Abstract

Louis Hyman argues that nothing discussed at the conference was new, setting the Philadelphia Social History Project, which linked census records on punch cards, against the cliometric overreach of Time on the Cross. Quoting a 1965 review of a Yale conference, he says the barrier to early computing was the drudgery of coding and debugging, which AI now relieves. He describes teaching students to move from trusting numbers, to rejecting them as ideology, to treating them as incomplete but useful, as historians treat every document. He concludes that AI changes the scope of history but not its method, and that inspectable data pipelines restore trust in research.

Chapters

Transcript

AI-generated transcript. It was produced from the recording by speech recognition and cleaned up with AI. It has not been edited by the speaker and may contain errors. The captions on YouTube are AI-generated in the same way. Each timestamp links to that moment in the video.

[0:17] Good afternoon, everybody. I have had the most wonderful time the last two days being with you all. I hope you did as well. We began this conference, how many people started off? 1,200? now we're at 1,600 people registered watching globally. And that's because you all came, your excitement, your enthusiasm, all the enthusiasm of all the presenters. I wanted to offer some brief notes at the end of our conference to sort of wrap things up and reflect a little bit of what I heard and also just a little bit of reflection and indulge myself in this way. Now, one of the first things I want to say is nothing we did here was new. As we've heard many times over and over again, we have been grappling with the nature of computers and history for generations now and what it means for our discipline. We have written books and held proceedings on these exact same

[1:21] topics for decades. This is one from the mid -1960s. We have thought about the ways in which the nature of computation can influence our storytelling, the nature of our evidence, the nature of what kinds of ways we are thinking about the past. now during our conversations in the last two days we focused a lot we've talked a lot of the boogeyman Time on the Cross that cliometric overreach that imperious work of history that's so doomed quantification computation to the dustbin of historiography this is a very negative experience that if we did not have personally we had through our historiography classes which obviously is an amazing experience for everybody. And it's important to learn from historiography. But I'd like to offer a counterexample, a positive example from that same moment. Because even as we talk about Time on the Cross, there were

[2:26] so many other things that were going on at exactly the same time in computational history. And this positive experience, the one I've always found so inspirational was the Philadelphia Social History Project, which you can see behind me. You can even download it today if you want while I'm talking. You can download it from ICPSR. You can see here all the occupational codes that are very fun to talk about with undergrads when you ask them, what does it mean to be a stevedore or a hustler or all these other terms that perhaps they're unfamiliar with from labor history. The project, if you're not familiar with it, took its aim to create studies of the family and social structure, of the development of community institutions such as the church, school, and beneficial society, of migration and social mobility. As its director, Theodore Hershberg, explained, the project wasn't simply to discount or ignore traditional textual

[3:27] records, archives, to augment them, to complicate them with quantitative research. And I haven't heard anyone today say anything about replacing all the ways we do history. They're saying, hey, let's do this thing too. Hey, this might be useful sometimes for some of our questions. This is exactly how people approach this in the Philadelphia Social History Project. Now, the project drew mostly on census data, which at that point was still only written down. They transcribed the data into punch cards, which was itself a project, 500,000 individual records in Philadelphia, 24,000 businesses and all their variables, almost 12,000 black households described by the Pennsylvania Abolition Society. And the analysis of all this data was made possible by what was then called a sophisticated linkage program that will instruct the computer in tracing specific individuals, households, and firms from census to census and within each census.

[4:30] And that kind of analysis, whether done in the 1970s or more recently in the Living with Machines project, would not have been possible without computer aid. were the dozens and dozens and dozens of scholars and machine operators and transcriptionists and all the other people that worked on a project like this and the millions and millions of dollars of funding. It was a high-cost, high-skill project. And this is his dissertation. If you think dissertations don't matter, I would encourage you to read Theodore Hershberg's dissertation. It's extraordinary. And it is a remarkable text insofar as it addresses so many of the concerns we have right now, even though it was written in 1973. The Philadelphia Social History Project surfaced experiences that had been hidden before. The hidden histories contained in numbers. They evoked lives of people enslaved and free, understanding their lives as they went to church, as they went to work, as they lived in houses.

[5:34] They did all the things that people do, excavating for the very first time the Black Northern experience, something that had been occluded from history before. But here's the problem. It was really hard. It was very expensive. It required people to punch all those cards. It required people to know how to program all those statistics. Doing a correlation took a lot of time. It was very hard. Consider this review of a conference proceeding held at a place called Yale, which is a college in America, for those who haven't heard of it. It's from 1965. It's a little north of here. Consider this review of a conference proceeding. Wait, should we have a conference proceeding? Would that be fun to have? Yeah? Should we get a conference proceeding going? Can I get a yeah? All right. All right. Very good. All right. All right. sponsored by Yale University and a grant from IBM. Can you imagine? That sounds amazing. Now, one of the things it said in the review

[6:39] is that one of the frequently repeated observations in this book, you do not need to know much about computers to use them. We've all heard this, right? It's your own field and problem methodology that you must be expert in to produce worthwhile results. Again, no one said anything new in the last two days. the computer helps those who know what they help what help they want this is very similar to the language a user on ai the next paragraph however is quite different i think this is the kicker the thing that makes it very different today than during the first epoch of computer computing history the contributors to this volume who are working with computers and scholarly research are still doing much drudgery, but of a new kind. They are coding their materials so it can be punched, getting it punched, proofreading it, defining the routines it is to be put through, checking for bugs in the programming, keeping track of the processing that various sections of the materials have gone through, organizing the potentially infinite flow of output

