Abstract

Jo Guldi previews her book The War Over Rent, which applies data methods to a century-long struggle over rent control, beginning in Ireland in 1881, where land valuators such as John Murray gathered evidence for the new Land Court. She argues that bureaucracy can work as repair and that public data helped end a century of conflict, which complicates a Foucauldian suspicion of quantification. She then explains the data work behind the book: counting debates in Hansard (the record of British parliamentary debate) from 1806 to 1911, then using topic modeling and similarity search to find the land valuators and Patrick Sheridan. Searches of prison and newspaper records with large language models (LLMs) supplied character detail, and she calls the overall approach the tapestry method. The talk asks which numbers count and who produces them, and what it takes to keep claims inspectable and contestable when AI systems mediate our encounters with evidence.

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.

[0:05] Okay, I think we'll go ahead and get started. Okay, we've had a lot of jargon, we've had a lot of computation. Are you ready for a proper history talk? Like proper, you might not know that you're into 19th-century Ireland, but I'm gonna get you into 19th-century Ireland, okay? So the year is 1881. John Murray is walking across Western Ireland. He measures fences, he counts chicken coops, he examines roads. He asked who planted the hedge. He notes the quality of the soil. His son doodles in the margins while they ride on the train to the next town. So why was the British government paying someone to count chicken coops in 1881? Because by the time Murray was walking across Ireland with this notebook, Ireland had become one of the least governable places in British Empire.

[1:05] There were evictions like this one, which used a medieval technology of war, a battering ram, to cave in the walls of the cottage so that those who were behind on their rent could never live there again. There were boycotts like this one where so many people refused to work with the landlord because of the rent increases that the cattle had to be herded into town to market for sale by a police escort because no draftsman could be found to herd the cattle onto or off of the railway in Claremorris. the boycott also by the name is so named because the cattle belonged to a landlord named Captain Boycott that's the origin of the term his cows there were assassinations like this graphic like that presented in this graphic rendition of the assassination of the viceroy

[2:07] of Ireland Her Majesty's representative in Dublin on his way to work one Sunday morning. And the democratically elected representatives of Ireland in Parliament had been thrown into prison. So after they were classified by Parliament as terrorists, after reports about their Irish speeches circulated. At the heart of this matter was Ireland's rising rent and the problem of eviction. After Thomas Malthus preached high rent as a means of controlling excess population, Irish rents had surged and eviction had become embraced as a political means of social control. High rents kept the Irish tenants mired in poverty. Karl Marx pronounced the worst housing conditions in Europe.

[3:10] Irish historians, meanwhile, pointed out that the land had been stolen in the 17th century. So something as simple as property law and the collection of rent had stopped looking like justice, and it started to look like an unfair inheritance of centuries of colonization. The question facing Britain was no longer, how should the rents be collected? It was, how do you govern a society that no longer believes basic institutions like property ownership are legitimate? John Murray, walking the Irish landscape with his notebook, belonged to a new profession, the valuators of land. Valuators weren't collecting statistics for economists. They were creating something much rarer, evidence that both landlord and tenant could recognize as legitimate. They were inventing a language of economics that both parties could read. Murray's calculations were designed to inform proceedings

[4:15] at the Irish Land Commission, also known as the Land Court of Ireland. Created in 1881 by Parliament, the Land Court was Ireland's best hope for ending more than a century of open warfare and guerrilla violence. Because of the court, tenants were empowered to take their landlords to court over escalating rents. The court was instructed to entertain the evidence Murray was collecting, evidence that the tenants had contributed to the Irish economy with every fence and chicken coop. Taking trains and wagons across Ireland, the evaluators were assembling evidence behind rulings that reduced Irish tenants' rents by an average of 20%. The rulings by the land court became the first rent control in world history. And together, evaluator, tenants, and judges transformed Ireland from a nation of beleaguered tenants to a nation of owner-occupiers who had capital to invest in their own enterprises. So nowadays we tend to argue about rent control as though it's simply economic

[5:20] policy. Should the prices go up? Should they go down? Does rent control distort markets? But 19th century Ireland suggests a different question entirely. What if the rent control wasn't originally designed to change prices? What if it was designed to restore When prices have been too high for ordinary people for too long, people stop trusting prices. When contracts have been abused, people stop trusting contracts. When judges are in the pay of one class of the population, the courts become illegitimate. People withhold their trust from the entire system. And at that point, fixing prices isn't enough. You have to build a new source of legitimacy. And that's why collecting data everyone could trust mattered. Of course, the evidence alone doesn't change history. People do. And Irish citizens didn't turn to the valuation of land first. That's the end of the story. First, people start responding to

[6:25] injustice and mistrust by organizing, some with violence, others by changing institutions from within. In Ireland, those different strategies sometimes took the form of individuals in the same married couple. I want to introduce you to Patrick and Elizabeth Sheridan, hotel keepers in a tiny town in the West called Tubbercurry. This isn't their portrait. This is a portrait of any random household in Ireland, but the scene that I'm about to sketch was also probably repeated in many small towns across Ireland. So we'll take this as a representative of two hotel keepers in a small town. Patrick Sheridan's family was evicted when he was a child. Early, he sought an alternative to landlord rule in the belief that legitimacy comes from force. He joined a movement called the and became the leader of a secret cell in an international network pledged to a mooring British empire through

