Recorded on December 2, 2025, this video features a talk by Maximilian Kasy, Professor of Economics at the University of Oxford, presenting his book The Means of Prediction: How AI Really Works (and Who Benefits).

This talk was part of a symposium series presented by the UC Berkeley Computational Research for Equity in the Legal System Training Program (CRELS), which trains doctoral students representing a variety of degree programs and expertise areas in the social sciences, computer science, and statistics.

The talk was co-sponsored by Social Science Matrix, the Berkeley Economy and Society Initiative (BESI) Tech Cluster, the Berkeley Institute for Data Science (BIDS), and the UC Berkeley Department of Economics.

About the Book

AI is inescapable, from its mundane uses online to its increasingly consequential decision-making in courtrooms, job interviews, and wars. The ubiquity of AI is so great that it might produce public resignation — a sense that the technology is our shared fate.

As economist Maximilian Kasy shows in The Means of Prediction, artificial intelligence, far from being an unstoppable force, is irrevocably shaped by human decisions — choices made to date by the ownership class that steers its development and deployment. Kasy shows that the technology of AI is ultimately not that complex. It is insidious, however, in its capacity to steer results to its owners’ wants and ends.

Kasy clearly and accessibly explains the fundamental principles on which AI works, and, in doing so, reveals that the real conflict isn’t between humans and machines, but between those who control the machines and the rest of us.

The Means of Prediction offers a powerful vision of the future of AI: a future not shaped by technology, but by the technology’s owners. Amid a deluge of debates about technical details, new possibilities, and social problems, Kasy cuts to the core issue: Who controls AI’s objectives, and how is this control maintained? The answer lies in what he calls “the means of prediction,” or the essential resources required for building AI systems: data, computing power, expertise, and energy.

As Kasy shows, in a world already defined by inequality, one of humanity’s most consequential technologies has been and will be steered by those already in power. Against those stakes, Kasy offers an elegant framework both for understanding AI’s capabilities and for designing its public control. He makes a compelling case for democratic control over AI objectives as the answer to mounting concerns about AI’s risks and harms.

The Means of Prediction is a revelation, both an expert undressing of a technology that has masqueraded as more complicated and a compelling call for public oversight of this transformative technology.

Maximilian Kasy received his PhD at UC Berkeley and joined Oxford after appointments at UCLA and Harvard University. His current research interests focus on social foundations for statistics and machine learning, going beyond traditional single-agent decision theory. He also works on economic inequality, job guarantee programs, and basic income. Kasy teaches a course on foundations of machine learning at the economics department at Oxford. Learn more at his website.

Podcast and Transcript

(upbeat electronic music)

[WOMAN’S VOICE] The Matrix Podcast is a production of Social Science Matrix, an interdisciplinary research center at the University of California, Berkeley.

[MARION FOURCADE]: Hello, everyone. It is 4:10, so per Berkeley time, we will begin. So thank you for joining us today.

My name is Marion Fourcade. I’m the Director of Social Science Matrix and one of the faculty affiliates of the CRELS program, the, which is the main sponsors of today’s event. The UC Berkeley training program on computational research for equity in the legal system, CRELS, brings together doctoral students from across the social sciences, computer science, and statistics to explore issues at the intersection of computation and equity.

So we are delighted today to welcome Maximilian Kasy, professor of economics at the University of Oxford, who will discuss his new book, The Means of Prediction: How AI Really Works and Who Benefits. In this book, Kasy offers a very accessible yet incisive account of artificial intelligece, intelligence. Rather than framing AI as an unstoppable force, he invites us to consider the power structures and ownership arrangements that determine what AI systems do, whom they serve, and how their objectives might be governed democratically.

And please note that today’s event is co-sponsored by the Berkeley Economy and Society Initiative, especially, specifically the Tech Cluster and also by the Berkeley Department of Economics, where Max got his PhD. But before I turn it over to our guest, let me just mention the final event at Social Science Matrix this semester. It will, oh, we don’t have the–

Okay, this is the lineup for next, for next spring. But the, we have a final event this week. On Thursday, we will welcome Alexis Madrigal a journalist who is very well known to many of you through KQED and The Atlantic.

You know, he was the author of the COVID Tracker and did many really important projects. And he will give a, a Matrix lecture on Oakland that is entitled “To Know A Place,” and that’s on Thursday at 4:00 PM. And then you can see here the lineup for next spring.

We have a whole bunch of events. I especially want to spotlight our, you know, our Matrix on Point on corruption in America and on higher education under attack. Let me now introduce our speaker briefly.

So Max Kasy received his PhD here at Berkeley and joined Oxford University after appointments at UCLA and Harvard University. His current research focuses onsocial, the social foundations for statistics and machine learning going beyond traditional single agent decision theory. He also works on economic inequality, job guarantee programs, and basic income.

He does, you know, there’s a very wide range of topics and he teaches a course on foundations of machine learning at the economics department at Oxford. So without further ado please welcome Max for this lecture.

(audience applauding)

[MAXIMILIAN KASY]

Thank you, Marion. Thanks for having me, and thanks everyone for coming today. I am going to tell you today about my new book, The Means of Prediction, which came out a few weeks ago with University of Chicago Press.

And the book, as you heard, is called The Means of Prediction: How AI Really Works, so the “Really” part was inserted by my publisher, “And Who Benefits.” And all right, so to get things started, like if you’ve read any news in the last few years and especially any news related to AI, chances are you will have come across a story that, that goes something like, like this one here. The, the idea of this story is that we are close to the point of reaching artificial general intelligence or superhuman AI.

And in that point, AI will be better than humans at improving AI itself, and so you get this exponential explosion of AI capabilities. And that will maybe be nice in the short run, but then will be a major problem because the objectives of AI might be different from those from the humans. And so humans will want to switch off the AI at some point and because existence is a precondition of fulfilling its goals, the AI has to eliminate humanity first.

Right, so this is, this is a story that Silicon Valley loves. This is a story that Hollywood loves. I’ve counted, and I’ve found I think around in 40 Hollywood movies that all have the same plot line, going back to the ’60s with Space Odyssey or Terminator in the ’80s or more recently, Ex Machina, or The Matrix, or you name it.

