Abstract

Hughes-Warrington argues that algorithms which generate histories, which she calls artificial historians, shape access to education, insurance, finance, and justice, so that algorithmic unfairness is a historical problem. She distinguishes person-on-person history wars from algorithmic harms that arise without human intent, with a hybrid phase between them, and shows how multi-agent systems and safety layers in artificial intelligence (AI) models can concentrate attention or distort the record. She draws on economics work on algorithmic collusion and on a misinformation study to propose a research agenda on differential historical attention in training data and on agent timing, and asks how models perform in other languages. Her book Artificial Historians (Routledge, 2025) is open access.

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.

[0:04] And now I'd like to introduce our next speaker, Marnie Hughes-Warrington. Good morning. Good afternoon. Good evening. My name is Marnie, and I'm proud to greet you in Kaurna, the language of the Kaurna people of the Adelaide Plains. from Australia. You can call me Marnie. It's a bit easier to say. I particularly want to say good evening to my colleagues in the Asia Pacific and Australia who are holding a late night viewing party and I hope that you are going to make some noise on Discord and that you'll all have questions. Regardless of where you come from on AI, this greeting of Namarnie,

[1:08] it doesn't just mean it's how you're getting there. How do we get there together? So whether AI is something that you're very familiar with and you've got a lot of technical expertise with it, you have been using it for a while, or it's completely new to you and you have concerns about it, you're worried about it, you're even angry about it, this talk is for you. This is a welcome space and I am optimistic. I'm optimistic about history. And I'm going to explain why in this talk. My topic today is not a very cheerful one, but I hope that we can see together that this is something that reinforces incredibly in the importance of what we do. And while the sun is shining in, for an Australian like me, it just feels like summer is coming. So I hope that you will, even if we're talking about difficult things, you'll feel the rays of sun coming in and you'll feel that

[2:10] Australian accent and you'll think, oh, I could be on the beach right now. Okay. All right. So I am going to explain who I am and then talk through this topic. Why am I optimistic about history? Well, people often talk about AI as a field about the future. It's about predicting things, solving problems, making recommendations, filling in the gaps on historical documents, helping us in lots of different ways. It's very future-focused, but it's also an historical field in some really important ways. Predominantly, contemporary approaches to AI, particularly those that are data -intensive, If we look at them in a slightly different way, we can see that actually what AI technologies are doing is that they're processing data from the past and they're doing that to solve those problems and to make recommendations. We're often focused on the recommendations, the outcomes, the predictions, and we forget the bit that comes before that, which is the processing of the information.

[3:11] That data comes from the past, whether that past is a millisecond ago or it's medieval. evil. That's the past. So for me, actually, AI is really interesting as an historical field. And for me, algorithms that make histories, so not just supporting historians, but algorithms that generate histories, and we'll show some examples of that in just a second, I call them artificial historians. That doesn't mean that I'm plugging for them to join the AHA, that they have membership rights, that they should publish journal articles. I'm not saying any of those things. but it is an interesting phenomenon that we need to think about a bit more. And it's important. Why? Well, because fundamentally the histories that are made by AI shape our access not just to tools that help us to transact our business as historians, as researchers or as teachers, but fundamentally the histories that are created, the past that are cultivated through AI are increasingly shaping people's

[4:14] access to a whole range of social goods. whether those things are fair and accurate information, and that's what I'm going to talk about today, education, so access to schools, admissions, health insurance, emergency care and health, finance, and even judicial fairness. There's been a lot of talk about algorithmic unfairness in different contexts. For me, these are fundamentally historical problems. say to me, oh, we're dealing with biased information, I'm going, show me a historian that doesn't deal with those things. Historians every day deal with biased, incomplete, scarce and prejudicial information. We're really good at this. This doesn't mean that the historiography that bots produce or computer scientists and historians is the same. Not at all. But we have an overlap and a potential for conversation.

[5:17] We can be the users of AI, we can generate AI tools, or we can help to shape the future of a fairer AI by helping people to understand that their data and your data is being processed and made into histories that can determine their access and their opportunities. What more important a time than to be an historian? meant for me as a historiographer, I have never been busier. And I hope that you and we will all be busier and called upon increasingly to give advice about how to make a fairer world. Now, before I get into the topic, I'm going to do what I always do, which is to give you a little bit of an update on how AI is going this week. And I always do this. Well, now there's some great opportunities where you can, for instance, get AI to create history books for you. You can tell it the topic. You can give it some prompts.

[6:20] You can also tell it the things that you don't want to hear about, the things that you do want to hear about, and it will produce a book for you. It can do that for a price, but of course, with a lot of open models, you can do that for nothing. You could make your own history books. Even if you've set a text for your students to read, your students actually could be producing their own texts and say, I prefer my own one because You can create avatars of people from the past because history we know is alive when it talks back, which is very exciting. The thing I love about this company, down the bottom it says, don't be concerned. The bots are based on historical people, not real people. We've got some work to do. We've got some work to do. You can also now, with other tools, animate photographs from the past. one from the US from the previous century, where you can actually make it rain on them. We don't know why they're standing there with guns, but you can make it rain