[7:26] attacks on both sides of the Atlantic. He participated in an attempt to storm the British armory at Chester Castle in an attempt to get British guns and take over Dublin, the center of Irish English rule in Ireland. After being released from prison, he recruited other Irishmen for tasks like the bombing of London Bridge. Disguised as a Catholic priest named Father Murray, he single-handedly ran guns to Ireland from Paris. At Phoenix Park, he may or may not have been the individual who plunged a nine-inch surgical steel knife through the heart of the Viceroy of Ireland. Meanwhile, his wife, Elizabeth, raised by hotel keepers, is at home with five children running the hotel. She joins a non-violent movement called the Ladies' Land League that recruits middle-class and working-class women. And these women start counting. They start counting the rent.

[8:27] How much has the rent gone up this week, this month? And they make sure that those numbers get into the newspaper every week. So someone's paying attention. They organize peaceful rent strikes, where neighbors agree to withhold the rent until landlords recant. Elizabeth, in other words, seems to have come to the belief that legitimacy derives from public organization and transparent public evidence. So the Sheridans as a couple revealed something crucial about the strategies. not women are peaceful but the tactics of the Irish tenants gradually shift from attacks from violent attacks to the collection of evidence and this evidence-based process this process of exposing data to public view is becoming a political technology in Ireland in the rent histories gathered by women the newspapers all of these began to make invisible

[9:28] grievances publicly legible. Now, evidence, of course, only matters if we can do something about it. So it needs to reach those in power. Unfortunately, it does. The evidence that Irish land had been seized, that the rent was going up, reaches a series of people who can do something about it. A philosopher, J.S. Mill, an Oxford history professor, Goldwin Smith, who starts asking about the 17th century history of land redistribution. An aristocrat running a parliamentary committee who recommends the creation of a new court system to handle these complaints. Frederick George Ponsonby, author of the Bessborough Report. And then a poet who turned into a lawyer after all of his closest radical friends were thrown into prison on prison charges. John O'Hagan becomes the first judge of the Irish Land Commission. So these are the pieces that are in play. Now we understand John Murray's jottings

[10:29] about chicken coops and fences and soil. He's reporting to these institutions. They are charged with changing the land law of Ireland. When tenants request a hearing at the land court, they entered, often for the first time, an institution dedicated to making their experience transparent and unbiased. Forms assign each case a number be lost in the proceedings. And then when the case is scheduled to be heard, the judges visit the farm themselves and inspect claims about the quality of the soil and landlord investment or lack thereof. And the court took testimony and gathered from local neighbors about who built which barn. So this process shows us that bureaucracies don't just needlessly manufacture paperwork. They show us a modern institution in action, crowdsourcing legitimate information from wherever it may be. In this case, information from many sources about what the

[11:32] price should be instead of whatever price the landlord asked for. So this is a story about trust and distrust, a history of bureaucracy as repair. pair, and it's a story about data as the remedy to an age of distrust. And now we're going to start to pivot to a talk about data from a talk about history. In this 19th century moment, public data is what solves, puts an end to 100 years of open warfare, and it's data partially crowdsourced by peasant women asking their neighbors how much the rent has gone up. And it's partially data gathered by historians and published as accounts of land seizures. It enters the public record because there's agitation, and then it creates new institutions that are charged with administering repair on the basis of new data.

[12:33] Now, this is a very different account of data than that which we have been trained in for the decades in which the most known theorist of quantification is Michel Foucault. The suspicion of bureaucracy, the suspicion of quantification has led historians into a place where very often we take a simplistic account in which numbers are the instrument of capitalism and narrative is the source of the human. So I think, like most things, that binary deserves to be complicated. Numbers are just one form of communication through symbols, and they can be harnessed by numerate individuals for many purposes, as can narratives. the question is are modern bureaucracies serving a

[13:37] landlord class or are they democratic are they responsive are they dedicated to transparency can anyone collect data or is data solely the province of experts who are in the pay of one social class the book if you buy this book barely talks about the data behind this study, but that's what I'd like you to introduce you to now, and that's the account of the data, a different kind of data, the data that helped me find these stories. That's what I'd like to talk to you about for the rest of the day. This book starts in Ireland, but when the project began under the title The Long Land War, which is another book already published, I was starting with a much bigger question. What drove Britain from a nation practicing colonial land seizure around the globe to a nation pursuing land reform in many of its former

[14:41] colonies simultaneously? I read the India story in Andrew Sartori. I'd read Ewan Cameron and the Scottish story. I'd read about events in Ireland. I'd read about the land war. I'd read stories about Africa and indigenous rights in Canada and the preservation of Maori rights in the colonial courts. And so my question became about this global dimension. What would give me confidence that we were telling a causal narrative that helped us understand exactly the moment when the shift in property law occurred? So I started with the data that I had, a pretty good data set for understanding British Empire, the debates of Hansard, which are the House of Commons, House of Lords, from 1806 to 1911. And in there, I started looking for any debates where there was a nation mentioned in the header, in the debate title, Debates on the Land Law-Ireland, or