It’s always this story of man versus machine. So it’s usually a man where there’s like one heroic actor that somehow in a battle for life or death with a machine. And going even going further back, beyond like the more recent age of AI, it’s, this story has been told many, many times.

So just to give you one example, there was this Disney clip from the ’40s called Fantasia, where you have like Mickey Mouse who’s, who’s a magician’s apprentice and wants the, wants the broom to carry, carry water, and do his chores because he doesn’t want to do his chores, and then falls asleep, and then the broom dutifully just keeps carrying water until Mi– Mickey almost drowns and has to be saved by the master magician. Right? So it’s again, pretty much exactly the same story.

You could tell many, many other versions. And so these stories clearly resonate, right? These stories, and clearly touch on some of our deepest fears, right?

So, it really goes to the core of our existence. Starts with, like, automation costing our livelihoods, continues to being ruled by these inscrutable forces, losing our autonomy, and ultimately becomes a matter of life and death, or even the survival of humanity. And all of that happens due to some incomprehensible, inscrutable, almost divine forces that are coming at us, right?

So, clearly, that story does resonate on some level. I don’t think it’s a good story in terms of enabling decisions about AI for kind of two broad sets of reasons. One is it’s kind of a story that’s happening to us, right?

The machine is somehow self-improving and then eliminating us. There’s no human actors in that story. It’s completely hiding that there’s like a million decisions to be made along the way of what technology to build, to what users to put it, when to turn it off or not, and so on.

And so it replaces human choices and agency by a story that’s just happening to us. And second, in this story, there’s never, like, a conflict within society about what to do with this, technology. It’s always, like, the heroic human actor fighting with the machine, as opposed to talking about the different parts of society having different interests, values and so on that impact what happens with the technology.

And so I think it’s probably for these two sets of reasons that the tech industry also loves to tell this story, both in kind of the boomer or doomer variety of the story, that are actually surprisingly close to each other. This is not a story I want to tell in this book. Um, instead, I’m trying to, to make a different argument that runs through the entire book, and that argument goes broadly as follows.

First, I mean, I have to start with the question, what is AI? Again, a very, very mystified question, right? What is intelligence?

Maybe what is consciousness? And so on. What is general intelligence?

Who knows? If you look a little bit under the hood of what happens in AI machine learning, it’s a lot more mundane at the end of the day, and 99% of AI is based on optimization algorithms. What does that mean?

Optimization basically means there’s some numerical measure of success. Can be very different across different contexts. And the machine tries to make that measure either as large as possible or as small as possible.

So that’s called either a reward function or a loss function in machine learning and statistics, right? And that reward function can be anything from the number of times you click on an ad to the number of people deported by ICE to the probability of correctly predicting the next word on the internet. Those are all the different rewards that you can put into these optimization algorithms.

Now, that’s, of course, not, no great news to computer scientists. Computer scientists, engineers more broadly, are very good at optimization. That’s what they spend two-thirds of their time on.

They have come up with all kinds of ingenious ways of running optimization efficiently in terms of compute time, in terms of memory used, and so on. They are very good at diagnosing optimization failures figuring out when something went wrong in the optimization and coming up with fixes and so on. So from, coming from that perspective, it’s kind of natural that if there’s seems to be some problem with a technology like AI, they gravitate to thinking of it as something went wrong with the optimization and we have to fix the optimization failure.

But which is kind of the academic counterpart to the man versus machine story where I don’t know, if you go back to something like say Space Odyssey. Like here, there is the board computer on the space mission and that has some programmed objective that the astronauts don’t know, and the astronauts want to switch it off, and so it has to kill the astronauts. But something went wrong in the optimization there.

I don’t think that’s the key issue in most socially consequential applications of AI. I don’t think that’s what most conflicts around AI are about, most debates. I think instead the key issue is almost always not did we fail to optimize, but what did we choose to optimize and who got to choose what’s being optimized?

So that’s a very different question. It’s not that the optimization went wrong, it’s what objective function did we pick for the optimization? And closely related, who got to pick the objective function?

Now, we can talk about what the objective function should be, and I spend some time in the book on that. But there’s the counterpart to that is like, who actually gets to pick the objective function? And given how our society is arranged, it’s typically those who control the inputs into AI, The, what I, what the title of the book actually comes from, The Means of Prediction.

The key ingredients that you need in order to build modern machine learning-based, statistics-based artificial intelligence. And those are first and foremost training data, Data on which the algorithms can learn things. Second, the compute, the processors, GPUs, data centers cloud computing networks that are used to both train and then run those models.

And a bit more secondarily, the expertise that’s needed in order to develop and improve those algorithms. And also crucially, the energy needed to run those data centers, So, and all of this happens at a massive scale at this point. So if you just think of the compute, this year alone the big tech companies are investing around $400 billion only essentially in buying processing units and putting them in data centers.

For next year, another $400 billion are budgeted. So these are just massive orders of magnitude, and similarly for the training data where at this point basically all of the internet is part of the standard training data sets, including things like transcribed YouTube videos and all that. Um yeah.

So that’s the means of prediction. Then I’ll talk more about how, what role they play in, in building AI. Now, if you buy the premise that the key question is what’s being optimized here, and we might have some normative ideas about what should be optimized, and that could be very different from what’s actually being optimized, given who makes the choices, there’s a question of how to close the gap.

And so the ultimate argument that the book makes is that what we need is that those who are being impacted by AI decisions also have a say over what AI is designed to optimize, right? So we need democratic control. And the book tries to sketch a pathway towards such democratic control.

Now, when you hear AI these days, again, like an AI being a mystified notion, it’s also gone through like massive oscillations over time, what people think of when they, when you say AI. Right, so for decades, it was kind of this obscure niche field of computer science for a bunch of nerds. And then there were like some spikes, like in the ’80s, often forgotten.

There was a big AI corporate AI boom with like massive investments into what were called expert systems and similar symbolic AI systems that ended up as a bust because the systems didn’t work out. Then, sometime early in the 2000s, pretty much anything related to data was called AI. Now, public perception has kind of collapsed again and now almost everybody thinks of language models when you hear AI.

