Environmental historians—all historians, really—are grappling with an ever-growing volume of information because of the availability of digitized data and sources. As these datasets grow, they become impossible for historians to read and analyze without the aid of digital tools. Text mining and Artificial Intelligence (AI) are two strategies historians in the twenty-first century are applying to manage the volume. These new tools bring new opportunities for the collection of data, but they also provoke philosophical questions for the discipline: Why should we engage with large datasets? What is the best way to engage with this data? And in the case of AI, is it ethical to use it all?
Early in October, Western’s History Department hosted Artificial | Natural: AI & Environmental History. During this two-day event, students and scholars in history and other disciplines from universities across southern Ontario attended presentations by three distinguished professors working with AI: Marnie Hughes-Warrington, Jo Guldi, and Dagomar Degroot. The event also included a roundtable discussion with Dr. Hughes-Warrington and Dr. Guldi and a workshop that explored the practical applications of AI in historical research.
We—Mary Baxter, Charan Mandur, and Thomas Stroyan, PhD candidates in history at Western University—had the opportunity to sit down with Dr. Hughes-Warrington and Dr. Guldi to discuss a wide range of topics, including environmental and digital history, the nature of historical research, and the role of AI.
Professor Marnie Hughes-Warrington (University of South Australia) is a philosopher and historian whose work focuses on how machines write histories and the potential of AI as a historiographical discipline. She recently co-edited two books: History from Loss (2023) and The Routledge Companion to History and the Moving Image (2023).
Professor Jo Guldi (Emory University) was the first person awarded the position of “digital historian” in the United States. She applies machine learning, statistics, and other big-data methods in her research on Britain, capitalism, and landscape. Her most recent book is The Dangerous Art of Text Mining (2023).
The following transcript of this interview has been edited for clarity.
Charan: We’re sitting here with Dr. Marnie Hughes-Warrington and Dr. Jo Guldi. Could you two introduce yourselves really quick? And maybe tell me what your favorite animal is?
Marnie: I’m a Bradley Distinguished Professor of History Theory at the University of South Australia. My favourite animal would be the Australian fairy-wren because one of my Aboriginal colleagues and I, every time we see fairy-wrens, we believe that we’re talking to one another through the bush, no matter how far away we are.
Jo: I am a professor of quantitative theory and methods at Emory University although my PhD is in history and I have been a professor of British history—Britain and the world, Britain and its empire—My favorite animal is the armadillo. The Texas speed bump. A curious, burrowing mammal.
The Historian’s Role in an Age of AI Part I
Thomas: My first question is for you, Jo. You’ve written about the value of text mining as something that can help scholars work faster, particularly in keeping up with the statements and views of political or private institutions. And you specifically say that computational text mining implies the possibility of automating criticism and turning a one-time critical reading into a routine audit. Do you think scholars or even activists need to rethink how we stay informed and do reading?
Jo: We’re doing an excellent job when it comes to theories of political economy, theories of gender, theories of the environment, thinking about multiple scales of power and agency. By the “we” here, I mean people in the humanities generally and historians, environmental historians in particular. But we have been entering an information age for 200 years now, and in 2024, it feels to many of us that information is speeding up daily in ways that we, as historians, are having trouble grappling with.
So how do you keep up with that? And how do you write a history that matters and captures the public attention given how much the organs of dissemination and distribution are changing?
“How do you write a history that matters and captures the public attention given how much the organs of dissemination and distribution are changing?”
The conservative posture of retreat says, “well, we’ve got journals, we’ve still got the book industry, we make audio books, people sometimes make documentaries. I’m just going to give it my best go and write a traditional dissertation.”
But we have an opportunity as a field to lean into the tsunami of information because if our training as historians teaches us anything it is how to read and how to think about structures of information, including mass information. It’s just that the knowledge of techniques like text mining—counting words—is poorly distributed.
I would like to see other historians using these tools to create, as they say in the literature about auditing, tools for continually assessing: for example, how frequently have the newspapers of the United States, Canada, and Europe mentioned climate change in the last weeks, the last month, the last year? Is there a trend? Are we talking about climate change more and more as the storms increase with frequency? By saying that it can be automated, what I mean is that you can write the code, you can think through your theory once in order to do that assessment, and then you run the computer program again next week and next week and next week and you get up to date answers, which is a way of persuading potential leaderships.
We really do have something meaningful to say because we have the latest information like the stock ticker, like the weather report, like the latest statistics on COVID or the latest graph of land decimated by fire, firefighters and all of those other organs as a way that ways that economists and scientists produce information, which is up to date with the very latest happenings. We have more theory about what’s going on in society and politics than all of those other disciplines put together. We have meaningful things to say. We just haven’t applied ourselves to the kind of information that’s out there and the style of talking about the most up-to-date instance of the present, which is what consumers of information expect in an era of mass information deluge.
Stories from Data
Charan: What kind of stories, or what kind of narratives can we get out of processing large amounts of data?
