how can something impersonal like maximizing profit be biased? I just won't add any explicit bias. I'll take out all the demographic information and boom. There, I get that, that's what I will call fair. Or even I'll leave it in and just say, well, that's my highest likelihood of getting my loan back even though it includes all this demographic information. Well, I'm not trying to be biased. So therefore that's what fair means, not trying to be biased. And someone usually immediately in conversations, this is the next thing that people go to is like, well, we'll just pull that out. Pull the demographic information out. There's no features that are present that should allow this explicit bias. We've nipped the explicit bias in the bud. So that's what I mean by fair. We didn't put any of that stuff that you wanna be biased, don't wanna be biased against. Then a third possibility. Well, you could go for intergroup parity. I could basically say, I'm gonna make a threshold or a model for every group and just say, look, the same fraction of people in one group and the other group are gonna get loans or get positive or be viewed positively and that's what I'll call fair. I'm just gonna go for parity. The Google blog suggests another one which is equal opportunity, which is basically of all the people who would represent a positive measurement the same ratio or estimated to be positive. So it's like you're basically equalizing the true positive rates. I say, if you would really pay that loan back, you have the same chance of getting the loan regardless of what group you're in. That's another way to look at it. Others are possible too. You could equalize false positives, for example, which would say, if you're not gonna pay your loan back, some of you will still get loans even if you're not gonna pay it back in the end. And we wanna make sure that one of the groups is not overly advantaged in terms of getting loans when they kind of shouldn't get them. So that's another way to go. So there are at least like five possible things that different people at different times will call fair. And I say this not to muddy the waters too much, but to say that all of these are very implementable. I mean, the example in the blog is really nice. And if you go and take a look at it and play with the thresholds, it's a fun little way to take a look. But all of these are eminently implementable. You can make algorithms that do any one of these things and someone in the world will say that, well, that's what fair means. So that we need essentially best practices and standards to figure out conceptually, what fair we're gonna use when and why. We basically need that kind of philosophical and legal rhetoric around it. And I see it now being built, which is nice. So what are we gonna do about it? Even now, Fias is super possible. We try and be careful of some of that. Some of that revolves around good data scientists doing a good job of being data scientists and being careful. Some of it revolves around business leaders and others asking for the right thing and being cognizant of the places where bias can creep in. We should probably consider things like tests for differential treatment with broad use cases. If I have a large population of people and I try it on a large and diverse population of people, can I verify what the differential impact looks like rather than sort of waiting for my algorithm to go into the wild and then just waiting for some negative impact is the field needs some consensus on like which fairness metric we're using and when or are there an envelope of them that we're sort of attempting and what happens if that ends up not being the case. So what are we gonna do about emerging behaviors and dangerous AI? Well, that's again, like I said, more of the fun question but it still sort of keeps me up at night. We do need some kind of standards organization. We need some guidance that would drive legislation on what we're gonna do about strong AI as it gets stronger. First of all, in my personal opinion, the things that would precipitate these scenarios are as follows. One, if we develop self-referential reasoning that is in which the value function is sort of tied to the health of the robot itself. Normally, this is not the case. If I have a robot and I want it to move objects, it knows that moving objects is best but it doesn't necessarily have a concept of its own existence. I think that's one of the things that people are tinkering with that I would consider somewhat dangerous. Also the development of recursive self-programming. So normally that machine learning algorithm that lives in the guts of the robot in the head of the robot just has a bunch of parameters that it can tune. And there's static code around that that says things like here's your reward function. Here's the way in which you're going to explore the world. Here's the times when you'll use your best guess as to what's good. And here's the times where you do a random thing to fill out your uncertainty. We haven't really let AI systems muck about in that static code. And I believe that's another thing that might pre-curse some negative outcomes. That's again kind of my own opinion. So there's some other stuff but what should we do? Well, there's possibility that we have this AI runaway or we make a bunch of cars and they drive us around for three years and then on some lucky Tuesday they all figure out to drive themselves into the ocean or something with us in them. How do we stop that? Well, there's some things. One is this idea, these are mostly my own opinion and others may disagree but one checkpointing where places where human input is periodically required to continue training and learning. So I'm training an algorithm and rather than just training in