What I want to focus on here is, in some sense, intermediate. It’s not everything to be not everything related to data, but it is automated decision-making systems that typically are based on training data, and so decision-making systems that optimize some objective. And so under that header, there’s a lot going on already, and very much implemented in daily life, that is very consequential for those at the receiving end, going beyond the headline-grabbing large language models.

It starts with algorithmic management of gig workers in platforms like Uber and other platforms, where you have algorithms deciding who gets to work when, for what wage, and so on. If you apply for any big corporate job now, chances are, there’s some AI system that’s pre-filtering your CV and your other application materials to decide who even gets invited to a job interview. If you go online to any like search engine or social media platform, basically your entire sense of reality will be filtered by some algorithm that decides what you get to see about the world.

Much of which, of course, is governed by ad targeting, right? So that’s still by far the largest application of AI is maximizing the probability that you’re going to click on some ads and buy something. But it’s not just the commercial corporate applications.

It’s also state actors of various forms, right? So there’s the whole question of predictive policing and incarceration, where algorithms decide where to send police who gets to go free on bail or not. And now our ICE Immigration and Customs Enforcement bought a system from Palantir called Immigration OS, which is essentially targeting their immigration rates to put people in camps and deport them.

Or even one more extreme in the war in Gaza, there were algorithms called Lavender to decide who gets bombed and there is this set data predicting when people are home to decide when they are bombed. So yeah. Obviously, all of those, and there’s increasing severity, very much consequential if you’re at the receiving end of those algorithms.

And so I wanted to keep these, these examples in mind throughout the talk as, as kind of what I’m thinking of here, and not just large language models. All right. So the book has three major parts.

The first part of the book is essentially trying to be a non-technical, non-mathematical explanation of how AI and machine learning work. So that’s the kind of stuff I would teach in our, to our grad students mathematically, but instead, here trying to drop the math and just focus on some of the key conceptual considerations as a precondition for having a broader debate about what we should do with this technology. And I think that’s in particular interesting to learn about what economists would call the production function of AI, again, getting back to the title of the book.

So to figure out how the different inputs of like data, compute, and so on, map into how well these systems perform, and kind of who controls these systems. So that’s the first part. The second part is trying to lay out a framework to think about the political economy of AI.

So, thinking both about the normative side of things, like what should we optimize here, and the positive side of things of who in practice gets to control these means of prediction. What does that imply for who gets the profits, who gets to choose what’s being optimized? Thinking about which actors in society could be agents of change in the sense of possibly closing the gap between what is being optimized, and what we should want to optimize, and the role of ideology in the, in possibly preventing us from doing that.

Right, so that’s the second kind of big picture framework. First technical, then kind of this political economy framework, and then the third and largest part of the book takes these two frameworks, sort of takes them jointly to go over a range of debates that have been happening over the regulation and social impact of AI, things like data privacy and ownership, algorithmic fairness and discrimination, automation in the workplace– explainability of algorithmic decisions and so on, and tries to reframe these debates and put them under this kind of unifying umbrella of the frameworks that I put in the first two parts. But I also think this is also where I hope this book really makes a contribution, whether if there’s a lot of great books out there on AI, both from the technical side and from a more critical side.

What the book, I think, hopefully does is bridging the two and providing kind of a more unified way of thinking about all these problems that goes beyond the more journalistic account of here’s something bad that’s happening, and here’s something else that is bad that’s happening. All right. So, let me tell you a little bit about some parts of this first part of how AI works.

So yeah, again, trying to explain foundations of machine learning and AI without mess. So, what is AI? Again, kind of taking this intermediate perspective between everything created today, though, and just language modeling and kind of taking a more or less standard textbook definition of AI as automated decision-making to maximize some notion of a reward.

And, so, that’s in some sense almost the most important part of this, because it lays out the key framework of what AI is trying to achieve which is there’s some reward, some objective to be maximizing. There’s some space that we’re optimizing over, some set of actions that the algorithm might choose, and there’s some information that the algorithm is basing its decision on, like training data or some kind of prior knowledge that might be built into the system in some way or other. And in a way, like, in order to have a debate about a particular AI system, those are the most important ingredients to understand and to have a discussion about what’s being optimized over what set of possible actions.

Second however, we can go a little bit further under the hood. And AI in principle can be built in different ways. I already mentioned these expert systems that were more prominent in the past, or more broadly what’s called symbolic AI.

But over the last 20 years, the field has been pretty much completely dominated by machine learning, which essentially is AI using statistics, right? It’s more or less automated analysis of data in order to build systems that optimize something. Why has that approach dominated over the last 20 years?

I don’t think it’s because there have been any fundamental conceptual innovations that are all that different from what happened before. So, all the key ideas there have been around for 70 years or longer. But what has really changed at that point is, or like since the 2000s, is the scale of training data that are available and the scale of compute that’s available and affordable, right?

And it’s like once these things reach a certain scale, then things start to work. So the first turning point that was reached there was image recognition, right? So taking images and saying is there a cat in this image or a dog in this image?

You know, almost 20 years ago. It was the first big success of deep learning in particular. Then a bit later, there were the successes of gameplay with like AlphaGo, AlphaZero and more recently language modeling.

But essentially all these breakthroughs just happened at a point when the size of the training data and the available compute got to a point where various things could reach a performance that was satisfactory. And now when is this point reached? This is where a bit of theory about how these learning algorithms work is really helpful, and in particular a bit of theory around how so-called supervised learning or prediction works.

Which supervised learning, the idea is basically you have a bunch of data points that have some label or some outcome associated to them, and then from that you wanna learn to predict the label or outcome from it for a new data point. This could be a label that’s associated with the images. Is this a cat or a dog?

This could be for a sentence, what’s the next word in the sentence given the preceding words and so on. So there’s many different tasks that you can frame in terms of prediction. Any time you face these type of prediction problems, you face a key trade-off between what’s called overfitting and underfitting.