[15:45] Events in Kenya, and where in the text was some mention of eviction or ejectment, a longer list of words. And so here is the timeline, on the x-axis is time from 1800 to 1911, and on the y-axis is how many counts we get. of those mentions of separate debates where eviction is being mentioned. And we've colored them by nation where they come up. And keep track of this central green line because the Irish are the only colony that has political representation in parliament. They're not the only colony with a land problem or an eviction problem. They're just the only colony which has the power to send its own representatives to Westminster. 1880s and 1890s, the Irish continue to escalate, talk about the

[16:45] importance of eviction, even while more and more nations, Jamaica, the Blasket Islands, Bona, the Hebrides, Wales, West Africa, come into play of debates about similar issues of the voiding of indigenous property rights and the need for some kind of repair. So focusing on Ireland was an outgrowth of broadly engaging the secondary sources in the data. It wasn't the result of a pet interest in Ireland. I started doing this data work long before I went into the Irish archives. I was trying to figure out where I should start looking. And the evidence was clear that Ireland was doing something. Again, not because the Irish were smarter or had more justified claims than anywhere else in empire. They just happened to be connected. They happened to have representation in parliament. And so they were in a place to speak on behalf of a growing global chorus of

[17:48] complaints against English land seizures, putting something into action. My next question was when these events happened. And the 1880s have showed up already. but here's another experiment. What happens if we take Parliament and we put a green dot on the map every time somebody says the word land? Now in many ways this is a terrible experiment because you can use the word land in so many ways. You can talk about a land tax, you can talk about the land we walk on, you can talk about the green and fertile land of England, or you can land a boat. But watch what the evidence does. Basically, there's a timeline on the y-axis. We're going to move from 1800 to 1900. We have January through December. Then we have the week of the month. So one of these gray rectangles is essentially one page out of a calendar, one calendar month. And then within it, there are pixels that represent one day. So one green dot is a day. So sporadic mentions of land in the early 19th

[18:52] century. But what happens when we get to the 1880s? Basically every day on which parliament meets has a green dot in it. So there's been a structural change for how members of parliament understand the land question. So what about my characters? Where did they come from? And the answer is, again, systematic. Let's start with the evaluators. This is a topic model of all of parliament, a nested topic model. The computer is dividing all of the debates into sections. First, it's dividing them into four sections. If there are four categories, four topics, one topic is institutions outside of parliament, like courts. One section is how we present and do things in and then there's war. So within the finance section there are lots of discussions of land and the buying and selling thereof and the land tax and all of this looks

[19:56] sort of familiar if you know British history and then when you drill down into a 500 topic model the computer finds all sorts of individual debates about the land some of which are familiar if you've read the secondary sources what's going on in the highlands and some of which are unfamiliar and the unfamiliar Sections include these strange words like valuation. So of course I began looking up what is valuation. Valuation took me to the National Library of Ireland and looking for valuators took me directly to Robert Murray. So I'm indexing a problem, finding a profession that I didn't know existed, and then going to the source base. From 100 years of data, here's another similar report because at this time we didn't know how to use topic models so I was just throwing pasta at the wall. This is a process where I'm trying to teach the computer how to find the kinds of papers I want. I started with four canonical reports. The Devon Commission, Bessborough Commission, these are primary sources that I

[20:58] knew about. Stuff like this. I said to the computer, start with all of Hansard. Find me the stuff that's like the Bessborough Commission. The 25% of the debates that are most similar. This is a similarity measure. then I hand selected debates, did it again, so now I've got a much smaller corpus, and we did a topic model. So now I have a list of computer-identified debates. These are the debates that are most relevant to the study of land. And in there is one card in which there are a lot of personal names, like Lloyd and Shaw. And what this is about is another section of the manuscript that I didn't discuss. There's a moment in time in which Parliament starts looking into the land acts and discovers that the the contracts have been forged The contract saying yes, I will pay this much more rent is actually forged And if you look at the paper the paper is newer than the date of the

[22:00] contract So you can prove that These have been tried in landlord run courts so the peasants couldn't get a fair hearing. When Parliament starts looking into it, the cases are presented before Parliament, they get into the papers, the landlords have their names aired publicly, and the practice has to change. We agree that we have to agree about data. So my account about public data is based on these characters. And the characters aren't chosen at random, they're the characters who made it into Parliament. But Anne Shaw is a widow told that she has to pay more money by her landlord, Colonel Lloyd, and so the Lloyd Shaw case is one of these exemplary instances. Here's another case, a very simple topic model of all of Hansard. This was one of my earliest activities, and I'm highlighting topics that look

[23:05] like they have something to do with land. And one of them has a personal name that I don't recognize. And that is Patrick Sheridan. And once I have the name, I can do a full Lara Putnam. I can go into the archive. I can find all of the newspaper articles where somebody mentioned something crazy that Patrick Sheridan said that got him arrested for the fifth time for running guns disguised as Father Murray. And I think this is a fantastic character. He's so well known that he gets mentioned in Parliament, but he's never showed up in secondary sources except in a footnote. So Patrick Sheridan becomes a main character. Another exercise. There's a chapter on the land debates in Parliament. What happens in Parliament? Well, I could read all of them, and that would take years. There are so many debates. They're so voluminous. I did what I tell my students to do. I took the top 10 speakers in the land debate.