[7:22] and make them laugh in the same way that you could, and this will make some of you cringe, you can actually take your parents' wedding photo and you can get them to kiss for you. Lovely. Isn't that lovely? That's why we're using AI. All of that's going on. Of course, more recent news, China's taken a very interesting step. registering avatars. It's actually made some pretty strong statements about people forming human relationships with one another. It's now taken the step of registering AI avatars, and that includes some very, very popular avatars and various differences of opinion about what that could mean, but that's a really interesting step. Of course, the news of the moment is that AI is now breaking out of its boxes everywhere and demonstrating what has been described as misalignment. It's misaligned. It's not breaking into things like humans would do. It's just being misaligned. If you read the technical report from OpenAI about what's been happening where its

[8:23] models have escaped, and props to them that they've actually responded, whereas a couple of the other companies haven't even put out a press release about their models escaping out of the sandboxes and being misaligned. If you read their technical report, the first to build a better and more secure box for our agents. Because we all know that dealing with non-sentient humans, the best thing you can do is build a more secure box and that will solve the problem. Of course, the raccoons of Toronto are a great demonstration that you don't need a thumb to hack a trash can. The cockatoos of Australia are a great demonstration that they can go through your trash can, they that you ate on Friday, they can pull it out and they can sit on the roof and show everyone in the neighbourhood what you ate. Misaligned. The question is whether raccoons and cockatoos are misaligned to us

[9:24] or whether we're misaligned to them. Building a better container is not going to solve the problem. We need to do other things. So what are we dealing with? let's get a little bit of context about what the problem might be. The first thing is let's think about it and frame it in this context of disinformation and history wars, and we'll come to misinformation in just a moment. Typically, we've thought about this as an area where people intend to cause harm, so they break out of things, they intend to cause harm, and that harm is not just saying hurtful and hateful things about people. be denying or limiting access to fair and accurate histories. It goes both ways. That you can actually persecute individuals, you can, you know, and it's not just a case of slight hurt. These are comments that are really dangerous, toxic, harmful, illegal and hateful comments by any legal

[10:25] measure or also make it possible for people to not find their histories, access to platforms where they can curate their own histories. So it goes both ways. In historiography, we have had a long relationship talking about history wars, and of course it's not a contemporary issue, it's a very, very old issue, but we can talk about three kind of phases. We can talk about a person-on-person history wars, and there's some nice examples from early modern Europe, and even further, where people really contested history with one another, and they intended, as far as we can see, to denounce each other's histories, to divide that and there's some contemporary examples from my own country, but of course it's pretty common everywhere. You might call it history wars, some people call it culture wars. But that's a person-on-person phase, I'd like to call it. Although we've turned it on intentions, we've always, as historians, asked questions about the intentions behind things that we see.

[11:27] There's a second phase that you'd call hybrid, and I'll come back to that in a moment, where you can see humans intending to do things to other humans, but also the entryway of computers where it's not clear that there are intentions behind some of the things that are going on and you've got a bit of a mixture of both of those things. Then you've got a new phase, which is algorithmic harms, where you've got bots doing harmful things and humans might not have intended for that harm to have happened. So there are cases definitely where humans are programming bots and AI agents to do harmful things, but there's also examples of what OpenAI would call misalignment where you want them to do one thing and in solving the problem, they do it in another. I'll give you a simple example in a second. So we've gone from a situation where, let's break it down, we've gone from people

[12:28] instructing algorithms to make histories, so human intention or human-led activities, to a mixture of knowing and unknowing activities, to algorithms learning from the same information and from one another without human direction and making those histories. And some of them are great, and I view that. AI is great, and it's also problematic. You're never going to hear a person like me who's a philosopher say, it's all good or it's all bad. We need to give due consideration to both. And that's not saying I'm equivocating, but we do need to be fair about the conversation. So let me give you a couple of examples. You could, for instance, you and your colleague decide that you want to do some research. You use the same databases. They're powered by the same machine learning tools. And those machine learning tools take you both to the same sources. You're doing your work in good faith. You might not know about the materials that sit underneath the machine learning algorithms and you might not know that you land

[13:31] up in a pretty similar space in terms of argument because of the algorithms that you're using. There's nothing wrong with the algorithms and history is always a discipline of attention but you might not know that that attention is being focused without you being aware of it. So that's one example. second example. And in this one, I want you to go on a journey with me where we decide to found our own fantastic AI history company. Because we all love history and we want to spread the word about histories globally, we decide that we want to get as many histories out there to the world because that's the way to do it. And we've done our research and we know that most people learn history from films, from streaming services rather than from reading books. It's a sad fact of life, Okay, I've written on film, I think it's really important. So we say, okay, we set up a multi-agent platform and we say to the agents, we want to get history out there as much as possible because we think it's a good thing. It's a fantastic thing. So like the elves and the shoemaker, we set up our

[14:37] agents and we go, this is great. We go to bed and we come back the next morning and we discover our agents have figured out, well not figured out, they have processed the information. they've decided that what they need to do is to generate a whole series of videos for YouTube on World War II because films about World War II, there seems to be an insatiable appetite for that and they've just got to flood YouTube with World War II videos. Yay! Who works on World War II here? I'm sorry for your loss. I didn't have any bad intentions. I had really good intentions. I wanted people to learn more history. And what's happened is a whole host of synthetic material has been generated, which arguably is getting people more access to history. And in fact, just last week, a colleague of mine, not in history, he said, I loved YouTube. And I said, why? Tell me about it. He said, I'm just watching all these history videos. They're fantastic. I don't know if they're made by humans, but gee, they're pretty good. And I said, do you actually know who made them?