And so, this is like the one part here that I want to do a bit of a deeper dive in a minute, because I think that’s really what tells us about what the role of the different means of prediction is in terms of building AI by these thresholds of AI being workable or AI based on machine learning being workable, are reached, and what the importance of controlling these inputs is. So yeah. Let me jump into that and then talk about the last part.

So, overfitting/underfitting, another name for the same thing is variance and bias. But a key idea is when you make a prediction, often you’re gonna make errors. You want to make these errors as small or infrequent as possible.

Label as many images as correctly as possible, or predict the next word in the sentence as right as often as possible and so on. And or predict like which ad you’re most likely to click on, whichever the task is. And so, you make these prediction errors.

Your goal is to make as few prediction errors as possible. Those prediction errors can come from two broad classes of sources. One is what can be called estimation error, or the variance of the estimator.

And that basically comes just from the fact that there’s some random variation in your training data. And it’s just, basically, that’s a problem that can be solved with more data. The more data you have, the smaller the variance gets.

And so the smaller your estimation errors in your predictions get. But then there’s a second source of error, which is called approximation errors, or bias, or underfitting. And there the problem is basically that your model doesn’t accurately reflect the underlying structure of what you’re trying to model there.

There’s some kind of complexity that your model can’t accurately reflect. And that doesn’t go away with more data. But that potentially does go away by making your model more flexible or more complex.

And so basically anytime you do supervised learning, you face the tension between those two where as you have typically, you have to choose how complex to make your model. And there’s all kinds of ways you can do that. It can be the number of variables that your model has.

It can be how long you let some training algorithm run. Various other ways. It goes under the name of regularization.

But what it does for you is doing a trade-off between underfitting and overfitting, or between variance and bias. And so, that’s kind of reflected in this picture which, without exaggeration, I think this picture kind of captures two-thirds of machine learning theory which is why I want to spend a little bit of time on it. So I mentioned, like, we have this choice of how complex to make our models, right?

Like, how big to make a neural network, how long to let the training algorithm run, and so on. And what we see is, the more complex you make the model, always you’re going to get better at predicting your data in the sample, right? In your training data.

Basically, if you have a bunch of images and a bunch of labels, if you make a model complex enough, you can always exactly correctly classify every image in the set. It might just be for completely spurious reasons, right? Maybe all the images with dogs have a green pixel in row 23 and column 20 and you’re, and the algorithm picks up on that and perfectly predicts in-sample, but becomes completely useless for new images because they don’t have this random pattern in there.

And that’s why that’s where the variance part comes in. And so that’s why the out-of-sample predictive errors, right? The errors for new observations different from your training observations doesn’t always get better as you increase model complexity.

And instead, there’s what’s called the generalization gap, this difference between in-sample and out-of-sample performance. And so any machine learning or supervised learning algorithm has this sweet spot in the middle for model complexity. That sweet spot shifts as you get more data.

Now, why is that relevant for us? It’s, again, it’s relevant because it tells us what the production function of AI is. It tells us how compute and data determine how well you’re going to do.

Because what happens as you get more data here is the gap between those two gets smaller, and kind of the minimum here shifts to the right. So basically, as you get more data, you want to make your models more complex. But if you can’t get more data, it doesn’t really help you to scale up your compute more and more.

There’s a limit. And so this is, this observation is really what’s at the starting point of the boom of AI in the last five years. Because what first researchers at OpenAI and then at the other big tech companies realized is these laws which we know from theory for a long time, they kind of looked empirically for, like, language modeling and for other tasks, how these things scale.

And in particular, they looked at what happens to their prediction errors for language. Say, as they changed the number of training observations, right? The amount of data they put into the whole thing, and what happens as they make the model bigger or smaller.

Basically, as this were more or less computed, the whole thing. And as they did that, they came up with what’s called scaling laws. So OpenAI was in 2020.

This is kind of a, a similar paper two years later from, from Google DeepMind. Very similar idea. They came up with scaling laws that look something like this.

And so this is the one math formula I have in the slides. It’s not in the book, but I couldn’t resist. This is basically the law that told them if we throw more and more compute at the whole thing, which here is reflected in the model size, right?

So that’s how big the neural network gets. We can make our prediction errors smaller and smaller in a predictable way. And so they looked empirically over a range of values for how large the model is, and how much training data to throw at this, how this changes, and then concluded it pays off to just scale up the compute threw everything they had at the compute, forgot about all other approaches, and just made these things go bigger and bigger.

And so this is what we’ve seen in the industry over the last five years. There, here, this is for language models. What you have on the horizontal axis, in both cases, the release date.

On the vertical axis, here you have the data size. And here, you have the model size, so basically, how much compute to throw at the whole thing. But what you should notice is this is a so-called logarithmic scale, meaning as you go from 10 to 100 to 1,000, 10,000, billion training observations, and here the same thing from a billion, 10 billion, and so on, number of parameters in the neural network.

And so you get this more or less linear growth of the last few years, meaning it’s exponentially growing how big these training data sets and models are. What’s interesting though now is, they’ve basically reached the limit in terms of how much training data they can possibly throw at this, because every possible text in the world has already been fed into these things, right? The entire internet is in there, including every transcribed YouTube videos and you name it.

And so to keep scaling the compute now it’s a big bet to think that the scaling will continue and the, at least what statistical theory suggests, this might not work out because basically if it doesn’t really help you to increase your compute or increase your model complexity if you’re to the right of the minimum here. If you can’t scale up your data this isn’t going to get you very far. All right.

So, this is, again, what the industry calls scaling laws, what the economists would call a production function. Basically telling you how well these AI systems work depending how much you can throw at them in terms of inputs of data and compute. And so this one is for language models, but the same thing applies to all kinds of other things and there’s a bunch of things we can take away from that.

One is the potential of this type of technology for different applications. Right? So as I mentioned, like the reason these things have taken off in different domains, I think is largely because they reached a scale where this was viable, where the number of training data in particular, and also the compute was enough to get a reasonable performance.

Worked out for language modeling. There are other domains where there are clearly fundamental limits on how much data we can get relative to how much complexity the underlying problem has. Right?