[24:08] I asked the computer to tell me which words are most distinctive of each speaker, which words are these speakers speaking at a higher rate than the other speakers, and here's a list of their favorite words. And then we can take someone like Mitchell Henry, who's a London dentist who gets a in Ireland. And then he gets elected and talks a lot about how landlords are great. They love helping their tenants. And why are the tenants so agitated? Well, these are his favorite words, immoral, arbitration. And we can call up all of the instances in which he says these words. So again, we're not just cherry picking the character or even the speech acts. We know exactly how to look for these instances that are most representative. So I'm moving towards a definition of some rigorous forms of survey in the archive and something more haphazard.

[25:11] I also tried using AI. AI can make hay with anything uploaded to Ancestry.com, including all of the British prison records. I got curious about whether some of my characters, like Patrick Sheridan or here Michael Davitt, did they have tattoos? They were in prison. This was like an 11th hour. Can I add some texture to my narrative? I would really like it if one of them had a tattoo. That would be great. In fact, they don't have tattoos, but all of the markings on the body are recorded when they're admitted to prison. And so a simple few hours line of querying on an LLM produced the handful of phrases by which I introduced Michael Davitt with a scar over his left eyebrow and a wart on his chin. It's there. prison record. Easy to find. So I would call this LLM sleuthing. The LLMs are amazing about finding a certain kind of needle in a certain kind of haystack, which are the databases that they have access to.

[26:11] Again, the LLM, when instructed to, found me this article about how the same guy, Michael Davitt, shrunk three inches when he was imprisoned after the Chester Castle raid. Too much hard labor, sitting on a stool, no proper furniture. He actually shrunk. This was published in the Boston Herald and then in a small Michigan town newspaper in the 19th century because there were enough Irish immigrants. They were curious about these affairs. LLM sleuthing could be very much like cherry picking. It's simultaneously very precise. You get very detailed information about the wart on the chin and about the shrinking, but not necessarily the full landscape. And by contrast, what I'm describing as my data practice, which is informed by these much older set of practices for dealing with text mining, I would call this the tapestry method for processing big data.

[27:11] And it's as if you're looking at a tapestry. Here's the Bayeux tapestry. And I can zoom in all the way from the picture to a single thread and then get curious about which color of thread has traveled the furthest in order to be part of this greater design. so just so trying to examine the big picture of events in global empire took me to Ireland in 1880 to the valuators to Sheridan to the Lloyd Shaw case in order to make sense of the global transformation of information that allowed the global land grab to be reversed at least temporarily and land to be given back to former indigenous owners and tenants. Okay, so thank you very much. Thank you.

[28:14] Questions? Chad. I'll back in just quickly because I think this is so interesting in terms of how you kind of turned on its head a little bit of using AI tools in terms of not helping us make sense of this huge corpus at a 30,000-foot level as much as helping you dive in at a granular level in order then to get back up to a higher level. And, you know, it's interesting in that earlier quantitative stage, you talked about how one of the things we learned is how that started to lead to almost a personalization in the same kind of way. Could you talk more about that in terms of how historians could think about AI tools and scale in your very iterative way of granular to overall

[29:18] 30,000 foot interpretation? Yes, thanks, Chad. So I think, you know, the primary problem of LLMs is that they're a black box. We're not allowed to see the algorithms that go in on the inside. The algorithms that go in on the inside have a family relationship to the kinds of data methods that some of us have been using for decades. But we're not allowed to inspect them. We mostly interact on the level of chat. And we can tell chat, transcribe this manuscript, create a database, all of these other tools. pattern matching apps. They're pattern matching, and we don't know how they're pattern matching. And so one of the things that we've discovered in the last 30 years of natural language processing with data is that, in fact, we've got excellent tools for doing something like saying, what are the top four things that they talk about in parliament? What are the top 500?