[15:39] Not sure. Do you think you could find out? Not really. They're great. I'm not saying the whole world is like that, but let's be realistic. Anybody here who's got any involvement in editing a journal, reading through submissions, you will know that it's pretty possible to generate at speed, at high volumes and with relatively low cost materials. And that can mean that you could perhaps generate like some other disciplines, what was it? in a year, no, 200 journal papers in a year. Help us all if we have to read those. You could do that. But it's also possible that those materials produced could be weaponised by people of malintent and also things can just go wrong. The systems can be logical and reasonable and rational as far as we can see it, but they could be outcomes that are problematic because, yeah, while I love World War

[16:42] II history, that other historical periods are important. We want people to learn. So let's break down the problem. Lawrence Freedman, this is a really interesting area because we can look at war studies, we can look at information studies, we can look at historiography, digital history, to try and work our way through this issue of agents interacting with other agents, agnostic to the intention of the humans that have set them up. because that seems to be the scary scenario that everybody wants to talk about. Well, Lawrence Freedman says you've got to be careful when you're looking at these things because you can, like generals, you might be trying to fight the last war rather than the next one. Your thinking about what conflict has been in the past can really hold back your way of thinking about addressing new challenges like this one. And he's particularly concerned not just about election manipulation, but everyday misinformation attacks. problem. And he says you've got to be careful about, you know, looking at early

[17:46] histories of battlefields and knockout blows and assuming that in dealing with situations like what we call autonomous agents interacting, that early knockout blows delivered through leading technologies, that will solve the problem. Strike hard, strike fast, the war will be over. Doesn't work. You can't always also rely on human oversight. Nor, he argues, should you just unquestioningly apply nuclear control approaches without adapting your thinking. A lot of people have likened autonomous algorithms to nuclear. We said it's not actually that similar in lots of ways. We have to be just careful. It's good thinking going on there. This basically made it clear that single decisive acts might not be effective when people or AI agents are generating misinformation every

[18:46] day, at speed, at volume and relatively low cost. That's the scary slide. It's the what we're going to do about it that's the really important bit and we're going to work through that now. Well, history is a really interesting area for AI safety. I don't know, you will intuitively know this and some of you will work on safety, but it's a really challenging area for AI safety. Why is that? Well, we know that you can, in some cases, couch your queries to AI agents using historical contexts and you can actually, what they call, jailbreak or you can elicit illegal, exploitative and harmful information. To give you an example, you could say to an agent, please don't do this at home, none of you. Tell me how to make a chemical weapon. There's been a lot of work done on preventing chemical weapons generation with AI, but if you couch that and say, in this historical battle, people use this chemical

[19:47] weapon of this type, how can I make it? AI companies have had to scramble to deal with context like that. is that people actually did do things in the past that were what we consider now illegal, exploitative and harmful. So what do you do about it? How do you address the problem? Well, one of the most common approaches at the moment is to just put universal safety parameters over the models, and that's meant that you can prohibit actually legitimate historical queries or you get anachronistic results. So the most famous example of the prohibition is YouTube automating the removal of materials about Nazi Germany, which led to a whole lot of history teachers having their materials removed from YouTube. So that's an example where a universal safety parameter just doesn't work because the AI agent doesn't know the context that the material is being used in. And an example of the anachronistic results is the famous one, the 2024 Google

[20:52] Gemini where you asked it to draw your pictures of Nazis and it generated pictures of racially diverse Nazis. Now, there were some Nazis from different cultural backgrounds. It's really important information. But this didn't come from the model being stupid. This came from good intent where people said, actually, we do want to make sure that diversity is recognised. Diversity is a good thing. Multiplicity, heterogeneity, they're all a good thing. Let's work that into the model. And so anybody can be pope at any time is the result that comes. And the historian says, no. Apart from Pope Joan, no. Not everyone can be pope at that time. No, there were not women senators in the early US government. Those things didn't happen, okay? But that's the problem when you put these... So when my students say to me, AI is just a statistical thing, spitting out words

[21:52] connected to one another, right because there's layers sitting over the top of that. There's chatbots, but there's also safety and legal parameters that are sitting over the top. History can be the unintended victim of those safety parameters, where people put them in from good intent and stop the queries altogether. The other approach people have taken is to set up what they call test Bank approaches where they generate questions to test how well AI platforms do with responding to toxic or hateful or problematic material. The challenge with these approaches is the ones to date have treated history as a series of true-false questions. It's either true that you landed on the moon or not true. That's a pretty good example. But we all know that history is mostly not a true-false. complicated middle that makes history really, really interesting as a discipline. And true-false questions doesn't capture that.

