What Can We Learn From The Histories of AI: A Conversation With Stephanie Dick

0:00Welcome to the Harvard Data Science Review podcast. I'm Liberty Vittert Capito, feature editor of the Harvard Data Science Review, and joining me as always is my co-host and editor-in-chief, Shaolin Meng, statistician, founding editor-in-chief of the Harvard Data Science Review, and the person who has probably made more people rethink everything they thought they knew about big data than anyone alive. Today, we have a guest who approaches artificial intelligence from a direction most people in this space rarely encounter, through history. Stephanie Dick is a historian of mathematics and technology with a focus on artificial intelligence, and someone who has spent over 15 years asking a question that turns out to be surprisingly hard to answer. What is AI, really, and where did it come from?

0:51Her work traces the origins of artificial intelligence through automated theorem proving, early facial recognition, and the first law enforcement databanks in America. And what she finds every time is that the technical choices were never just technical. They were choices about what counts as knowledge, who counts as a person, and what we're willing to let a machine decide. She also co-edits the Mining the Past column right here at the Harvard Data Science Review. So in many ways, she is already part of our family. Stephanie, we are so glad to have you here with us today.

1:29Well, welcome, Stephanie. And as a historian of science, which is a concept I guess most data scientists probably have a rough idea, but they may or may not know what did you actually do. So if you can just start by telling the broad audience what you desire, like what do you actually do as a historian of science? What a great question, Shelley, and it's wonderful to be with you. Thank you so much for having me on the podcast. So history of science is, I guess, a strange discipline that lots of folks don't know is a standalone sort of academic undertaking.

2:04And I, in particular, am what I call a historical epistemologist, which means I'm really interested in the history of what counts as knowledge and also the history of how and what we know in different moments in time. So, for example, it was 2,500 years that the Western world believed that the right way to know about nature was to watch it take its course uninterrupted. And then there was a rupture in European society where philosophers and theologians and aristocrats were debating how we should know the world and came out believing, in fact, we should do experiments.

2:49We should intervene in the way the world works and not let it run its course in order to understand it better. And historians of science want to know what's happening in those moments, what are the ruptures that change how we think we know, the practices that shape our knowledge. And my own research has been all about how the modern digital computer, the prospect of automation, different forms of artificial intelligence have shaped how and what it is that we know.

3:19So that's a bit about the field and where I've positioned myself within it. I'm curious now, sounds like you're doing both the philosophy and the history. So what does a philosopher do that you don't do? I mean, I come from a philosophy background and there is certainly an entirely branch of philosophy that is also interested in knowledge questions. Epistemology belongs to the discipline of philosophy. But historians of science tend to hold the belief that the answers to our deep questions about knowledge actually change over time.

3:57So whereas often philosophers want to come up with more sort of generalizable once and for all theories of what knowledge is, historians of science are open to the possibility that the answer to that question actually changes in some really fundamental ways. So I like to frame it as in we are doing very similar work to philosophers of knowledge, but we give historical answers rather than philosophical answers to some of our core questions about what knowledge is.

4:31You've argued, and correct me if I'm wrong, that, you know, history offers real tools for navigating today's AI landscape, which I think a lot of people don't necessarily believe. And it's not just context or cautionary tales. So for a data scientist who thinks history is for humanists, where's the most practically useful thing your work teaches them about AI systems that they're building or using right now?

5:06I think we're hearing all the time from the people who develop technologies that the technologies are going to change the world in some fundamental way. We are told that the internet will change the world. The printing press will change the world. The computer will change the world. AI will change the world. Changing the world is really difficult and complicated, and often it happens in ways that we don't expect and can't anticipate. And one of the, I think, most significant gifts of historical understanding

5:40is a sense of what really causes change in the world, what kinds of interventions transform a society or transform people's perspectives and opinions. And we find ourselves, I think, in a quite unprecedented historical moment right now, wherein it's difficult to anticipate where we're going. And I think that historical understanding can also offer some really significant grounding

6:10in unprecedented and disorienting historical moments. And I say this to my students a lot, but a lot of what gets called innovative is actually, historically speaking, quite conservative in the sense that these inventions reproduce the social order around them. They recreate many things about the society they were introduced to. They create more wealth and power for people who already have it. They reinforce the ways that we're already going about doing things.