Jo: In my talk later today, I’ll mention a Kai Gehring and Matteo Grigoletto article that I think is really an excellent example of grammatical level comparative analysis applied to history. Their article discusses how a team of European computer scientists looked at tweets about climate change and government versus market, and they looked for subject-verb-object triples. You can count the number of times that somebody says something like “government will save the climate” or “corporations plant trees” versus “governments cause global warming” or “corporations cause global warming,” or “fossil fuels cause global warming.” You can count all of that over time.
Their method suggested that at least in the United States, whether people blame the government, think of the government as the savior of climate change and corporations as the problem, or think of corporations as the solution depends on who is in the White House. And so that’s an interesting finding. I’m sure it’s not the entire story. They just did a first pass of this method. Other people can apply it.
My research tends to focus on the debates of British Parliament, the debates of the U.S. Congress. In U.S. Congress records since the 1960s, I’ve looked for the speakers who attacked environmentalists the most. Going off of Merchants of Doubt author’ Naomi Oreskes‘ thesis that the problem is climate disinformation, I was able to use text mining to trace a history of attacks on environmentalists. And it starts with former Senator Ted Stevens of Alaska in the 1970s, who basically gives a vocabulary to those attacks, so very early on.
That’s a case where text mining allows us to be very specific and then dive into one case with more and more details. Using that database and then not going to Ted Stevens’ personal archive, there’s only so far I can go into that case study, but I can pinpoint which senator to look at, which senator is doing the most attacking. And it’s basically far and away, Ted Stevens through the early 2000s.
I’ll also be talking about more recent work about how the idiom of climate change has changed in Congress. What I find is that maybe the Naomi Oreskes’ thesis is less important than whether members of Congress talk about present time or future time. There’s been a push to talk about future accounting—with people arguing that we need a global coalition to contain climate or to measure the future harms of climate and to make repairs now to limit carbon or we’re going to be paying out the news for floods and hurricanes till the end of time. Meanwhile, others ask do we need to talk about the present economy as in who has jobs in oil right now or who has jobs in coal right now?
So there’s been a kind of war between those two scenarios, and it’s not actually arguing about the science, it’s arguing about whether we care about the present or the future in our accounting. Not many people in Congress are saying Well, I don’t believe in science.” There are one or two. But a lot of people are saying, “I don’t care about the future,” and that’s a bigger problem, that’s a different issue.
On the Danger of Conflating Correlation with Causation
Charan: How do you stop yourself from making an improper correlation between quantity of data? I feel like when you’re doing quantitative analysis, people usually fall into this trend of “two things are correlated just because my data shows that kind of thing.” How do you tell yourself, “Just because two things are happening at the same time, or your trend shows it, doesn’t mean that was the sole cause of it.”
Jo: Digital historians need to be as cautious as traditional historians have been about ascribing causality. And that actually sets us aside from a whole body of literature on causality in the quantitative social sciences in the hands of folks like Judea Pearl, who see everything as a causal problem where you can use test and training data but they’re not looking at the bigger context of asking about the specific places where causation can be accurately ascribed. So this field of the new causal reasoning—assigning some math behind potential causal reasoning, has transformed medicine. It is a big deal, and it is transforming political science right now.
It is coming for history. But in my conversations with clear-thinking political scientists, they have been saying to me, “We need historians to be in this discussion because history has, as a tradition, a much higher standard for understanding causal relationships and which connections are real than political science or medicine has. And we need some accurate parameters around this new math of causation that’s being applied willy nilly. We’re not sure what to do.”
“If you’re trained as a historian, it’s something you know about. It is a debate you should be weighing in on.”
So this debate is coming down the line. I certainly don’t want to be the only historian involved in this debate. I’m not. Historiography should be there, the theory of history should be on the front ranks. But also these are things that every historian, political historian, environmental historian cares about. So keep an eye open. If you’re trained as a historian, it’s something you know about. It is a debate you should be weighing in on.
Time in the Age of AI and Climate Change
Mary: In both of your writings, I was struck by your ideas about time. Time underpins all history, and particularly with the study of environmental history. How is this ever-increasing volume of data and AI influencing our understanding of chronology, periodization, divisions between past and present or past and future? How is it transforming our understanding of when history ends and the present begins, especially in a time of climate crisis?
Marnie: By training I’m a philosopher first and then historian. I’m trained as a metaphysician, and Aristotle called that the first philosophy. Most people think of that as the “I think therefore I am” philosophy, but it’s actually systems philosophy, the kind of philosophy that says, if you’re going to make a world and it works, how does it hang together? What rules do you need? Do you need an up and a down? Do you need a left and a right? Do you need a before and an after?
What makes a world functional, meaningful, emotionally compelling for people? These are really knotty questions you’ve raised about this and today I’ll talk a little bit today about the honey ant Dreaming for Australian Aboriginal people. People have really struggled to translate into English what we have clumsily called “the Dreaming.” Other people have called it the “everywhen,” which is this unbelievably beautiful inflection of the past, present, and future all together.