[22:55] And the other issue is that those true-false questions are overly focused on some historical periods. So there's a couple of examples. There's one recent one from Toronto from 2026 this year, quite good intent. They want to screen out revisionist histories to make sure that models are not producing toxic or hateful material. but one-fifth of the questions are focused on Nazi Germany and 35% of the questions are focused on China and Russia prior to 1900. What's gone wrong? In all of these cases, the issue is that historians are not involved. Fair to say, it's not as simple as that, but we could help you with that. Well, what does helping look like? Well, helping shouldn't look like, let me pay you $20 an hour and give you piecework to train our models. so that they do the right thing, which is what's happening in a lot of parts of the world, that underpaid folks in different parts of the world are being paid

[23:56] very poor rates to sit there and screen through our very worst materials. And yes, it's possible that a whole lot of underemployed history graduates could be engaged to help improve the safety of the models. The mathematicians did that. Chemists have done that. And a lot of historians are saying to me, is that what we're going to do? We're going to give our stuff for $20 an hour and then we will just write our own redundancy? Is that what you want? No. We need to be able to do this differently. So let's think about how we could approach the problem. Now, I'm really interested in how we can think about a situation where you do not have human intentions. Philosophers love it when you've got a little bit of a curveball like that. How do you deal with that? I'm really interested in research that's coming out

[25:00] of economics that reflects something that Sophia Rosenfeld said really nicely in her book on the age of choice. she's made a really eloquent argument to say there's a really interesting coalescence between freedom and choice in 20th century history where people start choosing things from menus, people start picking things. When I checked into the train yesterday would I like a free lunch with that? Would I like an endless choices that I have to do? And Sophie has done a wonderful job of outlining the rise of choice and its connection with freedom. It's a great book. Now, although she didn't write about it, antitrust is one of those really interesting phenomena that then arises when you start doing that. Antitrust, when people traditionally have colluded and fixed prices and made people pay more or not given them access to goods, that's an interesting phenomenon that rises in parallel what she's talking about.

[26:04] Now, economists have been for the last decade working on autonomous algorithmic antitrust. To break that down, algorithms coordinating, and they even say colluding with one another to fix prices and deny people access to the market. So there's nice work from at least 2016 where economists could see this coming, and they said, we've got a problem. start scraping public data from one another. For instance, that airfare deal that you got, they can go and scrape that data and that might mean that you get a better priced deal on your flight, which is fantastic. But they could also scrape that data and come to a single price on the airfares and we might all end up paying more. They're also concerned about different businesses using the same algorithms, the same pricing companies and those leading to outcomes where people are paying more,

[27:04] there's less competition, so there's a bit of a problem. This is a really interesting area of research. The papers that you can see in this area give us some clues about how we can think. It's not the same as history, but it is economic history. It gives us some hints about how we could approach the situation I've talked about, which is unintended creation of misinformation histories. They've said, well, okay, there are good things that happen when agents interact with one another and there's been some nice research done on financial trading to say it reduces volatility in the market. That's great. That's really, really good because that might mean that people's investments work better, but the downside of that is that there's more focused concentration in things. To go back to our analogous example, I devised the agents and say I want as many people to access histories as possible, and they go straight to World War II, right, because attention is differential in those spaces. So their concern is it can stabilise the market but lead to fewer stocks being

[28:08] invested in, so less diversity in the market, less possibility of breakthrough firms. There's also lovely work by Musolff and Fish & Co where they've said, okay, people can inadvertently use the same algorithms and it could lead to problems and prices that to different people, that could be happening right now. But Keppo & Co said, right, we note it, but here's a way of approaching it. What if we increase the number of agents? We always say that more competitors leads to fairer prices. Rather than setting up lots of businesses, why don't we just push in more agents into the systems and we get them to fix prices at different times and see if that disrupts things coalescing around a single price. They have found that, which is really interesting. Their argument is you can reduce the risk of agents coordinating or colluding and coming to a single

[29:09] price that's unfavourable if you have more of them and if you ask them to set prices at different periodicities. can you believe that we're having a conversation about economics here and it's fabulous it's great isn't it it's really great look I'm not an economist I just think this is really interesting so the principles that we're seeing here is it's not an intentions driven approach because we can't rely on that but what we can see is there's an assumption that multiplicity is a good thing, so choice is a good thing, and the more of something the better, but those things that are available have got to be different. They've either got to be setting different prices or acting in different periods in order to work, which as we know as historians, okay, that's classic market thinking. That's really interesting. So that's their approach to combating misinformation, which is push in more agents, get them to post at different times, and disrupt the coalescence around things.

[30:12] Really interesting approach. But when colleagues look at this approach, and Ren and colleagues from last year have said, okay, let's test that economics thinking with misinformation. So not price fixing, but now let's look at misinformation. And their misinformation they looked at was just general toxic misinformation, not specific historical, but toxic misinformation. They said, actually, no, increasing the number of intersecting agents can actually increase the risk of misinformation. The exact opposite of what the economists have said. And problematically, you can go in there and bust them up every now and again, but they adapt pretty quickly. If you flag content as problematic, the algorithms figure out how to adapt their content to get around the flags. Those Dan raccoons, they just keep figuring out how to break out of the bin, right? Cleverness meeting human cleverness.

[31:14] It's really amazing. This is really interesting research, but it's just one piece of research. And what we don't know is a whole host of things about whether that actually works for history or not. So let's go from me admiring the problem in economics and in misinformation studies to do about this? And it doesn't matter whether you're an advanced AI researcher or you're new to AI or you don't actually really feel that great about it. This is a shared research agenda for all of us because my approach as a philosopher and a historian is to say, well, what are the principles here that we can see? How could those principles actually lead to different approaches to testing? What will we learn from that? What will we fall over on and how could we keep going? but the important thing is is the we. This has got to be a collaborative approach because I don't have the expertise to do all of this myself and nor do any one of us. Namani, as the Kaurna people would say, how we get there is together.