[30:19] How do those line up in a nested model? that is all accessible to you in LLM, even if you've never coded before. You can say, you can go home right now and say, bot, find me the US Congress debates. Bot, find Guldi's article about nested topic models. Write me the code. Now, tell me what to do to run the code on Congress. And you can pull up your own version of that. You can do that with your students. It's totally within your reach in a way that it really definitely wasn't Ten years ago. I was trying to persuade people to learn coding so that they could too could run topic models But the bot on its own isn't going to come up with the topic model It's not going to compare the big view and why not? Because what the bot excels in is summary and translation

[31:20] It can translate your email into more polite language. It can help you write your STS article for an environmental studies audience. It can switch out rhetoric very easily. It can also summarize. So it can give you a summary of the debates. But it can give you 500 summaries because summaries can be different. If you want to pursue the tapestry method, then you have to have a clear conception of the different kinds of pattern matching that a skilled data analyst would use. And that's what we've been doing in the digital humanities is to break down those different kinds of functions of pattern matching that a humanist might want to pursue. What happens if you want to find the words that are most distinctive of 10 people in parliament? That's TF-IDF. What happens if you want to break down discourse into a series of topics? That's LDA topic modeling. What if you want to see how those words change

[32:22] their meaning over time? That's word embeddings. So now you've got a Swiss army knife. If you read the dangerous art of text mining, you don't have to learn code. You don't have to learn any math. You can just believe me and read along. And you've got descriptions of this Swiss army knife of digital history practice and how it works. And you can deploy that on any body of text. And then when you want to find your character, once you've got a personal name, you can say, bot, find me Patrick Sheridan. Find me every piece of information you can find on him on the web. And I'll do the same thing for the proprietary databases. Another question. Please. I did have a technical question about the topic modeling and different tools and their evolution in time. but I guess that your article was from 2018 and I was wondering how things developed since then. But I do have a mythological question, historical question, and that is, especially thinking back to graduate school days, oftentimes

[33:25] we would come with some idea based on a set of sources and our older professors would be... But is it the representative? Is the person that you studied representative of society, right? So I wonder if you have insights into making a particular representative or something that is saying more than about the set of individuals that we have documents, but actually using LLM to say, well, I have the evidence that it is representative, even if it's not spoken in exactly the same words or exactly the particular document that we're looking at. So it would look very different if the data set that I was using was an index of individual peasants in their lives. So I could have had a data set that's peasant diaries. Here are all of the peasant diaries. And then I might be interested in what are the experiences that all of them have,

[34:26] and what are the most distinct experiences that they have. Now, this is also a matter of counting. For the consensus, for what they all have in common, you really can just count top words. That's fine. What are the words that they all have? For distinctiveness, that's TF-IDF. There's actually a series of algorithms, but I group them all as distinctiveness. And it's just what you saw me use on the parliamentary proceedings. Now, as to how these algorithms are evolving, yes, all of them are evolving all the time. the computer science department is employed to make new models new efficient, newly efficient, to use the newest technologies. So topic models are being rebuilt and rebuilt. The first topic models ran on probability, then they invented topic models like BERT topic that ran on word embeddings. The most recent ones insert a transformer step in the middle where it asks the computer to rewrite every sentence in the original text

[35:28] in generalized prose. So I might say, the tenants in this estate were of the opinion that it was worth pursuing action against the escalating rent. And the computer might generalize that as saying, the tenants thought the high rent was bad. And then it can count more efficiently other instances where tenants say something to the bad if it generalizes the same. So that's the newest topic models with LLMs in the mix that have names like Loom. So by all means, use the latest topic modeling. I don't have a preference for old style topic modeling. I'm showing you work on topic modeling that I did 10 years ago because it was a slow process of finding the algorithms, articulating their use for historians, doing this global 200-year study. It took a while. Yeah.

[36:29] Another question. Yeah. So I have a question that's partly about history but also partly about data. And the history part is where does the Irish potato famine fit into this story? So, right, one of the, as you know, one of the great demographic collapses in world history really, right, happened in the 1850s and 60s. then the story that you're telling sort of emerges in some way in the wake of that. Yes. Sort of the pure, just like we'll say historical narrative part of the story and then the data part, and I always sort of enter into these things through thinking about space and place and geography, and so I'm wondering if you've integrated work like Paul L's Atlas of Making the Great Irish Famine, not the old work, that's like 25 years ago work, or like Humphrey Sobel's sort of brand new work with Irish censuses and creating an Irish historical

[37:33] GIS, and whether that sort of kind of spatial and demographic kind of picture, I mean, that's a very different direction than you're doing with the parliamentary material, but whether that sort of where are the people on the land in these decades, whether that's something that you have brought or consider bringing into your story somehow and integrating that kind of data. And then finally, a sort of just for fun question about Maggie O'Farrell's new novel, Land, which is exactly, so cool, we just finished it. Oh my God, it's exactly, so sorry to steal the floor for a second, but exactly on the topic of an Irish sort of proto-nationalist photographer hired by the British in the 1870s to create this map and trying to like hang on to Irish place names and Irish information.