the field at some point to implement what I have learned or what I've changed about the algorithm I need to stop and have a human being look at the process and make sure that we can still check off. Structure limiting that might disallow the model past a certain level of complexity and by that I'm referring to like self-referentiality in some of the stuff on the previous slide. This idea of algorithm isolation where if a robot is undergoing a learning process it can't share the information directly with the other robots without human intervention. This idea of again learning in the field I think is one of the more dangerous ones that if you allow, if any robotic system has multiple places in which its behavior can lie like in the case of the weed eater robot and there's sort of a great one and a not so good one the possibility of going from the right one to a not so good one in the field is there. And I think we can reduce that very much by instead saying we're not allowed to train as the thing operates. We can take data as the thing operates and we can train later as a batch and then we re-upload registered and understood AI algorithms. Yeah, the common practice should be I think to develop a standard set of test sets with common and uncommon occurrences. People are doing this with self-driving cars. The idea of saying I'm gonna come up with a bunch of challenging scenarios that I will always demand that my car succeed in and anywhere that we find an outlier that there's a scenario that we didn't see any kind of edge case we add that to the set so that it can't happen again and or things surrounding that. So the idea of checking out an algorithm and saying we're going to retrain this we have more data now it's gonna be better but let's check it with some end to end tests. Let's make sure that it can't just completely learn on the fly while it's operating and let's checkpoint it with sort of human gut checks to make sure that it's not flying off the rail somehow. There's still some other issues as well. If AI solutions are allowed to do jobs that humans do now, who holds the liability? I don't know. We have some standards issues. We need certification. Somebody needs to figure out who owns the liability. I think that unfortunate case with the Uber self-driving car that hit and killed someone I haven't heard the end of that. Maybe somebody can fill me in about what's going on there but as tragic as that is it probably will drive us in the direction of figuring out where the liability lies because right now I don't have a solid answer of where that is. Standardized tests for performance especially in AI that's gonna operate in the field but even for machine learning algorithms that are gonna make choices for people. What in the world happens how we remediate the situation if accidents do happen? How good an AI system has to be in a situation where accidents are possible to minimize those. Does it have to be better than humans? If so, by how much? These things just need to be defined. And even if humanity is safe, how do we distribute resources in the world where cost of goods and service plummets that's maybe slightly outside this talk but it's something that's on my mind. So keep that in the back of yours as well. There's the possibility that we by doing these we end up driving human productivity up which sounds like a good thing but there's also the possibility to reduce the necessity of people in industry and perhaps we can collectively create some kind of world where we don't sort of need to work as much as we do now. So as a legal and tech community we kind of need to get out in front of this and right now it's sort of the wild west. Very little of this is mandatory. Very little of this is absolute standard practice. Attempts of compliance with these ideas is sort of starts to be on the individual's implement. And I see some of the, especially some of the larger employers start to develop policy. I think people in the legal and philosophical space should definitely be involved in those conversations and maybe even in the middle. Cases like the Uber car accident will be more important. We got to define liability and we can expect the next decade to like necessitate a lot of these things. I can't imagine we go another 10 years without a sort of real regulatory body that everybody knows about. And we should brace for some of these issues, expect a couple of significant legal cases and just get ahead of it wherever we can both in terms of bias and in terms of strong AI to head off any kind of potentially dangerous fallout. So that's all I have. I'm happy to take any other questions and just thanks for hanging with me you guys. So two quick things, number one, there are two major initiatives around the access to justice technology side of these particular questions. One of them, Avachi along with, I believe it was Anasteel and others came up with some principles for especially AI algorithms in this space that have been kind of distributed nationally. I'll get links to that and add it. And then here in Washington state, the ability to appeal decisions and audit what goes on with machine learning is something that is being looked at at any court rule or actually it's in the stages of being principles at this point. It was a court order that is being considered here in Washington state. It looks like Claudia's got a few questions also. Claudia, you are still unmuted so please feel free to ask them directly. Okay, thank you. I was very interested. Thank you, very helpful and insightful and honest presentation. I wanted to look at the slides where you have the remediations and I don't know if you can answer it but just to get a sense of the scope costs. Yeah, like, I think it's the second to last where you were saying