[7:43] that is available. This drudgery was the barrier. This drudgery was the expensive part. This drudgery is what led people, I think, to abandon this because you had to do all this drudgery and then still do all the other work to get a job, to keep your job, to be recognized in the field. But it's this part that has changed, I think, that AI has relieved us of the drudgery of so much of that step. In fact, so much that now we are all worried it's too easy to do history research. That's essentially what our anxiety is. It's too easy to write a history paper. It's too easy to analyze a data set, right? And so this is something I did a couple years ago. and it's something to worry about this is me analyzing the philadelphia social history project in ChatGPT a couple years ago it's not even easier now somehow

[8:46] we're complaining that it's too easy to do all this work and by doing so we are losing rigor now if you've been in the conference the last two days like i have i think part of what you have heard is simply how fallacious that is. For the last two days, you've heard all the different ways that many rigorous historians just like you, who share your values and evidence and storytelling and veracity, truth, talk about how they're asking big questions, as big as the Philadelphia Social History Project, but also how much easier it is than it used to be. And this video you're watching is me doing a very naive analysis of the Philadelphia Social History Project. And what do I mean by analysis. As you're watching this video, I want you to be bristling. I want you to say, Lou, I cannot believe you're showing that in public. You are a shameful example of a historian. And I am, but not for this reason. I am for many other reasons, but not this one.

[9:46] What do I mean by naive? I mean, it's just trusting the numbers, putting that number up there as if it were true. we should not trust these numbers this is just the very first step in a process of understanding what these numbers mean it's a very important very useful first step but it can never be the last step and no one i have heard in the last two days thinks this would be the last step in any analysis. Did anyone here hear that? Can I hear a no? No, no one heard that in the last two days, no matter what people tell you. Now, in my class, I refer to this slide all the time. I refer to it in the last talk I gave, where I encourage students to do this kind of naive analysis, to treat the numbers as if they're by inspection true, reflecting a perfect account of how reality is, complete and objective, that modernist fantasy, that oracle nature that is the number.

[10:51] And then I encourage them to reject all that, invert it, say, all these numbers are our ideology. They are just an expression of power by an institution, whether it's the census or any other kind, or a hospital measurement or anything else like that. they are subjective. Objectivity is impossible. And then I encourage them to take the next step and think about, okay, what if these numbers are incomplete? What if they're not a total representation of how reality is? What if they are a map, but not the territory, as a certain Frenchman like to talk? How can we still make use of them if they are incomplete and useful. In fact, that's what we do with all of the documents we use as historians. All of our documents are incomplete. They are all biased. They're all subjective in some ways. They're all born of institutions or moments or events, ways of seeing the world.

[11:51] And in that way, data is no different. And yet we would not reject the archive in its entirety, would we? No, we would not. That is literally what we do for a living. Okay. So we move through these stances as we do the data analysis. We move through these stances as we move through the archive. And this is what I have seen in the last two days. How do we as historians move through these moments of analysis, thinking about how we can better reconstruct the past? There's been all kinds of correspondences. Claude helped me pull these themes together. read every transcript of every conference proceeding yesterday, pulled together the common themes. I want to be very transparent about where I am and not using AI. I'm using AI. I did not happen to watch every minute of the footage from every session yesterday. But this is what it pulled together. It pulled together all the different themes and correspondences of our work.

[12:52] And the fact that it is overlapping says to me that something is happening here. where you're coming together, creating a new set of clear and shared values about what computational history could mean and how we can push it into the world. We can help and write better and more ambitious history. All these themes we encountered are not about the imperiousness of new methods. None of these are Time on the Cross. They're the Philadelphia Social History Project for a new century. They enable us to ask more capacious questions, but we should not treat the AI as an oralist. We should treat it as a tool, or rather a research assistant, overenthusiastic, who really, really, really wants your approval, loves reading, and will probably dump caffeine on himself in some liquid form. All right? This is a useful research assistant.

[13:56] Insofar as you trust your research assistants, like I trust Charlie. Charlie works for me. Yu Yi works for me. They're both people I believe in. Isabel works for me. I trust all these people. They're not trying to lie to me, but they make mistakes. I make mistakes. Do you make mistakes? Yes, you make mistakes. We all make mistakes. All right? So it's not the right approach to treat an AI as an oracle. Ask a machine a question and just accept the answer. But that's the problem. Most people treat AI that way. Most people treat history textbooks that way. And we all know how incorrect that is. The right approach is to create inspectable knowledge, to use the AI to produce more verifiable data pipelines, to make our archives more transparent, to make our modes of analysis more reproducible, more inspectable, whenever possible, not all the time, but sometimes, when it's appropriate, when we're trying

[14:57] to do certain things. And used in this way, AI is not destroying history. AI is restoring trust in research in a time when people don't trust research as much as they ought to. It's restoring trust in our teaching. It's restoring trust in our ability to deliver a core skill to the 21st century, discerning what is and is not AI slop. That is, how do we make sense of evidence and create stories from it that move people, but move them in a relationship to the world as it actually has been? so ai does not change how we do history it changes the scope of it it's an extension of our values and our discipline it raises the value of importance of what we do as historians and how we judge documents machine reads the code keeps it honest but we still decide what it means And that is a

[15:57] capacity we teach as historians. And in evermore before, our kind of discipline, the historical discipline, offers something, I think, to students. Makes it more relevant than it ever has been before. So thank you so much for joining us all together in Baltimore at Johns Hopkins University. And hopefully we'll see you in the future again. Thank you. Thank you. We're going to have a short break, and then we are going to have the fabulous world historic Tim O'Reilly, visionary of Silicon Valley, talk to us about AI and its meaning for the political economy of Silicon Valley as part of the Margaret Levi lecture. Everyone is welcome to stay and hopefully will stay for that. And then afterwards, we're going to have a glorious reception as we had last night. So thank you so much. Thank you.