[38:37] A novel that like super well researched and historically informed. That's delightful. So if you don't mind, there's a hard copy right there of the galley. The book itself is out in March, if you don't mind passing it around. The book begins in the 1790s with Malthus and a new discourse over what rent means. Malthus thinks that the higher the rent, the better the landlord is doing. And if you reach maximum eviction, then that would be the most successful economy possible. So that ideology is actually already in place before the famine hits. And in the famine, and almost a third of the country dies of starvation and then fever, and then a third of the island leaves for other parts of the world. So the ideology of eviction, which accounts for that mass migration, is already in place. And the famine prompts, you know, there had been objections already to the rising rent and to this ideology, but the famine prompts the first parliamentary

[39:37] inquiries. and the parliamentary inquiries look at the worst of the evictions, the kind of clearance-like evictions where landlords in the west of Ireland are getting rid of all of their tenants but then many of the tenants are broken so the parliament is concentrating on investing in the landlords. Their ideology tells them that only landlords can create an economy so they don't know how to invest any other way. So that's the story in the 1840s and 1850s that takes place in the opening chapters of the book. And then the real action for me happens in the 1880s as they have to figure out how to build a newly legitimate institution and what kind of data will support that. I'm definitely checking out O'Farrell. You know, it's to my shame that I'm not more up. I know that you are the gazetteer goddess. And it's to my shame that I'm not more up on that because my own background is, as you know, in maps and geography.

[40:39] And I came at that, but then the kind of data that I started doing was text. And there was so, there's been this revolution in how computers handle text. And it was consuming to get up to speed on that. So, you know, I'm aware of some of these gazetteer projects and aware of the Cambridge Population Group, for example, and all of the amazing work that they've done with maps and demographic information. We have colleagues that are creating, including Ruth herself, who follow her work. It's amazing. We're doing such amazing things with GIS and technology. But I just point that way. In my dream, every history department department has one text specialist, one GIS specialist, one network specialist, and we can actually pursue hires because the shape of history departments is changing. We can't afford representation for every geography and every time period. Many of us cover longer and longer time periods in larger parts of the world.

[41:43] My mandate has been written in the world for a long time. And so, you know, we can't cover everything, but I think methods hires are promising to harness this kind of material for the students of the next generation. How are we doing on time? Fifteen minutes. Okay, fifteen minutes. That's great. We couldn't have a lot more questions. Please. I'd say about macroeconomics questions. So this rent crisis of, would you say, the 1870s and 1880s? That's the period of redressing it. The rent crisis is from 1790 forward and maybe for all of time. So how much of this might have been due to some of the factors? There's this story about the immense increase in agricultural production of the American Great Plains that that leads to the depression of agricultural prices that leads to all sorts of immigrants from Eastern Europe. How much of this pressure that these landlords and these tenants are feeling in Ireland is related to that story?

[42:45] Yeah, it's a great question. When I first started on this, I thought that it might be a Mike Davis story about El Nino events and famines and the intensification of global agriculture in the 1880s. Ireland is the breadbasket of British Empire. They're supplying most of the meat and butter for London during this time. And cheap grain from Russia and the Great Plains is competing. and Scott Reynolds Nelson has showed us how that's a increasingly a global competition which is forcing Ireland to be squeezed. Now Ireland was also the seat of scientific experimentation within British Empire. So in the same way that we tell the story about the land-grant colleges and the progress of agriculture here, there's a similar story via the landlords and the intensification of agricultural practice and geology being adapted for more intensive agriculture in Ireland, but all of it for export. So the Irish themselves are subsisting on potatoes, which is what leads to the

[43:45] demographic hunger crisis of the 1840s. Yeah, it's a terrific story. And there are all of these layers, there are all of these layers of global trade that you have to keep account of. But it's not as simple. In the Mike Davis story, the center is India, and India is in a demographic crisis and has huge famines in the 1880s because of the conjunction of El Nino and global competition over grain. So grain is suddenly expensive to import. It's not that simple in Ireland. What's driving this reform isn't that the Irish are suddenly hungry. There's a little uptick in the price of grain in the 1880s, but not enough to explain this. And the organizing isn't happening so much because people are losing their farms. Some people are losing their farms in the 1880s. But more what we're seeing is people whose parents were evicted when they were a kid in the 1840s and 1850s are grown up in 1880. And they're like, guys, let's join the militia.

[44:47] So we've got a childhood trauma case where the action is delayed for a generation. And ideas have to metabolize and spread. And then the life of the ideas in Parliament and philosophy in Oxford is important to that. In England, you have this movement of ideas in Ireland, and they meet in the Land Commission. Thank you. Another question, please. I got a little confused in the question here, because I was going to ask a question about how you disclosed your use of AI in the published version of the book. So it sounds like you've been writing your own code before AI to process this data. First off, incredibly impressive. Kudos to you. And so at some point did you switch over to using, I guess, more recent AI? That's question one. And number two, to what extent did you disclose your use of AI

[45:51] in that book? because that's a question that a lot of historians are wrestling with at a practical level. I mean, there's no real guidance on that either. We just have editors saying, please stop this AI slot. Right. I think, okay, well, what is that? Right, right, right, right. You know, a huge problem for the AHA. The Australian Historical Association has promoted more guidelines on how historians should be using AI in their publications. We're kind of in the Wild West in America, and so we have journal by journal. that work, please talk to me afterwards, because the American Historical Association has committees that are trying to solve this problem, and they need smart people. So yes, I've been doing data pre-AI all along. I showed you some AI-generated visualizations of data made the old-fashioned way, writing an R in Python, because the visualizations are better. AI just does the visualizations better. So we just scanned along those