The message, the moral of the story is when you’re a metaphysician, your emphasis is on finding connections, wherever you are is that space of “I am connected to the honey ant and to the place that I am living in.” I am implicated in the future, and I am implicated in the past, but there have been models where people have tried to iron out these things to, in philosophical terms, take a very loosely packed suitcase and make it very fine-grained, maybe to make us feel like we have a bit more control over the situation.
So you get some really lovely models of chronology, which talk to specificity around causality. But there are lots of cultures in which it’s, you know, coarse grained is not the right language, where people do have a sense of presence in the past and the future. Chinese imperial histories were not written in chronological order. They were written in place order.
But unfortunately, our historiography collapses in and we often tell a story, or we say that we have always put things in linear order. We’ve always thought about causality in these ways, and therefore believe “I must write those things” in that way. That expectation of a linear, successive order is only a recent moment in historiography, but it lives in the discipline; I don’t disagree with that. But I have never for a moment believed that we aren’t connected with the long time of geology and the universe, that we’re not connected with the worlds of animals. We are all in this together. That’s the system, the planetary system that we’re part of.
And if, in the cause of ironing things out, we write a definition of something that means that we don’t see the planetary crisis or the things that we’re having to deal with, then we have a problem. So as a theoretician, my job is always to say to people, “What’s the connection? What’s the system? What’s the world that we’re part of?” If we cast the world in particular ways that means that we don’t see it in other ways, then things can burn down in front of our eyes and we don’t know how to act.
I’ve always been interested in histories written by people who aren’t historians. I’m now interested in histories written by non-humans. We can’t pretend that we’re not connected to these non-human histories. And we can’t say that we made all these things, either. We can’t. I’m very interested in causality, but sometimes, like fuzzy logic, you iron it out, you give a list of numbers from 0 to 10, and you forget just how loosely stuffed the suitcase is and that, actually, not being able to iron it out perhaps is the biggest ethical call that we have on all of us. Right? If you write a history where you iron it all out, what work is there left for all of us to do for that future?
Mary: How do you practice that connection?
Marnie: Most people come to history not through professional history but by watching historical films and television. So we must go to the places where history is made and heard and not to decry it or dismiss it, but to understand it, and to place ourselves in those places with empathy, to understand where people are coming from. And in those spaces to understand, for instance, how somebody might hold up their mobile phone and film what is happening around them not just because they just want to be entertained by things, but as a cause to action. They actually do want to change the world because history has not spoken to them in lots of ways, and they feel like whatever is on the screen or whatever’s in the book is not about their lives.
“Every history is not a statement of what is; it’s generally a statement of what ought to be.”
I am very interested in how people take to themselves practices of history making and show others how they can enact change on all kinds of levels whether they’re writing a very local history or whether they’re thinking really, really at the planetary or the universal scale. And I’m very interested in the worlds of histories I think are as much about the future as they are about the past. It’s very hard to read a history without seeing some notion of lesson or orientation to the future. The very first written histories were couched in those moral terms and those ethical terms. So every history is not a statement of what is; it’s generally a statement of what ought to be.
Jo: I’ve also written about the longue durée as something which is relevant to the future, which can give us a shape of the change we want to see in the world. The long past helps us to think about the Anthropocene. If we think it was created by agriculture, it means one thing. If we think it was created by Britain’s investment in fossil fuels, it means another thing. If we think it was broadened by a system of economic globalization and the spread of capitalism that means yet another thing. Thinking with the past gives us goals to orient ourselves towards.
That kind of thinking is critical right now when we seem as a planet to be at a loss for remedies for the climate emergency. It seems to me that we are living in a moment of a true vacuum of ideas. There was a kind of flourishing neoliberal fit of ideas about cap and trade and carbon taxation and possible mechanisms for goading on innovation and engineering and the sciences that would solve climate twenty-five years ago, and we’ve been writing those out ever since. These ideas are not keeping pace with the problem, in fact, they’re creating new problems, and now we have AI.. AI is the number one new sector burning energy and creating demand for fossil fuel consumption right now.
We find ourselves at a sort of dead end, if we take this seriously. And this is still maybe taboo to say in most settings, but it seems to me that those market-based solutions have proven that they’re worthless. We need other solutions. So what are the solutions that look like some sort of an international alliance?
Well, we can look to the past for the history of treaty making. People who don’t know much about history will say, “We know everything there is to know about institutions and global government. Do you want it to be an authoritarian communist regime? No! But the United Nations is corrupt and doesn’t do anything. Therefore, what can we do? Nothing!”
The historian is more curious about the past because the past is way thicker than that. In between the polls of autocracy in China and capitalism out of the United States, there are infinite other models such as the history of the League of Nations and international treaty making in the United Nations. Many other forms of international alliance have been tried. Just look at the history of regulating fisheries. In 1945, we had no law to regulate who could fish where. We created a modern map of fisheries, which still has problems, but it has protected the oceans from being entirely overfished, and we can continue to build on that legacy.