So to take extreme cases are the same. You wanna predict financial crisis, there’s only one financial crisis of 2008. No amount of machine learning sophistication will allow you to do matrix statistical extrapolation from there.

Or if you take like diagnoses of rare diseases, right? If you only ever have like 200 cases of some rare disease that you want to diagnose based on, say, some imaging method. You know, again, purely statistical learning might not work out if the mapping from say images to the disease is too complex relative to the amount of data that you can have.

Right, so there’s, there’s clearly all kinds of domains where this, this purely learning-based approach cannot work out. There’s others where it has worked out, like gameplay, like language modeling. I think there’s also a lot of interesting more intermediate cases where it’s less obvious.

One interesting example for instance is, is genetics, where in the early 2000s like after this human genome project, there was a lot of excitement about now we are going to be able to predict all kinds of disease risks and cancers and whatnot from, from genetic data hasn’t really materialized. I think there’s good reasons for that which again you can think about in terms of this picture. It’s more or less a matter of counting where even if you sequence the ge– genome of all humans on Earth, and you allow just pairwise interactions between those genes, your, the, the amount of data you would need is very, very larger than what you could possibly get from all of humanity.

So I think there’s good reasons from that purely statistical learning cannot work out in a domain like that. All right, so this is just to give you like a bit of a flavor of why I think it’s interesting to think about the underlying statistics and imlications t has. Supervised learning is not the only thing that happens here.

Now, the key domain is what’s called online learning or adaptive learning. There, the key difference is for supervised learning basically the data are given to the algorithm and then it just learns some patterns from the data. But in many domains, the algorithm makes decisions and then observes consequences of these decisions.

All right, so that gives you a, that adds a key conceptual layer because it basically sees the causal effects of whatever it’s doing in the world and can learn something about these causal effects. And again, like, if you think about ad targeting, right? The algorithm can show you some ads, see who, sees who clicks on them, doesn’t, then from that learns about which ads will be successful in terms of, of maximizing corporate profits, say, right?

And so this introduces a second key trade-off they talk about in the book called the exploration exploitation trade-off, which is basically the tension between experimenting to learn something for the future, that’s the exploration part, or exploiting what you have learned to get kind of good rewards in the, in the moment. All right. Sorry, I should probably move on in the interest of time, but this is just to give you a flavor of some of the key ideas here that I think matter to think about the social consequences.

And so thinking about the scaling loss in production functions of AI in particular it tells us something about the potential for future improvements. It also tells us something about who controls these algorithms by knowing who controls the inputs into these algorithms, right? And so, in particular it’s the data and the compute that matter here.

And we see a lot of contests over property rights going on where they, I mean, creative products of people, intellectual property and so on being appropriated by AI companies. There’s a massive redistribution of the economy happening. Some bigger players can sue and get some money out of it, right?

Like, the New York Times suing OpenAI or something like that and then getting some, I don’t know what the sum was, but some decent amount of money for like them first illegally training their, their language models on the New York Times database. Many smaller actors can’t do the same things. Think of all kinds of like artists, scientists, and so on, all of whose outputs are, or coders for that matter, all of whose outputs are like part of the training data here.

And relatedly, and I’ll get back to that later, there’s also the massive question of externalities, especially when it comes to data related to individuals. So, by externalities economists mean like you do something that has consequences for others essentially that you don’t bear the costs or benefits for. And, that’s kind of core to machine learning, because anytime the data about you are used for training somewhere, that learns patterns across people, so it’s not really about you as the data point.

It’s about everyone else. Which again, has key consequences for how accumulation happens here. And I’ll get back to that later when I talk about data privacy and data ownership.

All right. So, I’m skipping here, like, the normative parts of the paper. The book has a chapter on thinking about kind of how could we possibly think about the, of a framework for what we do want to optimize as a society.

And obviously, political philosophy and economics and other fields have had discussions about this for 4,000 years. But it’s useful to remind ourselves of these discussions in this context. The counterpart to this normative question is like, who can possibly close the gap, right?

So, who can close the gap between what AI is optimizing and what we might think that it should be optimizing? And so, this gets to the question of who can be agents of change, so which individuals or organizations or parts of society have the interest, the values, and the strategic leverage to move things in a direction that might be beneficial here? Now, a lot of discourse around the ethics of AI is essentially, about telling engineers to be nice, right?

And I mean, there’s nothing wrong with engineers being nice, but I think there is strong limits on what that can achieve. Hmm. And kind of, if we go systematically, are these limits to, let’s say, like having like an ethics course in a computer science curriculum, how far that’s going to get us.

I mean, it’s great to have these ethics courses, but I think if you’re working at a tech company, you might be very concerned, say, about what your social media algorithm does to, like, public debate or mental health, and so on. At the end of the day, you’re working in a company that’s living off ad revenues. And so, that’s going to be what you have to maximize.

And so, I think a key call to people in the debate around AI, the social impact of AI, the regulation of AI, is to say that we should address much broader segments of society. And I think there’s all kinds of groups that can at least partially play a role of agents of change, right? So, there’s different kinds of workers that are tied into the tech industry.

There is, there’s of course the, kind of the core tech workers, but then there’s like, all the click workers who are sitting in, in container farms around the world, in places like Kenya or Venezuela or elsewhere, that do a lot of the work of generating the data used to create these AI models. There is the gig workers who are working on platforms governed by algorithms that determine the working conditions, whether you’re driving for Uber or you’re like, in an Amazon warehouse or s– some other place like that. There’s consumers whose reactions to what tech companies do has important influence, if only in the form of marketing or things like that.

But if you’re relying on consumers for your profits, then you’re afraid of having scandals. You’re afraid of having a shit storm around what you’re doing. And so, that imposes some constraints on what companies can do.

But then, more broadly, there’s the question of media and public opinion and parts of the state, different regulators, different components of the law. A lot of existing regulatory instruments, whether it’s competition law, privacy law, intellectual property law and so on, that all can have some bearing here. And that all can help to move the needle in terms of how power is distributed over who gets to choose the objectives of these systems that more and more govern our lives.