[46:51] visualizations. I started using LLMs for a lot of things two years ago. By that point, the manuscript was largely developed, and I had already been printing out manuscripts and cutting and pasting them. So there are tiny bits of the manuscript where I do things like, I use LLMs as like a thesaurus all day long. for this, move the clauses around to be more compelling in this sentence. It's a fantastic editor with guidance. So far as disclosure, I wrote a general AI editing, data use, research is now the boilerplate on all of my articles. I have extremely high standards what my data looks like because I've been working on this for so long and because for

[47:52] historians to accept the work of digital history it's been so important for us to document openly and transparently which algorithms we're using and why so you know I have books books about that this summer I published an article called the paradox of the present with environmental history and it's basically about the conundrum of how we talk about AI use. So especially if you're writing environmental history, some of my work is environmental history, you have editors and readers asking but what about the data centers? What about the water use? And so the paradox of the present is that many of us think that these tools can give broadcast publicly accessible democratic access to data. And my telling of history is one in which democratic access to data allows ordinary people to give an account of themselves and demand institutions be responsive.

[48:53] So that's compelling at the same time that these same technologies create enormous burdens of care, which are not equal. So it's a paradox. I can't hope to solve it. I think disclosure is important. There was a wonderful talk in the earlier session where the historian was holding up guidelines for transparency, but again, we should talk after. I guess one more thing, sorry. Yeah. So, in the old days, when someone would publish a book and they had a bunch of graduate students helping them out, you know, they were listening to their documents. Yes. And that's kind of like another way to think about it. Maybe disclosure has negative connotations to it. Yes. Maybe it's more of a, when we start publishing things with the help of AI, it's like an acknowledgement. I want to thank Claude ChatGPT for going through this data for me. Here's how Claude ChatGPT did it, or whatever, since we're

[49:55] basically using it as a research assistant. Long love letters to ChatGPT in the future acknowledgement of books. It could be. It's kind of a wild west and we don't have any clear guidelines and it goes from one journal to the next. And one of the reasons why I'm here is because there are no clear guidelines. I'm just trying to figure out how everybody else is using it. Just to kind of test the waters to see what the future is going to be like. Thank you. We should follow up. Rachel? Yeah, I'm just like kind of a follow up to that question is like even for graduate student shooting range and do we I mean when we're disclosing it or like is it a statement of purpose uh is it do we put it in technologies or do we put it in like bibliography yeah yeah like what I mean and I'm curious what anybody thinks like my advisor and I are trying to figure out the same thing did you do you want

[50:58] to comment directly on that yeah I should let someone else in, though. But I was just going to make a look. I wonder if one of the things that this looks like is having methods appendixes in our books, which is what I did in both of my books. I mean, they were pre-LLM, but they relied a lot on data collection and data analysis of different kinds and with collaborators who are smarter and better data and GIS people than me. And so my books both include sort of co -authored appendixes that are sort of, I don't know, 10 to 20 page, like fairly detailed. Here's the data, here's what the data collection was like, here's the data model we developed, here's the kind of data analysis we did, and here's how that fits into the rest of the book. so you know I could envision that like I mean the data description that you did in this talk was phenomenal and like to be able to have an appendix like that

[52:02] sitting at the end of the book there is indeed such an appendix it is much briefer than what you saw in this talk this talk is essentially the seed of what will be another another book I've tended to separate the methods volumes because there are a lot of historians who want to know about that who might not care about and the actual factual work in the books. We have another question. I was just curious, if I could speak with the next book, do you feel like there will be more LLM usage because, okay, well, the algorithm is getting so much better I don't need to use the old methods anymore or I could just reproduce them. So is it going to be like, okay, you talked about the Tapegian method in previous, but now it's like, well, we're doing the Zutte's and things that have improved so much in such a short period of time. without a lot of theory, even though we don't know what's under the lid. Yeah. Yeah, well, I'll tell you from my own experience. I've lived long enough and done data practice and history for long enough to eat my

[53:03] words. So when I was a young assistant professor, I used to say things all the time, like, these are so labor-saving. I can study so many things at once, and I can write books so much faster. And the older faculty were like, yeah, yeah, yeah. 15 years after I said such things. And like, finally, for the first time today, I presented a total account of all of the myth methods that are used to establish my new thesis about the 19th century and make it more robust. Did it save me time? Kind of, except for the part where like, it expanded into a 200 year five book project. And I had to invent I don't know if that was time-saving, maybe it was extremely productive. So, you know, I think it's worth bearing that in mind, those of you who are embarking on large language models.