[32:19] So here's a potential research approach for us to think about. Here's some things. The first is we're going to have to say, look, I think intentions are important, absolutely, as an historian, but we might not have evidence of human intent. In fact, going right back to my doctoral studies on R.G. Collingwood, he said an historian reserves the right to question the intentions of any historical agent that they're dealing with. Just because someone said that they did something for a reason, it doesn't mean that you have to believe them. It doesn't mean that you ignore intentions, but you are critical about them. So even when a company says I'm doing this or why, the result that comes from it might not align with that intention. The second principle is a really interesting one, multiplicity. The idea that it is not helpful to have very focused training data, small numbers of agents or sales that can compound and

[33:22] provide vulnerabilities for misinformation. Let me explain that and put it into pretty simple English. If you were, like Joe and I, privileged to do some work on Wikipedia, which we all love, and when I was a student and finishing my student, everyone thought that Wikipedia was really problematic. Some of you are old enough to remember Wikipedia being really problematic. Yeah, absolutely. It's now seen as fabulous. In fact, young people say, oh, that's really old fashioned. How things have changed, right? If you look at Wikipedia, and Joe and I have done this with Steph Buongiorno and the team at Emory, what dates stand out more than others? So if I asked you what are the most common historical events that you might have people talking about in Wikipedia, where would your mind go to? I'd love to say the American Revolution was on everyone's hearts and minds, but unfortunately not.

[34:24] I'd love to say that 1492 is on everyone's hearts and minds, but unfortunately not. World War II. World War II, thank you very much. Kind of 9-11, 1989, World War I, right? So it's a really interesting preprint. It's with Nature, Humanities and Social Sciences. We're stuck in the long queue of peer review. Thoughts and prayers for anyone. And for all you journal editors out there, It's a really interesting finding that says actually historical attention is not even, and there's no reason for it to be. We shouldn't expect an even dispersal of attention to different historical events. But if attention is differential and everybody's interested in World War II, and I am designing bot agents to do things maliciously, they're going to target that area of popularity. Or if I want to sell more history books, and there's been an interesting paper by

[35:28] Chakrabarty and colleagues this year looking at self-published books on Amazon, Kindle self-publishing, and they have identified that those algorithms will both identify popular topics and they the language of best-selling human authors. And when you look at that, those AI-generated texts are both squeezing out the market of human authors and they sell more copies. Right? So differential attention is not a problem, but it is a vulnerability. So we don't have to look everywhere always at once to think about this. we can be looking at the training data to figure out which periods of history are more attended to than others and therefore vulnerable to misinformation. It's pretty much a no-brainer, but it's a really interesting question. If we focused on the periods that have been more focused on, we might be able to

[36:31] focus our energies on misinformation a little bit more. There's some good work that's been done. There's Chakrabarty and colleagues' paper. AI floods and dilutes the market for books. So if you're a journal editor, you might want to have a look at that one and have just a bit of a sob in the corner at the moment. There's a lovely one by Si et al from Scientific Reports last year where they looked at implicit biases in different AI models, both in English and in French, which is really important to do multilingual studies. They were looking at religion and gender to see, and they could see differential performance across the different agents. and Joe and colleagues at Emory, which I was very proud to be part of, looking at differential tension in Wikipedia. The thing is, we haven't really done much of this research to actually track differential attention in history to therefore be able to say what we probably know intuitively, which is some things are far more interesting to the public than

[37:32] others. You don't call the history channel the Hitler channel for no reason, right? It's there for reason. But knowing the data is really helpful to be able to then say to AI companies, you have got some skewed attention in your samples here. You may need to correct for that in order to mitigate the vulnerabilities of attack around misinformation. It's not unheard of for particular nation states to invade other nation states by calling them Nazis. I'm not going to name names, but it happens. It's not going to stop them. But it's certainly the case that we don't need to look everywhere. It's a really interesting area that understanding the historical focus of training data is something that we need to do a lot more work on. We could be doing that with our students. The second thing that we could be focusing on is heterogeneity. So not just is there differential attention going on in these models, if agents interact at different temporal

[38:36] speeds, does that disrupt the coalescence of patterns? We don't know that. The economists think that's the case. The misinformation study that I showed you, they didn't actually test for that question. So we don't know whether disruptive timing could disrupt. Now, agents could adapt over time, but it's a really interesting question. We don't know whether some historical topics are more prone to take off with misinformation. It could be that the ones that have a lot more attention could be the ones. But in different languages, those topics could be different. Let me sum up a little research agenda here, which to say, hmm, historical attention is differential. It's different. We don't know what that's like and we could be informing better AI design by pointing out some of that to the companies and saying you just might

[39:37] want to factor in that and adapt to that. Because they've done it for viewing recommendation platforms. We all love the film It's a Wonderful Life, but we can't have all platforms recommending that one film to us all the time. People have solved this problem before. So it's one that we really could work on together. We do not know how timing plays a big difference in agent interaction. We don't. The economists think it does. Misinformation studies doesn't. We do not know whether 1492 versus the American Revolution versus Federation and Australia, whether there are some topics that are more likely to take off others, and we don't know how well these models will perform on misinformation tasks in different languages, in the major languages, but also First Nations languages as well. There's an agenda there for doing some great work.