6:43And so in that sense, some of the very new-looking technologies we have on hand might not be as transformative or innovative as we hope. So historical understanding can both help us make sense of the moment that we're in and make sense of how we arrived in this moment together. And it also offers us a sense of what kinds of interventions lead to more meaningful transformation over and against others. So kind of counterintuitively, I think that historical understanding is really the handmaid of innovation

7:16and of creating new futures because it's through understanding our history that we can really look at the world around us and recognize it and that we can have a proper theory of social change that allows us to go forward and make the changes that we want to see. Well, as a specific example and stuff, that I had a great fortune to hear you talk about the three acts of AI and love to have you to share with the data science community

7:48about your take on the historical development of AI, particularly these different acts and your new thinking about Agendic AI. Thank you, Shali. You know, it's my favorite thing to do is to give a three act historical arc for the artificial intelligence moment that we find ourselves in. And I think for lots of people, AI was just a kind of esoteric or science fictional idea until 2022 with the launch of ChatGPT.

8:18And I think it is really important, if not essential, to realize that AI has been more than one thing. It is more than one thing today. And you can look through history to see and to recognize that kind of pluralism of intelligences that we're talking about when we talk about AI. So Act One, where I often begin this history, is in the wake of the Second World War in the 1950s and 60s.

8:48And during this time, the hope was that artificial intelligence and this new technology of the modern digital computer would be able to improve human judgment and human decision making in the context of an increasingly uncertain and scary global situation with the possibility of nuclear war, with the possibility of geopolitical confrontation, between the United States and the Soviet Union. It was in that context that we first hear about AI.

9:22But in that moment, what intelligence itself meant for the early AI researchers was a way of thinking. Every moment in the history of AI is also a different moment in the history of what we think intelligence itself is. So defense intellectuals and logicians and mathematicians in this moment try to theorize human intelligence as being about our capacity for right reasoning,

9:52for rationality, for inference and deduction, for a kind of rule-bound processing of information. And they hoped that they'd be able to describe the human reasoning capacity as a set of rules and that they could impart those rules for reasoning to the computer, which by following them would reproduce human intelligence. And there were some initial successes with this paradigm, especially in areas like chess and logic. But it turned out both that human intelligence seems to be a lot more complicated

10:27than just our capacity for reasoning, and also that this approach would only really work in highly formal and rule-bound domains, which most of the world, as you were just saying, Shelley, is not like that. Things are messier than we might hope. So a second paradigm of AI starts to take shape out of a critique of the first one. Some researchers started to say, well, human intelligence must be a lot more complicated than just our ability to think in a certain way.

10:59Surely our intelligence is also a function of what it is we know. And the second act, the second dominant paradigm of AI was called expert systems. And it was all about trying to get human expert knowledge out of human experts, again, into a set of rules, usually for navigating a conditionally branching tree that captures something of the learned experience of a human expert. And this worked really well in certain domains.

11:31Again, for example, folks who've ever used TurboTax, that's my favorite example of a kind of legacy expert system that takes you through a branching set of questions about if you're single or married, if you own a home or you don't, if you filed taxes last year or you didn't, and so on. And that tree is supposed to bring you to, you know, an optimal tax claim without having to go to an accountant. But just like the earlier paradigm, this one struggled too,

12:03in large part for the reason, again, that human intelligence, human knowledge turned out to be really resistant to extraction and formulation as a set of conditionally branching rules. And so that paradigm also gave way to Act 3, which late in the 20th century took us away from a desire to reproduce human intelligence according to formal rules in the computer. And instead, the hope was that the computer

12:34could find a different path towards intelligence by coming up with its own rules for what to do based on patterns and correlations in massive data sets on which they're trained. And so with Act 3, I think the most significant inflection point is that we stopped trying to model human intelligence. Some people even threw out the very idea that human intelligence is rule-bound in that way. And we moved towards this kind of data-driven pattern recognition style of intelligence

13:07in machine learning and large language models. So I think it's so helpful to know that these different paradigms of AI have been at work in the past and also that they still are in the present. It might be that in a lot of situations, an automated reasoning engine is the right choice over and against a machine learning model, for example. And also to see that it's disagreements about the very nature of our own intelligence that have propelled a lot of this history as well.