[54:06] I am experimenting, I'm tinkering, but a lot of what I'm using it with is to say, look, we just figured out a lot of the things that computers can do with language. LLMs can do it faster, LLMs can do it better, but these are still the processes, the basic building blocks of how you organize text. So we'll do it on repeat. So I still have to finish the next book that explains this methodological method, how you get to a distant reading of property in the 19th century. Yeah, and a lot of that's gonna be really old research, but not irrelevant, not at all irrelevant to the new methods. It's not conducted through an LLM, but there's still things that LLMs can do ask for. Yes? So I teach at the high school level and I'm always looking for ways to effectively prepare the kiddos that we're sending to you. I'm curious in this kind of research how you tackle questions

[55:11] of linguistic relativism, changing meaning of words over time, slang, different meanings from one community to another. I mean, you mentioned some of all the meanings land can have. How do you try and build that into the model you're creating? So in a different book, The Dangerous Art of Text Mining, I give a soup to nuts methodology of these language techniques in modern history. And it talks about my own experience growing up in a town where everybody learned math and nobody learned history. And I got curious about history, but I knew how to code. And so how I got onto this journey. And there I recommend the method of word embeddings as a way of understanding how words change over time. And I give examples like, I give examples of how words changed in the history of Congress debating since the 1870s.

[56:13] And we can see every five years how the context of an individual word shifts. practices that I've long taught with. You can find many useful teaching examples there. And all of these practices are super useful if you have a large language model. Because again, you can just say, computer, I want you to recreate this body of work using word embeddings to show how words have changed. Here is the data set I have, or here is the data set I want to gather. Now help me find the words that change their meanings the most. You know, if you want reference the book. I see your hand. I want to ask for another question from somebody who hasn't spoken. Sure. Yeah. So I had actually two questions. One, so I was thinking about the graph where you have references to rent in Ireland, but also in other English colonies. References to eviction. To eviction, I'm sorry. Yes. So how do you get data for the other places other than Ireland?

[57:17] And how do you deal with the question of parallelism between different data sources? So this is a body of work that deals with an indexical, important, influential data set, which is the debates in the House of Commons. So this is only words spoken in the House of Lords or the House of Commons. That's all that's in the data set. And each of those debates has a debate title. As the debate is advertised, we're going to debate the land law hyphen Ireland. We're going to debate events in Sebastopol. So you can use that as a certain kind of metadata. It's not perfect metadata. I could devise an LLM process to annotate it with better metadata. What is the country in which each of these debates is concerned? But I had just the title. So I use the titles as a rough index of which colonies are under debate, in which we mention tenants, eviction, etc.

[58:22] And, yes, another question. I have a question about expandable AI. Yes. Expandable AI, because you mentioned that in your talk. So when we use an error, that is black box. But before LLM, for example, the topic modeling, can you consider that is more explainable and reliable? I agree with you, especially use of topic modeling as an example. So when we teach students, we need to teach them LDA, how they work and the theory about that and the mathematics behind that. But if we remember, there is a period of time before LLM and between this LDA, there was a time after 2018, BERT and the Transformers started to appear. We have a lot of BERT that appear.

[59:24] And then there are topic BERT that cannot be explained by LDA, but not black box at all, but semi-black box. Yes. And then, but now, suddenly it turns out, because the model are too big, for training a bird, we know what we use as the training data to train it. Yes. We kind of feel like they are more reliable. But now, especially for those co-source models, we cannot see what they are, our training data, so we consider them as bad boys. But there is another group of people trying to study even in the new one level of the LLM to see, hey, you see when this query, this context input, and then this new one trigger, and then this light up, and that's why that input come out. So this kind of another way of explaining what is going on to a lot of historian and humanities scholars, we consider it still black

[1:00:30] box. To STEM people, they would consider this is already kind of explained. So I would like to know from your perspective, how should we handle this? especially in the future I can consider like because we already see topic work is much better than LDA and then as you mentioned that other new semi LLM and also topic modeling mix can be better then in the future how we should explain to students or other our fellow colleagues. So I'm going to be really brief because I've noticed that we're over time and I want to make sure that we all get the coffee we need to continue our work. So in brief, when I teach these methods, increasingly I learn to teaching a process. And that these processes have ancient origins. And this is also really important to remind our students of. And if we're teaching something like comparing different accounts of history,

[1:01:31] that's as old, at least, as Eusebius. and concordances, extracting concordances to see where a word appears, proofing against the text. Again, Eusebius, they are ancient writers, ancient historians who are trying to admit these processes. So it's important for the students to know that these are ancient processes. They're part of the building blocks of knowledge. They're how we decide what's right. And that there have been attempts to do them in computers for a long time, and they're getting better. So that gives you at least, you know, in a lecture in which you introduce something like clustering language into topics. You can talk about at least three points. How did we do this when we were in the documentary turn after Lord Acton when we're synthesizing? How do we do this when topic models were new, the first time somebody realized that you could do this with a computer? And then what's the state of the art? So three, yeah, because historians and most

[1:02:33] social scientists are never really going to care about the tiny improvements in efficiency that motivate our computer science colleagues. So that's not worth knowing. The basic function, am I finding the most distinctive instance? Am I breaking it into topics? Am I counting references to dates, references to phrases? Am I looking it up? That's what we need to convey as the new methods of text in history. They are the old methods of text in history, in fact. Okay, we should all get coffee. Thank you very much. These were excellent questions. Thank you. You were a fantastic audience, and there's plenty more to talk about, but we have time to interact on another day of the conference, so please find me. We can talk about AI disclosures and other methods in the field, research as well.