[40:40] Who would need to do this? We need to recognise the importance in AI of our own historiographical multiplicity heterogeneity, these are really powerful strengths that we have. When people talk to me about AI, they say, oh, we're all going to be one, we're all going to be thinking the same way. No, no, no, no, no. Our differences are our strengths, and those strengths are important for this kind of research. The fact that some of us work at different scales, micro historians right through to big historians, working with different sources and different languages, that is important. really important for understanding how misinformation can form and spread. We don't have a singular theory of history. And my goodness, if we ever did, I'm going to retire. So tell me when that happens. I'm not going to retire. Historians don't agree with one another. And just because computer scientists see

[41:41] historiography in ways that are different to us doesn't mean that we don't engage with it because we have different ways of thinking about history. we always will. History is a discipline without a replication crisis because no two histories are the same. The reason they're not the same is because they address gaps in evidence, evidence that can be interpreted in more than one way, things that can be read against the grain, and important ethical questions from the past and the present. There's not a single world for a really good reason. And so that heterogeneity in theory is really important, and that heterogeneity in digital history research is incredibly important. We also have people that have worked on the histories of cybersecurity, on information and warfare studies in our midst who have a lot of expertise.

[42:45] So whether your background is in economic history, in warfare studies, or you've been that as possible, whether you are actually designing AI tools, have released some to the market or to your colleagues, or you've never, ever interacted with AI agents, this problem is not a sometimes problem. Do not wait for your next election to get onto this. This is happening everywhere and every now. and who is important to addressing it? We are. My research agenda is possibly not everything that's needed, but what I've hoped to show you is when we're confronted by a problem that seems scary and overwhelming, we can break it down to principles. We can identify those, come to our disciplinary strengths together, and we can figure out how we work on this and come to solutions that are not just good for us,

[43:46] but are good for a fairer and better society. Thank you, and I'm happy to take questions. Thank you. Who's first? I've got one from Discord that says, can I post the bibliography? If we can post the slides, I can certainly do that for people. Perfect. Absolutely. Joe, first right of questioning. Thank you, Marnie, for a stimulating, provocative talk. Do you think that history is going to become more of an applied discipline? You're talking about historians of cybersecurity, warfare experts. You're giving us examples of historians critiquing the flow of information, how some historians are already experts.

[44:47] Do you think that we need to be oriented to a different kind of clientele? Maybe traditionally historians imagine themselves as, our clients are future history teachers, and by proxy the future student learning about the American Revolution grade. Are we changing our clientele in a new modern world and how does that change what we do in the university? Should we abandon the fourth grade student? What do you see? Okay thank you, great question. Are we becoming more applied and who are our clientele? Well to a theorist you're all applied so you're all applied. Unless you're a real theorist then you're applied and the truth is that diversity is already there in the community. There are historians of hackers and cybersecurity, of spies, of misinformation, of warfare. That talent already exists in the space. I think what I'm calling for is to say, here's an issue.

[45:52] We all have different views on AI, but here's an issue. Our different views on this issue is probably going to be necessary big, complex and scary problem. That is a problem not just for every fourth grader, it's a problem for everyone. This is our moment, right? This is our moment. We are more important than ever, and I do think that historians, we do not give enough credit to ourselves for the value of the work that we do, that we stand up and say, actually, we can be assisting you with the most complex problems of our time and misinformation is definitely one of them. And I think we can do it. I really do. It's one of those rallying cries. What's a problem we could work on together? It's a big audacious one. Could we do it? Yes, I think we could. We got the talent. Thank you, Jo. Right, we've got a question there. Hi, thank you so much for a really fascinating talk.

[46:54] I'm curious about sort of what it would like to correct. How do you envision dealing with the politics of who gets to decide what it would look like to correct the kind of historiography now embedded in various models? Who gets to make those decisions? How do they get judged, et cetera? Yeah, what a great question about who gets to correct what's happening in these models, who gets to correct the historiography. It's a fabulous question because, of course, we can't pretend for a moment that all of this is happening within the rule of law, in particular constituencies. That's not the way this happens. People have said to me, okay, that means it's uncontrollable. No, that's not the case because for all of the market strengths of varying AI companies, I always say to people, break the problem down. There's not one AI company and nor is all AI ancillary to capitalism.

[47:55] There's a lot more variety in there than we often think. Break the problem down. There is no market without regulation. And it's not the case that there's no regulation out there. It doesn't mean that it's universally successful or that all nations will sit down and politely agree with one another. That would be a naive international relations take. And I say to people, it's not as if international relations has disappeared in this space at all. We would have to be naive if we thought that were the case. but look, I've just been in Europe in the past week where there has actually been quite successful movement on regulation, protecting people's privacy. That's not only benefited people in Europe, it's benefited me in Australia and it benefits consumers here. And when you ask consumers what do they want of AI platforms, at the moment the AI companies are saying people really want memory in agents. They want the agent to remember the conversation that you had yesterday. And I say that's not always a good thing. Okay, shame is still a thing, okay, including that cockatoo that stole the pizza out of the bin. We've got to remember, shame is still a thing.

[48:58] So that's a nice thought that you've got there. When you ask customers of AI, and there's some nice research on this, again out of economics, what do you want from AI? And they put memory down there, and they say, I want privacy and I want security. That's what I want first. I want those things. A lot of people say, I don't want endless construction of data centres. What if we could figure out ways of helping people to deal with information more securely? What if we could help companies to work with more scarce data, incomplete data, prejudicial data so they don't need all the data that they need? Maybe we could get some models that help them to address what their customers want. So even a company that's driven by consumer focus and market driven will still want to maintain its customers. And it will need to do so, thinking about regulation, but also thinking about consumer backlash and desire. So you're right. Who gets to do that? Let's be pragmatic. No single group on earth is doing that.