13:39So you co-edit the Mining the Past column at the Harvard Data Science Review and work very, very closely with many data scientists. So when you look at the history of how data has been collected and trusted, or in a lot of cases, not trusted, what patterns do you see repeating themselves over and over again and certainly going to be repeating themselves in the current era of really large-scale AI data training?

14:15I think a huge part of the history of AI leads us to this kind of sad desire to bypass ourselves. There's this hope in the Cold War that whereas, you know, people are too limited, too slow, too hungry, too emotional, too arrogant, too whatever, to make decisions, that computers can be trusted in some way. And similarly, there has been a long arc of hope and opportunism and excitement and optimism

14:48that data-driven computing technologies are also going to be able to solve some of our problems by bypassing human bias or by just making decisions based on the data. And I think there's a lot of hope that data-driven decision-making can do better than human decision-making in a lot of environments and contexts. And I certainly think that's true. But the main takeaway from the history of data, I think, is that all data are human data.

15:20We actually don't get to bypass ourselves in some of these fundamental ways, even when working with data-driven tools. And there's a professor at NYU, Lisa Gittleman, who has a fantastic quote and book called Raw Data is an Oxymoron, by which she essentially means that all data that we have reflect and represent a set of human values and priorities and decisions. Someone has to decide which variables are important,

15:52what we should be counting. There is always more than one way to count. For example, in the history of the census, and we have a column at Mining the Past written by Dan Buk, who is a historian of quantification and statistics and the census. And the census offers us a perfect example of this, where different data were gathered about Caucasian men between the ages of 25 and 45 for many, many years

16:22than was gathered about everyone else for the reason that white men between those age ranges were seen as both the military and the economic capacity of the nation. And so the most sort of raw data that we have about the population has actually always been a reflection of a set of values and priorities and the worldview of the state at any given moment. And all data are like this. They carry traces of human decision-making,

16:54human values, human choices. There's always other data that we might gather. So I think one of the things we try to highlight in several of the Mining the Past columns is this profound lesson in humanizing our data and seeing them as reflective of human decisions, human values, human worldviews, and to recognize there are always other ways that we might count other variables to consider, other frameworks to see the world through.

17:25Well, thank you, Steph. It is true that most data scientists may or may not realize how rich these historical lessons are. But you have also worked directly in this line of work, right? I understand that you have research projects examining the New York State identification and the intelligence system, right? And where now this gets to a point where, as you've said, philosophers always emphasize there's no such thing as raw data, right?

17:56And one of them, for example, I think you studied how the categories of, say, like criminality itself is technologically constructed, right? Because, you know, it's a human-constructor concept. So can you speak a little bit about how the lessons learned from, say, 1960 systems for the currently there are lots of debates about the, you know, predictive policing algorithms, right? There's a lot of those issues in a very intensive debate.

18:27There's all kinds of arguments. How the historical lessons you have learned, you have investigated, can shed some lights on those topics of really great current interest and the future? Thank you so much, Shauli, for the question. And this is one that I could talk about for a really long time. And I think there are two main takeaways for the current moment. And the first lesson is that there's no automation without transformation. And that every time we start using computers to take over tasks that people were doing before,

19:00whether that's investigating crimes, whether that's identifying license plates, whether that's matching fingerprints, all of these tasks were redefined while we tried to develop computers that could do them. So this project led me to look at what was actually going on in the algorithms that were developed as early as the 1960s to match license plates, to match fingerprints, to match faces or photographs of faces. And doing something like photograph matching

19:32with the computers of the 1960s was a really significant technical challenge. These machines had, you know, 60 kilobytes of memory and were incredibly limited in their operating capacity. So when New York State started trying to develop an algorithm for matching photographs, the first thing they did was completely redefine faces and completely redefine what it would mean to match two faces. So in the context of this early system for police data,