And so, I would think of this as kind of the short or medium run answer, is like, thinking about all kinds of different actors in society playing a role in shifting power. I think there’s a normative end point here that I would strongly argue for, which is that those who are being impacted by the decisions of an AI system should have a say over the objectives of that AI system. And since, again, it’s all it’s kind of by construction a collective thing in these algorithms, it has to, it can’t just be individual level.

It has to be a democratic or collective control. And so we have to build some form or other of institutions where people have a say over the technologies that impact them. One thing that might stand in the way of that is ideology.

Now, ideologies is a charged word, obviously. We often just say ideology when we mean the opinions of people we disagree with. But there’s also a more specific and useful meaning, I think, of the term, the way political theory might define it, which is that ideology is a representation of the world that might do one of several things.

It might represent the interest of a particular group in society as the interest of society at large, kind of obfuscating the fact that there is conflict of interest within society. It might take things that are actually choices that, that somebody or some institution is making and that are contingent, meaning that other choices could be made, and pretends that those choices are an objective necessity, kind of in the spirit of, of Thatcher’s famous dictum of, “There is no alternative,” right? There, many times, there is an alternative.

And third ideology often takes social relationships and represents them as technical one and correspondingly presents the solutions as technical ones. And so, I think this general framework of thinking about ideology very much can be useful to think about all these debates about AI, including the ones I started my presentation with, right? This whole like, intelligence explosion, man versus machine story.

Does all of these things, right? The man versus machine story, there’s– It’s the discussion is literally posing in the popular and in the academic version, how do we align the machine objectives with humans, right? They’re never in this description, do we see that there’s different humans who might want different things because of inequality, because of different values, because of whatever it is.

The story of the intelligence explosion is a technical version of this objective necessity, one. So this is just something that’s happening to us at most it can happen to can hope to get a little bit ahead of the curve as opposed to realizing that it’s humans all the way down in some sense. There’s choices to be made at every step of the process.

This is not something that’s just happening to us. There’s a more political version of this inevitability, or no alternative story is the political one or geopolitical one. In the U.S., it’s often something like, “If we don’t do it, China will.”

And therefore, we can’t do anything, we can’t regulate because great power competition. Other parts of the world, say in Europe, it’s often we can’t regulate because we have to catch up with the U.S. Makes it seem like it’s inevitable that those who currently control tech have to be in power because otherwise, somehow, the country is weakened. And there’s this technocratic argument that, it’s all very complicated.

Only experts can understand the idea for democratic governance is impossible. Just leave it to industry to self-regulate. And I think, I mean, all of those is if you listen to what’s the techie CEOs tell politicians, they very much push on all these stories as a way of arguing against any form of regulation or intervention.

And so, I mean, part of what the book tries to really do is make the case it’s actually not that complicated. I’m sure you have to invest a little bit of time to think through it, but the basic ideas are actually fairly broadly accessible and, say, knowing that Facebook maximizes clicks on ads and that it could maximize different things and when it chooses what it shows you. You don’t have to have a computer science degree to understand that, and to– And similarly for other domains.

All right. So that’s kind of broad outline of the political economy framework. And then the last and biggest part of the book is taking these frameworks and applying them to various debates about the social impact of AI and how we might want to regulate AI.

And so in particular, I’m going through like five, five different topics here that have kind of received attention, different, different fields. So some of them more in computer science, some of them more in the law and in regulation. Others more in economics.

But, I, I try to apply this unifying framework if I’ve tried to describe before to, to think through these. And so, yeah, let me go through each of them in turn. Excuse me.

Um, so the first one is this value alignment debate, right, where I’ve, I’ve already talked a little bit about that, but that’s kind of the, the slightly more academic version of, of the man versus machine story, which typically goes something like we have, we have an optimization algorithm. The optimization algorithm is maximizing something, some numerical measure of success. That numerical measure of success might be actually something we care about, but it’s not everything that we care about, and if it’s really good, it’s maximizing this incomplete objective, then things can go really, really wrong.

Right? So I’ve already mentioned the Mickey Mouse story. Another popular version comes from, from this philosopher Bostrom who, who has this paperclip story about the machine that’s programmed to produce as many paperclips as possible.

But then since we don’t need infinite paperclips, we might want to switch it off at some point to prevent that it has to kill everyone. Mm, which actually I mean this is a bit more of an, a site for economists, but there’s a, there’s an interesting parallel here actually between what engineers might call reward designs or picking what the numerical measure of success is and what economists have spent a lot of time thinking about, which is incentive design in, in fields like contract theory or mechanism design, where the question is what Yeah, how, say, to the design pay for employees or something like that.

And in that context, economists have actually also recognized problems that are quite similar to this value alignment argument under headers like multitasking or more specifically things like teaching to the test. And so the teaching to the test story is something like you have standardized tests, and you might use it, use those to reward teachers or select teachers. And those tests might even be perfectly fine in terms of measuring something that is genuinely valuable or part of what you’re trying to achieve.

They just don’t measure everything, but maybe they measure math skills, but they don’t measure the quality of social interactions in the classroom or something like that. But then, if you have high-powered incentives attached to the test, then you might put all the effort there and say the quality of social interactions might really suffer. But it– and I think that’s interesting, and it’s something to carefully think about in differing domains where important outcomes are hard to measure, then we might want to refrain from using high powered AI there in the same way that actually most companies refrain from using high-powered incentives for their employees for exactly that reason, presumably.

But, again, the bigger question is not how to align the algorithm, it’s the human controlling it. So, it’s not that the Facebook algorithm is not exactly achieving what Mark Zuckerberg wants. It’s what’s good for Mark Zuckerberg is maybe not necessarily good for, for everyone else.

Right. So it’s a question of how do you align the objectives of those controlling AI with the objectives of society at large? Um.

All right. Next one, which I think is really fascinating and important here, is the question of privacy and data ownership. Right.

So, I mean, data comes in all kinds of forms and applies to all kinds of things. But in the most socially consequential settings, it’s often, data that applies to individuals, right? So if you think about all this applications I started out with here, right?