these are things that we can do. Yeah, like these are things that we can do to minimize harms and really put out a good product or two. And my question is like, because I think that us in legal tech have no idea what these things cost in the for-profit market, like, you know, what does it take from when you have like a gum and tape system like the stuff that you see at the Smithsonian, right? Where you have the first whatever that went to the moon and it's all foil and tape and stuff like that. And then you have the space shuttle. I think that we all wanna get to the space shuttle. But just what is the cost to create some predictive thing like compass or like recognition? How many millions were invested in that? That's one question. And then the other one in terms of remediation or taking the steps like you were saying, go back and sample. I know that for the NIH all of us initiative, that's it's a 130 million genetics collection again because they have realized that 80% of their sample is basically why European genetics. And now that we're moving to cancer treatments and vaccines that react specifically to genetic code, they realize that they're excluding, you know, a significant portion of the population. So just in some of these things, like, what does it cost to do something from gum and tape to actual product? And then in terms of the remediation or suggestions you're making, what is the cost? Like the cost of going back and re-sampling or the cost of who's gonna cover certification? Who's gonna cover standard developing standard test and how much would it cost and how effective would those be in minimizing cost? So I guess I'm gonna take, what I say about cost is gonna mostly be guesses, but I'm gonna make some estimates here. My really quick answer is that the good news is, well, that although usually making one of these algorithms tends to be like relatively expensive in terms of people time, retraining it is not literally, but like almost just the cost of data. So like if I take- So a supercomputer is cheap, two million? Yeah, well- How much does it cost to rent a supercomputer? Like three million at University of Washington? Well, so that's just it. So normally this is measured loosely in like overall processor time and a lot of people now use distributed systems like Amazon web services and so forth to run this stuff. I know that and let me make sure I catch this right. Well, let me answer that in two parts. The first part is what do I do in the first place? If I wanna make like a new facial recognition algorithm and I might have to train it on the data set that has a tremendous number of faces that are called out and like some individual has actually put a dot on that explaining where the faces and so that can be very expensive. The creation of the data set initially can be really expensive and sometimes people use mechanical Turk and other sorts of approaches for this, but I wanna say that's, actually wait, let me do this and try and take a wild guess in my head. The data itself, like if I wanted to start from scratch and I wanted to have millions of faces, I'm gonna be able to do, this is gonna be a person can do maybe a handful of hundreds an hour and then thousands, like 20,000 person hours or something like that, just like 10 person years or time. I think the answer is the data might literally be, the initial data, if we wanted to go from scratch and bear in mind that I keep saying that for a reason, if we wanted to go from scratch, you might be talking about like in excess of a million or $2 and just making the data. If you wanted to literally start from nothing, if you wanted to make facial recognition algorithm, then the training would be probably a substantial fraction of that again. So I think the answer is Euro would be quite... Five million. Maybe a few million, but there's a lot of hope and the reason is because in the very modern era and like the last handful of years, we have now worked very well with this idea of transfer learning and in transfer learning, what you do is you basically take a machine learning algorithm that's been trained on a dataset. Let's say you've trained it on a dataset and the problem is that it's mostly like Caucasian faces and it's padded other faces. So you go, okay, well, that sucks. So we need to add lots of other faces and we need to retrain the algorithm. So if you can get past the cost of data, the nice thing is that you can take the old algorithm that isn't too good on everybody, but it's like kind of okay, then put a new dataset on it and you don't have to train it for nearly as long because you have something that was kind of like close to a correct answer. And so we have a mechanism by which they can take that close to a correct answer and only train it a little bit so it gets to really right answer. So in modern times, because many of the facial recognition and image analysis and natural language processing and so forth algorithms are relatively well canned up, that the training cost has gone way down in like the last three years. And it might be the cost of a dataset plus like no more than like days on a large computer, and by large I don't mean like super computer, I mean like big computer with a bunch of GPUs. So all of a sudden your training cost goes way down maybe thousands of dollars or something to remediate this. So in some sense, the real problem is that these efforts will then live and die on data. Let's say that I did this and I was a camera manufacturer to keep picking on this scenario and as a camera manufacturer I make this dataset and it turns out that it's just everybody from my home town nearby. So they all look kind of like me and it's not good enough and I need to fix it. Then where I can see the legal pullback from them would