20:03faces are redefined as a set of distance measurements between points on the face. So the outside of the eyes, the outer points of the mouth, the various points along the hairline, taken in pairs to become a set of measurements. The early algorithms developed in the context of this police databasing system would calculate and aggregate the differences between those distances across photographs and propose that the smallest distance

20:34was the likeliest match between two faces. And this is a too perfect example of the fact that often when we redefine a task so that the computer can do it, we are oversimplifying it or reducing it or taking out so many of the relevant variables. We often make hard problems solvable by really simplifying them so that we haven't solved the hard problem, we've solved the easy redefinition of the problem.

21:05And that was very true in this case. And Woody Bledsoe, who designed this algorithm, knew this was hugely problematic for recognizing faces, but the New York State police did not. When the algorithm traveled to them from the University of Texas at Austin, where Bledsoe was working, the nuance about this technical approach sort of fell away and the algorithm was seen as proof that police would be able to automate identification,

21:35automate mugshot matching using these technologies. So it highlights both that there's a transformation that goes in to making our tools work for us and take on these tasks, that often that transformation is reductive. And also, in this way, the transformation sometimes involves computer scientists building their literal selves into the tools that they are developing. In this case, Woody Bledsoe and his team

22:06came up with what they called the standard head. And that was a set of three-dimensional measurements of a head that they would assume of all heads in their database in order to simplify the calculations that allowed the matching and comparison of those distance measurements. And that set of assumptions, the introduction of the standard head, that's what made that problem solvable in the 60s. And even though we have way better tools now and we have way more sophisticated sensibilities about automation,

22:38we're still engaged in these transformations whenever we automate and introduce AI. We have to pay more attention to what those transformations entail. And we have to make those transformations with intention instead of just introducing what are often poor assumptions to make hard problems easier to solve. I work in facial recognition and facial shape analysis. And so it was funny when you said that people have literally put themselves

23:10in the system. My face is very much literally in the system. But I think it begs the question that you're now really embarking on this sort of large-scale program. And I believe it's called the ritual and algorithm that explores sort of the entanglements between mathematical, psychological, and occult theories of the human mind in the 20th century. And that's a pretty striking combination.

23:42What drew you to the occult as a lens for sort of understanding computing? And what does that reveal the purely technical history would miss? So this project really started for me when I was reading the letters, the archival letters, of one of the 20th century's most important logicians, Kurt Gödel, who immigrated from Hungary and was based at the Institute for Advanced Study at Princeton.

24:12And I learned from those letters that Kurt Gödel had been reading Carl Jung, who is one of the most important students of Sigmund Freud and one of the most important psychoanalysists of the 20th century. And Gödel was getting really interested in some of the occult ideas that Jung was interested in, including this idea that Jung has of what he called synchronicity, which was the idea that our state of mind,

24:44the state of mind of a person, might actually shape the outcome of events in the world in ways that we don't have a good scientific explanation for and that look like occult phenomena. And Gödel very famously has other interests in the occult. He writes about ghosts in the context of the 20th century and certain other sort of mystical phenomena. And this led me to a really interesting question, which is, you know, what does this central thinker

25:18about formalization and deduction and mathematical proof and certainty and their limitations? What role does the occult play in his understanding of the human condition? And so to make a long story short, I went along my way reading Gödel's letters and then reading some of this surrounding texts of Jung's and many others. And that led me down a bit of a separate rabbit hole, which I've called Mid-Century Men's Searching,

25:51which is that I uncovered this list of about 50 books, starting with Carl Jung's Man's Search for Meaning. And I found about 45 other books, all with the same title. It's Man's Search for Meaning, Man's Search for God, Man's Search for the Soul, God's Search for Man, the dual search of man for God and God for man and man's search for meaning and man's search for... And it all led me to this sense that, of course, the middle of the 20th century

26:22is this existential crisis where people are looking for meaning, they are looking at the human condition in new ways, and they're all at the sort of core of all of this anxiety are a set of questions about where does meaning come from in a world full of so much uncertainty and suffering and so much of what looks like injustice. And I now am convinced that artificial intelligence