Thinking about gig work as a job candidates, about people using social media, about people being incarcerated or bombed. That’s all data about individuals. And so, there’s the question of how do protect this individual’s data?

And that question has received a lot of attention both in the law and in computer science. So, in the law, you have things like in the EU, the General Data Protection Regulation. And I believe California actually has adopted a version of that.

Which more or less is about individual property rights over individual data, right? So you have the right to know what’s being collected about you. You have the right to be forgotten.

You have the right, maybe, to use services without sharing your data. So far, so good. There is kind of a computer science counterpart, which goes under the name of differential privacy, which is basically how can you design mechanisms or algorithms where, for an individual, there is almost no consequences whether or not you share your data.

Right? So, the technical definition is that there’s to be a bound on the likelihood ratio of whatever is coming out of the algorithm we serve with all your data. But really, what it means is it essentially makes no difference whether you share your data or not.

And so now imagine you have those two things together, right? So you have individual property rights, but it doesn’t, you, there’s no consequences if you’re sharing the data. So then you might as well share them if you get some convenience benefit, like some nice online service to use or something like that.

That doesn’t do anything to prevent any harms from AI. It doesn’t do anything to generate any benefits from AI. And the reason is what economists call data externalities.

Right? So another way of putting that is learning and machine learning, it’s all about patterns across individuals. It’s never about the individual data point.

Right? And so this is very similar to what happens with climate change, right? If you emit CO2, essentially that’s nothing to whether or not you’re going to experience some climate disaster.

If all of us emit CO2, might very much have consequences for all of us. With data collection, it’s the same, like you sharing your data has no consequences for you. All of us sharing the data might have consequences for all of us.

Positive or negative consequences. And so, because learning is fundamentally about these patterns across individuals and not about individual data points, that means that individual privacy or property rights fundamentally cannot prevent any harms from AI or generate any benefits. And so, to have the control over our data, we necessarily can’t stay on the individual level.

We need to have more collective or democratic governance. So this is one of the reasons why I think it really doesn’t make much sense to think of this in terms of markets where you have like an exchange between individuals ’cause it’s really at this collective level. All right.

I’m just throwing a bunch of topics at you now, just to give you a sense of how this framework, I think, can allow us to rephrase different debates. So, another one is the question of workplace automation and the labor market. So this is one that, again, economists have paid a fair amount of attention to, and that also looms large in the public debate.

And so this is a question that obviously doesn’t just apply to AI, but has applied to many technologies in history where a key distinction that’s important is between average productivity and marginal productivity. So let me explain what I mean by that. When you get a new technology, basically, it can shift how much we can produce with a given amount of inputs, whatever the sector domain is.

And it’s true whether it’s a steam engine or railway or electricity or AI or anything else. And so if it wouldn’t increase like the total output for a given amount of inputs, in some sense, there’s no reason for profit-maximizing companies to use that because they could just stick to the old technology and do better. And so because of that, what’s gonna happen is unambiguously going to increase the average output for given the number of workers, say that are working.

But there’s a part of that that is very much ambiguous, which is not the average but the marginal output. And so to give you an example of this, which is the type of picture that economists might draw where you have, say, the number of workers working in some firm, and the output of the firm. The more workers are there, the bigger the output.

And as we move from an older technology to a newer technology like AI, say, unambiguously, we are gonna get more output for given inputs because otherwise, why would you use the new technology? However, what is very much ambiguous is how that marginal output changes. And so this is a type of picture that would correspond to automation or to economic growth without shared prosperity, because basically what determines what share goes to workers, and what determines how many workers are hired is not so much the average output, but the output for an additional worker.

And that’s true in competitive labor markets, but it’s also true in markets with a lot of market power of employers and various others things. But this is not fate. This is very much ambiguous, right?

There’s nothing intrinsic to the technology that should make the slope bigger or smaller. It really depends on choices again, and something we’ve seen. See many examples in history.

There’s a point that incidentally Daron Acemoglu and others have also been making over the last few years. There’s a lot of choices and that’s especially true when you have a general purpose technology like AI that you can use for all kinds of purposes. You can throw in any objective in there that you can measure.

Right? And so whether which way it’s gonna go is very much going to depend on who picks what the technology is being used for. And so this is I think where workplace democracy really makes a difference.

Like are you gonna use the new technologies, new tools in a way that replace workers, that depress the wages of workers? Or are you going to use it, them in a way that maybe make work more interesting, make complementary to the skills of workers rather than replacing them? Again, the bigger, big picture message there is there’s a choice.

And who makes the choice matters for what choice is being made. All right. Another topic.

Fairness and algorithmic discrimination. It’s also one that has rightly received a lot of attention and debates about the ethics and the social impact of AI. And one thing that’s been happening over the last few years is there’s literally hundreds of definitions of fairness that people have proposed in that literature.

And there’s been a fair amount of attention paid to supposed inconsistencies between these different definitions of fairness. I think that’s actually quite misleading. I think by and large, the majority of these definitions capture more or less the same idea in slightly different mathematical form.

And the idea is, of these definitions of fairness or the absence of bias or the absence of discrimination, is that somehow people have the same merit, whatever merit means, are supposed to be treated independently of the membership in some group, some protected category, say by race, gender, or something else. And this is something where I think economists have had an unfortunate influence on what’s been happening because economists since going back to Gary Becker in the ’60s have had a leading definition of fairness called, or of discrimination called taste-based discrimination which Becker explicitly defined as the presence of non-monetary motives. So what does it mean?

Imagine, like, an algorithm for hiring or deciding who gets invited to a job interview. When would this criterion decide that the algorithm discriminated? Let’s take gender as an example.

If according to this criterion, the algorithm discriminates if you could have had higher profits by hiring more women rather than men relative to what you did. Right? That’s the definition of discrimination here.

But what does this definition do? It basically starts out supposedly with the interest of a disadvantaged group. In this example say it might be women.

But it’s kind of an amazing, I think sleight of hand because within half a minute we go from, like, the interest of the disadvantaged group and we’re actually not asking there about the interests of that group anymore. We are asking about the interests of profit maximization. Right?