What if we take that law of fishery commons and apply it to the atmosphere? We know so much more about the governance of natural resources as commons than we did in 1990 thanks to the embrace of Elinor Ostrom‘s research on the global commons, which was built on a history of Indigenous administration of natural resources in peasant commons for fishery, forestry, and grazing around the world. We can write the modern history of the commons and that should be a blueprint for administration because these are stable economic and political institutions that have a robust history. We just have to step out of the unipolar sort of boiled down to the bare bones, Sam Huntington, Francis Fukuyama version of history, where history tells us that either you’re Russia or you’re the United States and that’s the end of the story.
“We need other models, and historians need to be talking in those circles, in those mainstream venues, in the news, on the panel with the political scientists, and with the scientist and the engineer, talking about climate solutions, and in order to do that, we have to be a little braver.”
It’s not the end of the story, this is the very beginning of the story. We know that market messianism has totally failed the climate crisis. We need other models, and historians need to be talking in those circles, in those mainstream venues, in the news, on the panel with the political scientists, and with the scientist and the engineer, talking about climate solutions, and in order to do that, we have to be a little braver. We can’t just do animal history. We can’t just do local and place-based history. We have to think about planetary history and the organs of planetary governance.
Marnie: Agree. I would just add as an observation that I come from a country that has been unable to sign a treaty with its Aboriginal Peoples. And that also is important to say. If we cannot come to an understanding with the oldest continuous living culture on the planet about how to nourish the places we’re in, then we have not seen what we’re being called to do.
That point of view doesn’t necessarily make me an activist. What it makes me is someone who cares deeply about ethics. And I think often people use that word activist, which is fine. I don’t mind that. But let’s not confuse deep groundedness in ethics for activism.
The Historian’s Role in an Age of AI Part II
Thomas: Marnie, you’ve made the argument that we should accept artificial history makers. And you talk about how you see artificial intelligence as a historiographical discipline and that machine-generated histories will be more than simply imitation of human-generated histories. With these ideas in mind, where do you see history as an academic discipline and as a profession going?
Marnie: We’ve been decentred a few times before. People began making historical films in the late nineteenth century. People have always produced histories in lots of different formats. And so professional approaches to history should be contextualized in a much longer history and a much more global history where there’s always been lots of variety.
At what point would we recognize AI as an historian or AI as historians is the really critical question. Some people will collapse the question down and say AI is just a human-created thing—you’re really just talking about humans playing with shadow puppets; it’s really just about humans. Other people will collapse it down to capital and say you’re just writing the history of capital. Or you’re just writing the history of autocracy in those spaces.
I don’t think I am because, interestingly enough, if you look at the way in which Gen AI—and AI is a big cluster of technologies, let me be clear—if you look at the way it’s writing history at the moment, it is stripping out the conditionals. So the “ifs”, the “buts”, the “maybes” which professional history has used to indicate its ethical tentativeness and to call people to action. If you look at the way history is written by machines, they don’t use that language. Why?
The bulk of the public probably find the way professional historians use “if,” “but,” “maybe” really quite problematic. Yes, Gen AI models generally scrape language from all kinds of places. It might not be highly specialised language. But there also can be trouble dealing with these kinds of ways of talking that can generate safety issues. If I say, “X might have happened,” I can be ethically responsible in saying that, but “X might have happened” might also undermine confidence in something. It’s very hard to navigate those things. So we are already seeing a stripping out of that logic, in a way.
“I think [AI] already has changed history making and the logic of history, to be honest with you.”
I’ve always believed that history has got a kind of hybrid logic. It’s propositional, so it’s about what’s true, what’s not. It’s modal. It gets us to think about possibilities. But it’s also what R. G. Collingwood called question-based or Erotetic logic. It’s open ended. It wants you to think about the potential of action or evidence. And AI has not got that Erotetic element in the way that historians have traditionally done it—in the way of the footnote. I think the interesting thing is how that question-based element is going to re-transpose itself into new forms of expression. So I think it already has changed history making and the logic of history, to be honest with you. Okay, go for it, Jo.
Jo: It’s so good to hear Marnie’s comments on this because she thinks so rigorously about the elements of logic and history in a way that I haven’t done. I feel like I my approach to LLMs [Large Language Model] has been much more in the muck of what the machines can and cannot do.
Right now, there’s a very clear line between what LLMs can and cannot do for history. If I ask ChatGPT or another LLM to give me a history of the American Civil War, it will generate impressive paragraphs. I’m very skeptical of those paragraphs because I know that the material that the GPT is trained on, which is everything on the web between 2015 and 2018, has lots and lots of student summaries of best-selling books about the Civil War which have been the one of the number-one arenas for American publishing for the last forty years.