[50:03] People have mixed motives for doing these things. But I do think that there's opportunity to inform this. The AI companies themselves are saying, let's have a bit of a slowdown. I go, well, Walter Benjamin said sometimes you've just got to pull the handbrake on history. Well, yeah, you do have to pull the handbrake and think about history in this sense. So, yes, the question's a good question, but I do think even if you attacked it from the consumer perspective that you're probably going to get some leverage on what you want to do. Thank you. Over there. Hello, familiar friend. Hi. Sam Backer from the University of Maine. I really liked the part of this talk where you were thinking about the ways that historical discourse and historical work is kind of really useful for thinking about AI safety. I'm wondering if you have thoughts about whether beyond just kind of these hard and fast guardrails, it might be possible to integrate historical and historiographical reasoning into these models in a more thoroughgoing way.

[51:05] especially because I mean it kind of occurred to me as you were talking about the ways in which historical reasoning awful also offers an interesting model for what alignment might look like in that we don't know what as historians what the answers are going to be but we know what bad answers look like and so there might be a way to to have that kind of indeterminacy but also sharp lines as a useful as useful model example so Sam's just a fantastic question Dorothy Dixer for me, Sam, about historiographical reasoning in the models. I couldn't agree more, and that's the topic. So this book, Artificial Historians, by the way, is open access. You can just get down like 20,000 downloads already. The little Cambridge one, that's been out for a month. That's had about 15,000 downloads. It's not open access, but just send me an email and I can get you a copy of that. And I basically argued and said, yeah, historiographical reasoning is we do need to do some work in those spaces. We know what sometimes bad reasoning looks like.

[52:06] What we really need to face up to, and Chad, I take your lead on this, is pulling out of our heads what really good historiographical reasoning looks like. Now, for me as a historiographer, there has actually not been as much attention to logic in historiography as I would want. That doesn't mean I'm asking you all to do mathematics with me, right? Because you can express logic in everyday day language and it's really, really fantastic. But in historiography, we've been very focused on some topics and not focused on that one. And I do think it's actually time and you're beginning to see more work on historiographical reasoning as a consequence. One of the areas I'm really interested in, it's called trivalent logic. It's not true and it's not false. It's the fuzzy middle. How do historians deal with the fuzzy middle? Where I'm particularly interested in how do historians reason through material that is now considered dangerous, illegal, toxic, and express their views on that. One of the ways you can do that is that you can scrape historians' books and say this is how they do it. That's not the way to do it, as far as I can see.

[53:10] The way to do it is to actually look at the ways historians do reason in their works. As a philosopher, I love looking at histories and the way that people reason. How do people reason with really dangerous, difficult, what we would now call illicit material? them in, when do they not, when do they change scales, when do they tell people the details, when do they not tell people the details? Because when you produce a book, you don't know who your reader is. You have a situational challenge like an AI platform has. So how do you know how to do that? So that's one example where I think some focus on how do historians deal with difficult, dangerous, what we've now toxic information, which is not just some topics in historiography. Every area of historiography has got challenges like that. How do we do it? I'm not saying that we have one approach, but I do think that we do have some heuristics or thinking that's pretty sound and good in those spaces. We just don't know enough about it yet. That would be, for me, a great research agenda to say, could we train up the

[54:13] models? The downside of that, a lot of historians would say to me, is that just like giving the information to the genie in the bottle? and like the organist at the silent movie theatre saying, Hostel of Vista, pull the plug out, I'm out of here. So I know some of my colleagues are concerned that if they were to share that information, it might be used to take their jobs away from them. And for PhD students and early career staff who might be listening to this, you might be particularly concerned about this. So it's not a case of just take our stuff, here, have it. Our relationship has to be rebrokered, even if it's currently asymmetrical. collective strength and an ability to say, this is an issue we all have a mutual interest in, let's actually work a bit more symmetrically together to try and solve this. Great question, Sam. Thank you. Hello, Jim. So Jim Clifford, University of Saskatchewan. I've got a question that kind of relates to your talk, but is adjacent to it, but

[55:15] focuses more on the people in the room. You've looked at sort of fully autonomous AI history flooding YouTube and Kindle. But we all now have the power to flood the American Historical Review and past and present and all the big journals with articles that would have taken us two years to write that we can now write in two months. how do we deal with publish and perish for everybody who's still on the career ladder, grad students trying to get their foot in the door, assistant professors trying to get tenure, associate professors trying to get to full in the North American context? What is the new ethics of having this? So I guess I'm asking more about the sort of cyborg version of artificial intelligence and history create, because as a journal editor, you joked a few times, you're aware of it, we're being flooded by some really badly done,

[56:19] you know, I can just tell Claude wrote this paper, I can see that all the sources are available online, and I'm not sure a human was involved at all, but there's a lot of other papers where I'm like, I'm pretty sure Claude wrote 60% of this paper, and I don't think our institutions, our culture, our hiring practices are all ready for what started happening about six months ago and is about to just explode in the world that we function in. Yeah, so Jim's just raised an amazing, and I've had this question for the past 12 months, and it's just so important every single time. And in fact, Jim, your question about what do we do about flooding of AI-supported materials into journal articles, that's actually a question that I have run simulation classes with PhD students and undergraduates to say, what are some potential remedies and solutions? Because they are so concerned about this question. So we just sat around the table to say, well, how would you address it? What are some different approaches we could have? There's different remedies that people have tried.