We are asking, could we have higher profits by doing something different in terms of hiring? Which is a very different question from what does it do to gender inequality or inequality more broadly. But that’s something that runs through this entire debate on fairness and discrimination is asking basically defining bias as a deviation from whatever the objective of the algorithm is.

And that’s the fundamental way in which all these definitions of fairness actually coincide is by basically creating merit with whatever the objective of who’s in charge is, right? So in the hiring example, it might be the, say the contribution of an additional worker to profits, which we call productivity. And then asking is that being maximized?

But it’s very different from another perspective on evaluating algorithms, which is to ask what are the consequences for the people being impacted in terms of some notion of their well-being or welfare? Right? So one question is can we rationalize the treatment in terms of the objective of who’s in charge such as profits?

The other one is asking what is the implication of using this algorithm, this objective for the well-being of different people and in particular for people in different groups. Yeah. All right.

One last topic And then I’m happy to open the discussion after that which is explainability, right? So which often is discussed also in relation to questions of fairness.

Like how can we explain AI systems, right? Like I’ve been talking about democratic control. To control something, you first have to explain or understand it.

What do we mean by explaining AI? And this question keeps arising. There’s a lot of confusion about it, and I think part of the confusion is that different people mean different things when they say explaining an algorithm or explaining AI.

And so broadly, I would distinguish three types of explanations here. One is explaining decision functions, right? So you have a trained neural network, say.

If it gets certain inputs, it produces certain outputs. And you wanna know what this is doing. So in the context of deep learning, this has received a lot of attention over the last couple years under the name of mechanistic interpretability.

But similar also in other domains and that’s essentially about understanding the function, right understanding the whole mapping from inputs to outputs after the algorithm has been trained. And that’s something that in particular AI engineers care about. And the reason they care about it is typically because there’s actually something that’s missing from the objective that they’re optimizing that can be in the form of robustness, right?

So I mentioned example earlier of classifying images as dogs just if they have a green pixel in row 23 or something like that. That would be a non-robust feature of the mapping and you wanna diagnose issues like that to get algorithms that are robust in terms of making predictions in new problem instances. Related are concerns about causality, right?

Typically, this type of pattern fitting has nothing to do with causality. But if you care about the predictions remaining correct after you intervene and change some of the features, then you really care about the causal structure. And that’s again something that they care about in interpretability here.

But, so that’s kind of the engineering concern is, and that’s typically done by taking a complicated function like a neural net and trying to somehow approximate it by a simpler function, different from what lawyers care about, which most often is not at the level of the functions, at the level of the individual decision. Let’s say you applied for a job. You got rejected.

You want to understand why. And you cannot, you might want to know that for a number of reasons. You might want to know it in order to contest the decision, right?

Maybe there was something discriminatory going on. Say the algorithm just rejected you because you’re a woman, and that’s both unfair and illegal, and you wanna contest that. And so you want, and therefore you care about the why question.

Or actually, in a different way, you might care about if maybe there’s something you could have changed about your job application that would have changed the outcome, and you wanna know that for next time, right? Maybe if you just got, like, this certificate in coding in that language, then you would have gotten the job. Something like that.

Why questions are kind of intriguing because the way we tend to think about causality in social sciences, life sciences these days is typically in terms of what if questions. And what if questions, that’s actually where the algorithms shine relative to humans, because those are fairly easy to answer, like for an algorithm that makes decisions, we can actually say what the output would have been if you had put a, given it a different input, because you can just rerun the same algorithm and give it different inputs and see what happens. So at least conceptually or technically, that’s in some sense an easy question, this causal question of what if.

But this type of explanation, it’s the reverse. It’s the why question. It’s not, would you be hired if this happened?

But like, why did you not get hired? And so one way people have tried to resolve that is by kind of flipping this question or connecting the why question to a what-if question. That’s what’s called counterfactual explanations.

And so the idea here is, say for a job applicant, you ask, what’s the smallest change of the inputs that would have resulted in a different decision, right? So what’s the smallest change to your CV that you could make so that then you would get hired rather than not being hired or something like that. So, it’s a nice idea.

What’s tricky about it is it all hinges on what we mean by small, right? And so I already mentioned, like, different motivations you might care about this not being discriminatory, then small means a change of things that shouldn’t matter, like your race or gender. Or it could be about small meaning things that that you could easily change, then it might be almost the opposite.

And so these things tend to be somewhat fragile to what you mean by small. And so again, going back to my overarching story here there’s a third type of explanation, which in a way is the most important one. It’s not so much about explaining the function, and it’s not so much about explaining the individual decision, but it’s about explaining the decision problem that the system is solving.

And again, what’s the decision problem? It’s what’s the reward, what’s the objective that’s being maximized, over what type of actions, based on what information or data. And this is the type of explanation I think that we really need in order to build democratic control of these systems.

Because in a way, once you’ve specified this input of the decision problem, it’s very much secondary of how exactly you’re going to go about solving this decision problem when this is something you can show in a fairly straightforward mathematical way. It’s, it almost doesn’t matter what kind of learning algorithm you use. If you have enough data and the algorithm is not too stupid, then it’s going to essentially end up at the same type of decisions.

And so say in the context of deep learning, people have noticed this many times. You can play around with the model architecture of the neural network. You’re gonna get more or less the same model performance, at least within certain bounds.

And so I think, again, this is the type of explanation that both is fairly broadly accessible, right? Saying that, I don’t know, Facebook maximizes ad clicks and it does it by deciding which of your friends’ posts it shows to you or whatever. That, that’s a fairly easy-to-explain thing, and it is the thing that really matters in order to then have a debate about, is that what we want as a society?

Do we want to structure our public sphere by maximizing ad revenue, or do we care about other things? And so I think that’s the type of explanation we need for democratic control. All right.

So that’s broad overview of the book. If you’re interested in, it’s available for sale. If you use the code UCPNEW, you get 30% off.

(audience applause)

(upbeat electronic music)

[WOMAN’S VOICE] Thank you for listening. To learn more about Social Science Matrix, please visit matrix.berkeley.edu.