So everybody is regurgitating the same historians. And if you’ve got a bot that knows how to summarize, and you say, “summarize Civil War stuff,” it’s going to have a lot of detail. That doesn’t mean that it knows what a primary source is, it doesn’t mean that it can do any of the structured comparisons between first-person primary source accounts from the past. It doesn’t mean that the AI can think because it’s just summarizing other people who have compared different points of view of famous historians commenting on the past.
For 10 years I’ve been text mining history, which means I take a well-understood, well-indexed set of documents filled with words, like the debates of the House of Commons and House of Lords of the British Parliament. We know what’s in them; we know who’s in them—Gladstone, Disraeli; etc. We know what they’re about—the Industrial Revolution, rights for women, etc. And we can text mine them. And then sometimes we’ll ask the question, “Can the computer tell us anything more?” We can do that with statistics that make some invisible things more visible, and you can make discoveries.
This kind of text mining is much simpler than the neural networks behind LLMs. In text mining for historical analysis, we are counting words over time: How many times did you say “armadillo” in the year 1820 versus the year 1830? That’s where we start. And we do it with lots and lots of other words. Then we use some special statistics. Text mining for historical analysis produces structured comparisons of sources.
So I can ask, “Who during the American Civil War asked about armadillos?” And this could be a totally novel question. Maybe no historian has ever asked it, but I can teach my bots to ask it. I have full confidence in their answer because the answer is real. The process is really simple. The analysis is really simple: It’s just a count. Whatever count it comes up with is accurate, persuasive, and informative for those sources. And I know exactly what sources are coming in.
I foresee a day when ChatGPT will be able to do the same thing. But, possibly because ChatGPT is not designed by historians for historians—it’s designed by computer scientists who barely remember their last history course—it doesn’t distinguish between a bot summary of blog posts about books (which summarized other books about primary sources) and the kind of computational analysis that looks at new documents from the past. And so it’s easy to believe that LLMs are thinking about history when they’re actually not capable yet of producing new knowledge.
The term that I use in The Dangerous Art of Text Mining is a cyborg historian, playing on Donna Haraway. What would it mean if the historian knew enough about these technologies to appreciate this difference between the cut-and-paste history à la blog entries from ChatGPT and text mining for historical analysis and combine them?
There are a lot of things that could be built. And that would produce a more ethical, less biased AI, a safer AI, an AI with footnotes—and AI that knew what footnotes and primary sources are. But that has to be built and it can’t be built unless historians and digitally trained historians are in the room where the building is happening.
We have to be in the room where the engineering conversations are being made. And at the moment, I’m sorry to say that most firms in Silicon Valley have a prohibition against hiring people with humanities degrees, so that even if you really know your digital history, the HR department will look at your visa and say, “Ah, this isn’t an engineer, data scientist, or computer scientist.”—which means that those Silicon Valley firms that do AI have cut themselves off from the major disciplines which know about the structure of texts and the structure of knowledge. We should be kicking down the door in the names of the truth, of functional algorithms, of dispelling the marketing and propaganda.
“We should be kicking down the door in the names of the truth, of functional algorithms, of dispelling the marketing and propaganda.”
While text mining for historical analysis, I’ve taken great pains in my writing to try to produce a type of historical praxis, which is well aligned with the high standards of proof in the historical discipline. We are, above everything else, the discipline of evidence. We are the discipline that always cares about where the document came from. Where did you get this source on history? Was it really written down by an Indigenous person, or was it transcribed by a missionary writing the words that they poorly understood in an Indigenous language? We as a discipline have argued endlessly about the difference between those kinds of sources and we have a special kind of knowledge that is very important for the world of information.
Moreover, with LLMs, there’s been a push from the industry to talk about the machines as actual personified intelligence. I think of that as a marketing strategy. Friends in the tech industry will say, “Let’s ask her,” meaning Siri or an interface for open AI or ChatGPT. Then they encourage users to address themselves to the computer. It’s actually built in, but underneath that it works as marketing because we have so many, so many films already, which help us to imagine what an AI consciousness might be like. Feminist theory has a lot to say about the personification of LLMs, and it’s important because if we’re not aware of the feminist critique of personifying an LLM, then we’re liable to not see the strengths and weaknesses of the technology.
Marnie: But that’s the transposition. It’s interesting to me, thinking about what kind of approach to ethics would you need in a world like this, where you’re dealing with machines that use the word “I.” There’s a whole movement called explainable AI, and a sub part of that called advisability where, if the machine uses the word “I,” or a clause at the end of the sentence says, “would you like me to go and do X,” it’s presenting itself as taking advice from you.
That instantiates more trust. Trust tokens, basically, get you to form a relationship. No wonder we think they’re wonderful and we don’t swear at them. We do all those things because there’s an apparatus there of saying, you know, “X” AI will bring explainability to the front door. If I explain how the apparatus works and I take advice, then the user will trust more because the user is skeptical in a lot of ways.