be how in the world am I going to go and make and expand a dataset and that's where all the costs are. So it seems like there's two things going on here. One is that there needs to be some type of transparency with regards to the datasets and there is a public positive to having open data and open standards to start being able to identify where problems or areas for improvement are. The second thing is that the technology is moving very quickly to where the cost is going down every few months. What we're talking about at millions and hundreds of thousands of dollars today three years from now may be a very small fraction of that. That it's not just Moore's law but that the algorithm improvement process is getting more efficient also. Yeah, well said. That is absolutely the case. And the transparency needed I wonder if there's a place and again I don't know who pays for it but certainly this idea of saying well here's a basis set there's a lot of different people in it and either they're synthesized or they're real but there's a lot of different people with a lot of different statistics and a lot of different faces and a lot of different bank accounts or whatever we know and so you have this algorithm it better basically be equitable across all these people but we don't really have universal datasets like that so people are just kind of doing their best at the moment. I don't know what the outcome will be of that. Yeah, it seems like there's a need for a kind of nonprofit standards agency that could do testing and certification and give feedback that would house some of these datasets and then put together reports as these algorithms are being used because without the auditing process that we wait for harm to show up in tens of thousands of people getting denied unemployment claims or something because an algorithm was optimized in order to save the state money not to actually validate whether it was useful or whether the claim was valid so I think there is a strong need for that creation and some of the cost of that could be bared as a positive that companies or individuals competing in that space would pay for some of the certification process. Yeah, that sounds great to me certainly that this idea that there needs to be sufficient transparency in terms of algorithm structure things end up being technical but there's technical people in the community who also want to dive into it and if I said exactly that if I said look, here's a definition of fairness I'm using and I'm either using this fitness function for the algorithm this is what I'm training to and then I'm penalizing it this way or I'm using this but I'm vetoing that or whatever that you sort of do for that to be part of the specification of a product and published somewhere I think is probably a good step. Yeah, I like that transparency in publishing and that could be written in on a contractual level to grants or to other funding that is available through states or federally. So, I wanted to say Oh, please. Just sorry, just one more comment I think that the evaluation of the harm right because this is all going to eventually end up in litigation and all of that the liability questions all of those are super important but before you get to that the most important thing will be to make sure that you have experts on the communities that are going to be using the tool making the harm assessments and really working closely with the data scientists because if you don't have diversity and a variety of life experiences at that level or people who have worked with the system oftentimes representing people to the system you're going to end up making a lot of assumptions and that will end up definitely harming people so I think that inclusion and diversity at the highest level at the decision making part when the data and the harms and the assumptions and the training is happening but really at the design level it's key otherwise you're going to have a majority group making decisions for technology that would impact in a context that is a historical context that would really harm people and so eventually either the tools will be catastrophic or they will not be legitimate and so the investment whether it's 10 million, 5 million a million in 5 years or whatever will be all for naught and then how it is with technology all of us who work on it know is if one thing doesn't go well then people don't ever want to take the chance again it's not like dating where you go out on another date with technology you get sometimes one chance so you have to get it out with credibility from the first day I think a lot of the concern here also comes in choice because I believe that a lot of these things are going to be adopted in efforts to save money to automate and consumers or clients may not even know that these decision making algorithms are being used in the background and without proper disclosure and options then people won't even be aware that there could be these type of side effects I wanted to say thank you so much for everybody that came out we will have the recording and the slides available for this thank you everyone for coming in speaking to us we should have a short survey that is automatically sent out after this we are setting our webinar calendar for this year here hopefully within about the next two weeks there will be a larger survey coming out later today asking what topics people would like to see if this is something that people would like to have either informal discussions over or more formalized trainings around please let us know in both that survey after this webinar and then in the general one as we are setting up our programming now and thank you Claudia also for the wonderful comments I like the discussion aspect of that thank you Greco thank you