26:52should be seen in large part as a response to that crisis of meaning and that existential crisis in the Second World War and that a lot of people who are active in AI research were also very actively asking these bigger questions about meaning. And one more step towards the project, this led me in turn to think about the fact that on the surface, what we call ritual and what we call algorithm

27:22are very similar, actually. Often ritual-like algorithm is highly disciplined. People who engage in ritual often have a set of very precise steps they are meant to follow or very precise rules or prescriptions for what they're meant to do. But ritual, through its history in religion, in the arts, in many parts of human society, ritual is a history of meaning-making. We engage in ritual

27:53in order to deepen our sense of meaning, deepen our spiritual relationship to the world or ourselves in some way. An algorithm, even though it looks kind of the same in the sense of bringing discipline and routine and procedure into what we're doing, it gets accused of doing the exact opposite. In all of these mid-century men searching books, there's this fear that algorithm makes us mindless, makes us passive, takes away our agency,

28:24takes away our sense of meaning. So I think AI then and now is a meaning-making project, or rather it is pushing us into a meaning-making problem. Shaoli has heard me say this before, but, you know, we used to write things down because they were important. We would write because that was the best way to save what we cared about and to communicate it. But now the written word is completely saturated

28:54with what AI has generated. I don't think writing is a good signal anymore for what we care about, for what matters the most. We are needing to find new places to look to make collective meaning. So no matter how much we automate, we will always be left with that job, that meaning-making job. And I think there's a lot to explore philosophically and historically about that. There's also a lot to explore about how some of our leaders

29:25in technology and mathematics and other parts of 20th century life, how they are making sense of meaning-making and the sorts of decisions that they have made about the human condition in that context. Well, thank you, Steph. You really make me think. I think that you have something really profound here. But in what sense that is something that the machine cannot mimic, right? Because, you know, people talk about AGI.

29:55I think I probably know your take on AGI. So let's get into a little bit, right? This whole artificial general intelligence. And it doesn't have to be human, obviously. But I think, like, you know, when you have enough data, let's say we all engage more and more into the meaning search thing, which we all do, but they become more pronounced. We will leave lots of data, a lot of, you know, trace how we search for it. We'll be writing more about them or talk more about them. And these machines,

30:27obviously, are good at collecting all our data, hopefully at some point, and then, you know, mimic the pattern, right? They could have, we could have just escalated, right? We do more, it does more, right? So at what point, or more kind of intrinsically, what is the kind of the search process of the meaning truly belong to human in the sense that it cannot be mimicked by any other intelligence? Well, maybe that could, it's just we don't know yet. Thank you.

30:57I love this question so much. And you're absolutely right that I'm an AGI skeptic, but it's because I reject the premise that we have AGI. I think there's this sort of idea that humans represent a kind of plastic form of intelligence that could work always and everywhere and put itself towards many different kinds of problems. And there's been this long desire for intelligence and machinery

31:27that's not domain-specific and so on. But for me, there are so many, and this is, I think, perhaps the core finding of the discipline of the history of science is that there is more than one way to know. There is more than one way to know. Out there in the world, people are making knowledge in really different ways that all have really profound value. And I would add to that that for me, intelligence is also a pluralism. I really love the discussion we had recently in the HDSR

31:59about whether or not it makes sense to talk about the planet as having a kind of intelligence. And if planet Earth is intelligent, it's intelligent in a way that has to do with relationships and the dynamism of ecosystems and the give and take and balance of the natural world. That's a really different way of thinking about what intelligence might be from the kind of logico-deductive, rational decision-making models of human intelligence,

32:29which are often actually quite brittle by comparison. And so I think that for me, this idea that there's something called AGI or artificial superintelligence or the singularity on the horizon that we're approaching is both incorrect, but also it's a fundamentally missed opportunity because I love to embrace a world where we think about a pluralism of intelligences, machine, human, animal, hybrid, and that our task