The EU regulations have really increased awareness of people’s stuff being stolen, people’s property being scraped, pictures of their children being taken, all of that kind of stuff, right, and not giving back in the way that people expect it to be. So “X”-AI has grown up in that and as part of that; that “I” is shorthand for ethics.
Now in my world, the front end of the apparatus is not enough, but nor is it satisfying enough to set up an ethics panel or to write a set of ethics principles. They can just be ignored.
And so for me, I’ve been interested in the question of what does AI as historian, as an idea, usher in? Is the need for the discipline of history to come back to what it has not perhaps paid as much attention to over the last century, which is its logic?
We’ve been very interested in our narratives. I don’t disagree. I think they’re incredibly powerful. A story in history captures people’s ears, but there’s a bunch of people out there that engineer algorithms, and those algorithms don’t speak for the ways in which historians do.
Historians have to come back to the places where they were, which is to place themselves back in the room and say, “Actually, we do know how to program history. We just might not use symbolic language or mathematics to do that because logic has moved in the direction of natural language.”
We have to be ready to grab the opportunity and to do something with it in order to say to people, “hang on a moment. Had you found yourself trusting the AI because it was using the word ‘I’ and offering a question mark? Did you understand the interrelation that it was setting up with you there? Could you re broker the relationship in a different way?”
Information Bias and All the Doors to Kick Down
Charan: We have discussed broadening history to a global scale, the longue durée, and collaboration. But some believe digital history is skewed towards certain types of information and more recent histories. Should we be kicking in this bias door?
Marnie: That’s the door that most people kick in first, is the bias door. I think that’s a door that’s very familiar to people in the discussion about AI. They go to it first, and that’s appropriate, because if you look at the data lakes and the poverty of data, globally, you’ll see that there are deep wells in some spaces and no wells at all in others. So: totally agree.
Think too about some other things that are going on, which is we are all implicated here in a global story of supply and connections. It’s not just you’ve got data points, but we are all reliant upon the mining of lithium in different ways, right? The hardware has a role to play here. I can have all the data in the world, but how can I access it if I’m not actually hooked into the system, or there are sanctions or regulations around that?
There is that geopolitical, global story that we’re all part of too. I think that needs to be understood. There’s a safety story here around particularly the rise—Anu Bradford says that in Digital Empires—the rise of the EU’s privacy legislation. Even the nominally capital centre of the world, the U.S., has come to corresponding to that regulatory framework because of consumer demand.
So, there’s a regulatory side of the story here, which you would see could have been highly unlikely. It’s interesting to me that the EU has exerted influence in those spaces. And then there’s a whole bunch of people who are doing kind of liberatory AI, who are just out on the plains having a great deal of fun building stuff. In Central Australia, two years ago, there’s a group of colleagues who created a whole bunch of Indigenous emojis because they were so angry with the global group that makes emojis. That global group won’t allow the Aboriginal flag to be an emoji because it’s not used globally. So there’s this rule that unless something’s used globally, it can’t be an emoji.
When people talk about bias, I say it’s not just about bias. It’s also about connections, capital connections, regulatory connections, but also these spaces, these “nourishing terrains,” as Deb Byrd Rose called them, of people creating ways of engaging with data and information that nourish the environment and also nourish people as well.
Jo: So, a short list of doors to kick in.
Marnie: There’s a bunch of doors in the corridor.
Jo: Or also known as rooms where I think I would like to see the upcoming generation of history PhDs participating in conversations: I think we need to be in the room where LLMs are being designed, and specifically in the room where people are thinking about what data is being used and how sources are being treated, and are there footnotes, and how do you trace back the facticity of any statement. Even calling errors of LLM analysis “hallucinations” is a curious choice. That’s marketing up the wazoo. That’s a euphemism.
Marnie: When historians get it wrong, we don’t say that they hallucinated.
Jo: We have another word for that: an error. And it’s devastating, and it needs to be repaired, and we take that very seriously. Historians should be the first to say, “You know, maybe you’re not persuaded that you should take this error in facticity about the Civil War seriously, but what about an error in medicine?”
At some point, the error becomes legally actionable. If you give me 1920s knowledge about medicine with the same credibility as 2024 information, that’s legally actionable. So those errors are a problem and temporality is important. The machine needs to be able to distinguish the evidence of yesteryear and the conclusions of yesteryear from the conclusions of today.
That is something that historians know a great deal about. Most computer scientists have not thought about how you figure out where the consensus is moving on the basis of textual documents, something we know a ton of.
A second door to kick would be the regulation of LLMs—where they can be used, how they can be used, what we trust them for. I’ve been in the room where people who are in charge of LLMs have said, “The LLMs are hungry; the market will take care of that.” And what they’re thinking of is a startup called Scale AI that was just valued at some enormous billions of dollars.
Scale AI has been using people in the developing South to crowdsource the identification of different kinds of documents. But that crowdsourcing of the guy in the café in Mexico, God bless him, getting paid minimum wage to comment on data artifacts, he doesn’t have the same kind of access or expertise about artifacts of language as the curator of manuscripts in the Bodleian Library.