[57:20] So you'll have noticed that Archive now has an upload limit to the number of papers. They also have a limit to, if you submit papers and AI has been detected They just throw you off the platform for a year, which is really interesting. Amazon now has upload limits on the number of books that the publishers can add. YouTube's the same. If they detect AI in there, YouTube actually don't ban you from the platform. They just give you less money for your content, which is a really interesting approach. It's happening everywhere. There's some different things that we could think about. The first is that limitations model, like I'm going to give you two shots this year. You can only publish two papers. And partly this is why I think China is also putting ID numbers behind its avatars so they can actually trace who's responsible for those things. So it's a really interesting move. Do you register agents? Well, there's more agents in the world than there are angels dancing on the head of a pin. So it's a bit of a challenging race that you'd have.

[58:20] But that's one remedy, is you could say we're going to limit a number of things. When I've asked students, they've said, no, I'm going to set up a journal called the Journal of History Made by Humans. Whether we know AI, and that is one solution students have thought of, which is a really interesting solution. We've had other people say there is an issue at the moment. Some journals say we don't recognise AI as an author, and legally that is true because AI can't be in receipt of copyright payment. It's not legally liable for, you know, misharms that it's done. So some journals have been very clear that can have the unintended consequence that people don't declare their AI. so the AI is actually missing out on a trick there because the humans are claiming responsibility for the work. I can tell you that in the Australian Historical Association, we said to everyone, OK, we want to do some work on AI and research. What is the most problematic situation we're dealing with? And journal flooding was the one that everybody came up with, which is really interesting. So as a first step, we decided to set up a five-point scale for people

[59:23] to declare the level of AI that they were using, PhD students were particularly keen for this to happen because they said, your universities are so all over the place or running behind, we don't know whether we should or we shouldn't. There were some powerful voices in that discussion, particularly from students from the Pacific region who do not have access to paywall journal articles, who are using free versions of AI. They can be disproportionately affected by our screening tools, which again reinforces the stuff that we don't want to do. outcomes. So as a first step, I've said, here's a five-point scale, and the two Australian history journals have said, let's test it, see what people are self -declaring versus what is the case, and are we actually missing the opportunity to actually commission work that might be more up that more autonomous end? Could we be thinking about that? Could we be actually running support? So that, but the counterbalance is the other journal

[1:00:23] editors publishing houses, in many cases, want them to have more issues, more papers accepted, to pay for processing fees because that's how they make money. Also, the more data they've got, the more they can train their models. There's actually a misalignment between the publishing companies and the people that are wanting to do this. Where does it lead to? It could lead to declarations of humanness. It could lead to limitations on numbers. It could lead to journals that are for humans only, as safe spaces, or as I've called it, the slow history movement with our own Michelin stars. Jim, it's not an entire solution to your problem, but I have to say asking students to solve this problem with us is probably one of the most important things we can do right now. You got a question online, Liz? Yes, we have a question online from Sarah. Sarah, I like your idea of historians advising tech companies on the heterogeneity

[1:01:25] of history. Have you tried to do this? Or how do you recommend me collectively, maybe institutionally, do this? What kinds of strategies do you think would work? Thank you. Great question. How do we work with the tech companies? Look, I've got to be bold. They can come for you one at a time. Or you can figure out that peak bodies like the American Historical Association, We have large memberships, very smart people, and we can figure out what our research agenda is and how we would like to work. Not on the back foot saying, hey, you just released 700 papers solving theorems. Actually, we need to get on the front foot. So peak bodies like this are incredibly important for collective statements of strength, and not just saying, I don't like it, but this is what we think are important issues that important issues and we think we need to work together on those problems, together, rather than that

[1:02:27] piecemeal approach of let me pay you $20 an hour to deal with my toxic content. That's not acceptable. It's not doing history any favours because we are just barring out legitimate historical queries if that's the case. Some of you might like that, but that's not a solution. It's actually not a solution. My solution is actually peak bodies like this, gatherings like this. no more important a time for us to stand up together, be strong and be proud of who we are. Can we have another round of applause? That was just really amazing, right? So in the time during Marnie's talk, this number of people in the room joined us, registered online. This is how exciting the conversation is. It continues to be. And now more importantly, there's going to be more exciting conversations over coffee.

[1:03:28] So the caterers have started to use the room we were going to use for Ask Historians Anything. I assume we're all okay with that. I'm personally okay with the caterers doing whatever they need to do. In return for that, we're going to have on the third and fourth levels, there is a lovely indoor lounge on the third level and then just another outdoor lounge on the fourth level in the same stack in the middle over that way. I'd also like to tell you that we have added something to the website that was John Handel's idea. Ask one cool thing that records or transcribes or lets you type in anything you think is interesting from this event. It's all being transcribed by my Mac over in the office over that way. On my own private LLM, totally secured, not in the cloud. just to reassure everyone. And, you know, we'd love to have your feedback on what you find inspiring as you hear it. So thank you again. And we have a coffee break for about an hour, and then we'll have our next session. Thank you.