33:00in the future is about thinking through how to put all these different intelligences into relation with one another to unlock value and so on. So I'm of the mind that there are many different things called intelligence and I'm very open to the idea that there are forms of intelligence that are very alien to us that could be at work already in our world or that we have yet to meet in the universe. and yet I think that you know human intelligence

33:31will also have some very unique specificities that are born out of the specificity of our experience in the universe. You know, in philosophy there's often a debate about whether the human mind belongs to the brain, whether it's separate in some kind of dualistic sense, Cartesian sense of the mind being immaterial or my personal philosophy which is that the mind is embodied. I think that human intelligence

34:02comes from it comes from our experience and there was this tragic way in which in the 20th century our bodies, our emotions, our spirituality, so many facets of the human being were deemed not just irrelevant to our intelligence but they were deemed an impediment that if only we were less emotional, if only we were less spiritual, if only we were less hungry, if only we were less tired, if only we were

34:32less sickly, if only we were less mortal, right, we could be so much more intelligent. But my view and this was also the view of Hao Wang who I mentioned earlier is that human intelligence is actually a product of the very specific ways that we suffer and live and feel and navigate the world. I think human wisdom which is profound comes from a lot more than just data-driven pattern matching. It comes from storytelling

35:04and narrating and trying to find and make meaning out of our particular experience. So I absolutely think that every other form of intelligence in the universe can engage in meaning-making behaviors and projects but that human meaning-making will always fall to us because we are the only creatures in the history of the universe who have had this particular experience of mind and body and planet and it's our job

35:35to make meaning with that in the way that it might be the job of other intelligences somewhere to make meaning out of their experience of the universe. And as you were saying, you know, I used to critique AI by saying, AI has no access to the world. None at all. All it has is the data we give it and the data is always, as we were talking about earlier, it's always partial. It's always incomplete. It always reflects human priorities and decision-making and even with

36:06agentic AI that can become increasingly physical, interact with its environment, generate some of its own data, stay up to date in terms of real-time data generation, and that's going to improve the behavior of our data-driven models. But I don't think it stands in for the embodied experience of life in the world. And Alan Turing, actually, in a way that I love, sort of gave us that idea two years before he wrote

36:37his article on the Turing test that says we should call a computer intelligent if it can sufficiently often fool us into thinking that it is a person in a kind of discursive test. And two years before he published that article that reduced intelligence to conversation, essentially, he said, you know, the surest way to make an intelligent machine would be to take all the parts of a man and reproduce

37:07them by machinery and then let this giant machine man roam around the countryside learning things for itself. And even Turing says it would learn a lot, but it would still have no relationship to sex or sports or food or desire or all of these other human things that Turing also thought shaped the human perspective on life, the world, and any meaning-making project we might engage in. So I think there's no way

37:38out. There's no way out of the call to make meaning out of our own lives, not through technology or anything else. That will be a constant struggle, and it's one that I'm thrilled that we're taking on right now. I think a lot of us are overdue for an audit of where meaning comes from in our lives, and I think that AI is encouraging all of us to ask those really fundamental questions about where meaning comes from in our relationships and our learning and our work. So this is a moment for meaning

38:09and for fundamentals, and AI is nothing but a support and a motivator for doing that incredibly essential human work. Well, thank you, Steph, and I just want to put you on the spot, hope you don't mind, to give the audience a one-sentence description, not definition, but description of what is intelligence in your mind as the way you define it. intelligence is relations.

38:45Okay, I like it. I like it. Um, you know, for our Matt Swan question, I think, you know, it's important to note, you know, your work spans automated proof, facial recognition, police data banks, and now this entanglement of ritual and algorithm. And really, across it all, you've shown sort of how technical systems encode assumptions about what counts as knowledge.

39:15You know, who counts as a person, what counts as intelligence. But if you could wave your magic wand, what's one change you would make to how AI systems are designed, governed, or deployed? So one sentence, one change, and what chapter of history most convinced you that that change would be necessary? Oh, I want this wand.

39:47I will wave my magic wand, and in this moment, although there are many things I would love to change about artificial intelligence and the context in which it's being deployed, I would say end the sycophancy, because I'm quite troubled