The curator in the Bodleian Library has collections of metadata that took hundreds of years to compose and those are still useful. You have to be a ninny who’s never taken classes in the humanities to think that the market is going to solve this problem of what kinds of data the LLM is crunching, and that it’s going to get to absolute truth without some understanding of the variety of different potential voices and the bias of what landed on the internet in the years 2015 to 2018. It’s grossly biased.
We have huge holes in our history, which historians are trained to identify and speak to. We speak to the varying levels of consensus about what’s right and what’s accurate. Silicon Valley needs an education about this. Kick in that door and tell them.
“We also need historians in the room about the public understanding of tech: when to trust it, when not to; don’t listen to the spiel that personifies AI, or when it calls the errors hallucinations.”
We also need historians in the room about the public understanding of tech: when to trust it, when not to; don’t listen to the spiel that personifies AI, or when it calls the errors hallucinations. That’s important for talking to the media about. It’s important that the media understand the bias of these tools with regard to archives and the facts of the past. So we need interlocutors who can just talk about that. You don’t need a degree in tech. You don’t need a deep understanding of the LLM to get some sense of where it’s going to mess up the history. We need historians to be experts in the public understanding of the climate crisis, and then in the public understanding of climate governance.
That’s a long list, but the good news is that no historian has to do all of those. In that list—and I’m sure there are more areas—where if you specialize in environmental history or the ethics of history, you’re needed as a public commentator in a world which has moved so quickly towards tech and then ricocheted back towards English as the language of programming. There’s a room where your understanding of archives, your understanding of environmental history, is desperately needed right now to fill a vacuum of knowledge.
Marnie: Lots of people express their anxiety about AI. Fair enough. I think it’s incumbent upon all of us who work in this space to help people to feel welcome here, for the discipline to not feel like it has missed the bus and it therefore is unable to do anything. You’re never too late to come to the story.
It’s really interesting, Jo, what you were saying. Aristotle, in his rhetoric, talked about the dialectic turning on what he called endoxa. He said endoxa are the ideas of the majority—consensus—but endoxa are also the ideas of the wise. Now the problem is that we, in a way, and this is the structure of the internet, have collapsed the endoxa of the consensus with the endoxa of the wise. And we seem to have lost the endoxa of the wise in that space. But we’re not going to get to that by people feeling anxious that they can’t program, or that it’s already happened, and they’ve missed out, or that they will need to have particular technical skill sets, or that the tech folk aren’t interested in us.
AI has a history problem. Because of its inability to distinguish between what happened in the past and the present, I’ve looked a little bit at jailbreaking AI and the challenges people are having because the values of the past are not the values of the present, and people having to put safety parameters over the models to try and prevent people from using the models maliciously. So AI has a history problem, and so therefore we historians are needed to address that. But we have to see ourselves in that space, and that’s for every student to see themselves, but also for every professor to see themselves and not to think that they don’t have the skills to be in this space or the expertise.
Jo’s right: it’s about helping you to transform and transpose your frame and giving yourself more credit, for what you can do. We have perhaps, over time, not heard what we can do and it can be very anxiety-generating and disempowering for people.
You asked me the other day, Jo, who’s invited; everyone has to be invited. Everyone has to see themselves as part of this, and everyone can do something about this. I don’t think that AI is a monolithic tragedy either; I don’t see it as monolithic. There’s opportunity, and I suspect the pendulum could swing back as people figure out, as they already are, “I don’t like the terms of engagement: they dehumanise me. Or it means that you can scrape pictures of my family and just use those. Or that you can take my data, and you can track me.” People are expressing discontent about those things, as they should. My sense is the timing is good for people to express dissatisfaction.
What we need to get on top of is the well-ploughed route around the GDPR the EU’s General Data and Protection Regulation and the EU regulations around privacy that’s caught traction. And then there’s a whole separate universe around green AI, which is, how do I use AI in a way that minimises water usage and power?
Those two worlds have to be connected. It should not be a trade-off between me saying, “Don’t track my data, and I also don’t want the server or the cluster farm to be denuding the environment that it’s sitting in.” Those two things belong together. The opportunity is now to pull those two discourses together. Presently, there’s not as much on green AI. It’s only really getting there, but the regulatory stuff from the EU has shown people the opportunity of questioning how we broaden that discourse out. And that’s not traditional ESG [environmental, social, and corporate governance], where I’ve got a whole bunch of corporate measures and then I just carbon trade or whatever I do. I don’t think that’s ultimately going to cut it with a group of people around the world who say this relationship is dysfunctional and who are trying to figure out how to reset the relationship with technologies here.
AI as Archive and Historical Object
Mary: We’ve talked a lot today about AI as being a tool, as something to exchange with, as something to build relationships with. What about AI as being an object of history and telling its story?
Marnie: You’ll get a little bit of that today in my presentation, looking at algorithms, which have a very long history, but then saying there’s a particular moment around the ’60s where people start building search algorithms and figuring out how resource intensive they are. I will also talk about the rise of biologically-inspired algorithms, because we have boundless admiration for the way that ants work, and bees, and glowworms.
We’ve created this whole new digital bestiary, which is just incredible to me. It’s like the rewilding of AI in the most chaotic kind of way. So for me, there has been a kind of environmental turn in AI in the strangest possible way, a kind of swarm intelligence—all these computer scientists have been finding animal inspiration.
Maybe, cynically, unintentionally, whatever, there’s this whole complete, incredible universe of the biological organizing the way we think, generating results around the human past. And so there’s just this fantastic question about whether ants are writing history at the moment, which I just think is really, really interesting given that’s outside the boundaries of ethics. Insects are not inside the boundaries of our ethics.
So there is just an interesting historical moment where AI transposed itself into a space around heuristics in the 1960s and then biology comes to be seen as a major source of inspiration for these trade-offs. Now my question is, how do we come to a realisation about what that is ethically, and then reset that environmental history for AI in a slightly different way?
So yes, it has its own history, without a doubt. I just think it doesn’t realize it’s writing its own history. There’s lots of individuals in co-optive relationships with one another, producing all kinds of things in the AI world. And then some people have tried to make sense of that through swarm intelligence. I’ve said, well, maybe, maybe not, but even if that’s not the case, we can make sense of that, make people aware of that history and actually use it to craft perhaps a better future for AI.
Mary: Is AI a form of archive?
Marnie: Yes, it is. People have said to me, “But AI is very biased.” And I go, “As if humans aren’t.”
Let’s think about all the archives in the world that there are, even Hansard. I love Hansard. But, you know, the first woman to speak in the British Parliament was Australian. She changed herself to the grille in the Ladies’ Gallery. Muriel Matters is her name. She was protesting the fact that women didn’t have the vote. Women had the vote in Australia. They didn’t have it at the vote in the U.K. Right? So she’s in the House of Parliament and she’s chained to the grille and she falls through and she makes a noise, and she becomes the first woman to speak in British Parliament. Isn’t that the most awesome story you’ve ever heard?
Marnie: The Hansard archive is a very well-behaved and beautifully curated archive, but it’s got moments of complete rowdiness which we might not see in that space because we were looking at it for a particular view and then this one moment of rupture where a woman falls through because of architectural weakness takes first prize.
Archives had all, have all of these strangenesses about them. They’re never anything but that. The French revolutionary archives are incredible. Jules Michelet himself is fascinating—what he chooses to write in his diaries is obsessive around jellyfish and about his wife and all, but just really interesting obsessions that people have archivally, which means that you can never pretend for a moment that you have an objective full record of what we claim to be the case. That’s what makes them such fertile and nourishing grounds: they never come without an angle, but they can be read with other angles in other spaces, and that’s what makes them fabulous. So yes, AI is an archive, but it comes with all kinds of angles that you can read in lots of different ways.
Jo: That’s an amazing story.
Marnie: There’s a reproduction of the grille that she fell through in the South Australian Parliament. It’s not even the real thing; it’s just a reproduction. New Zealand was the first place to grant women the vote and South Australia was the second.
Jo: Women weren’t allowed in the gallery around Parliament. They had to be up in an attic tier behind the grille looking down. The wives and sisters of parliamentarians would end up there.
We were able in our research to find the names of some twenty women who spoke during the 19th century in Parliament, including Mrs. Disraeli. Others were women who were interviewed on the floor of Parliament after testifying before various commissions.
That is part of the value of data. It would have taken you forever to find those women manually because they’re needles in the haystack. If you go in with a research question, the data can be very supportive and nourishing. But it’s never one narrative of Parliament that you should get. If, n-grams only gives you one narrative, then it’s a broken tool because it says, “Here is the one true history.” A trustworthy practice of text mining for historical analysis obeys the rules of historical analysis, which say history is infinitely perspectival.
Marnie: For me, it’s this tension between all of that and public perceptions of history, which if you look at the American Historical Association surveys and then people saying, “No, there’s got to be one account: give me some certainty. Stop quibbling and arguing and give me the truth.” And so that’s the collision course we find ourselves in right now because AI scraped the natural language of people who perhaps are not comfortable with the perhaps and the maybes. It’s not just a battle for capital.
Jo: It’s driven by various political interests. If you’re talking about the legacy of the 1619 Project, the American public needs answers: “Was slavery the cause of disunion?”
Marnie: “Just tell me what happened.” So, if we want to put AI in perspective, it really is another moment where a professional group of people who ask, quite rightly, to put a question mark on things for a whole heap of good ethical reasons, finding themselves in a context where the question mark has been taken away.
Charan: We should probably wrap up. It was a lot of fun. I want to learn whose doors I have to kick some more.
Marnie: Just kick them all in, or actually Charan, just fall through the grille, right?
Mary Baxter, Charan Mandur, and Thomas Stroyan