1.1 Getting Started with Goals and Rules for History Making

‘Can I help you?’ is a wonderful question. It assumes that whoever or whatever is trying to help me has a sense of what I am trying to do and that they can assist with that goal. I might be trying to walk across town, or I might be wanting to make a history. Conversely, I need to decide if I will accept that offer of help. Importantly, accepting an offer of help from AI does not mean that I must be able to code or to use specialist software packages. Nor do I have to be a theorist of history. If you can learn to code, use specialist packages, or explore the theory of history, I encourage you to do so. We can, though, make a start at explaining the relationship between history and AI via learning from your experiences. Let me explain with an example focused on designing an artificial historian using a history assessment rubric.

Search for examples of a history rubric or a history scoring guide. You can conduct your search by looking at printouts you have, by lodging a query with an internet browser, or by asking multiple generative AI tools to make one for you. The result of your search will be charts or tables with criteria and achievement-level descriptors. If you had never seen a history before, would you be able to make one based on the rubrics you found? Puzzling over the instructions, you may gain a glimpse of what you are meant to do, but you might be unclear what the resulting history is meant to look like, let alone what you can do with it or whether it is good. You might discern, for example, that you need to generate an argument based on evidence but lack a detailed description of what that means. You might look to your experiences, for example, in arguing from evidence to find the best route to walk to a place, and I might look to my experiences in arguing from keyword search results to follow this instruction. The instructions very likely will not tell us whether either or both of our approaches were right, let alone good.

In a famous thought experiment, John Searle argued that being able to identify and write Chinese characters using instruction booklets does not make you a Chinese speaker, let alone someone who understands Chinese. Understanding Chinese is far more complex than that; it requires intelligence. His thought experiment was designed to raise questions about whether AI is intelligent (Searle, Reference Searle and Lee1980). Making a history is also complex, and you will have discerned differences of view in the rubrics I asked you to find. AI does not know what a history is or know how to write histories. Arguably, though, AI has not needed to know or to be intelligent to support historical research or to make histories for quite a while now.

Early approaches to AI in the mid-twentieth century turned on instructing computers with rules for problem-solving. Think of this as entering all the instructions in a history rubric. These instructions were expressed at first in mathematical or symbolic logic and later in many of the globe’s languages, or what are called natural languages. This work also recognised that problems (such as making a history) could be separated from detailed methodology steps. A key insight from this approach is that a problem can be solved using different methods. In simple terms, I can make a history using navigation algorithms or keyword search algorithms and opt for the one which solves the problem best with the least computing time and power. This insight was particularly important given the processing and memory constraints of early computers. Early computers, for example, used magnetic tapes to store data. These tapes stored data in chronological order of input and had to be searched in that order. It would take the invention of laser discs to be able to search data regardless of when it was stored. The use of graphics processing unit (GPU) chips as well as central processing unit (CPU) chips also meant that parallel processing of solutions for the same problem could be undertaken simultaneously, as well as the processing of a wide range of tasks. The efficient combination of problem definitions and methods for solving them remains important in AI today, even with ever-expanding processing and storage capabilities and future possibilities with quantum computing. I will explore some of the implications of this for history making in later sections on how artificial historians can help or can harm us.

For the moment, I want to emphasise that this instructional or symbolic approach requires you to describe a history with enough detail and certainty that a computer can execute it in the manner of Searle matching and writing Chinese characters. As we saw from our rubric example, however, defining a history is not a simple task. Nor can we even count on histories being clearly labelled as such: think of the last history you watched or read and check whether it even had the word ‘history’ in its title. We should therefore not be surprised that early examples of AI in the mid-twentieth century did not make histories; they solved agreed and describable (and thus computable) problems such as code breaking; moving pieces on a checkers, chess, or Go board; and locating inventory in a warehouse (Walsh, Reference Walsh2025, pp. 45–71; Charniak, Reference Charniak2024, chs 1, 4, 9; Stokel-Walker, Reference Stokel-Walker2024, ch. 5).

Yet regardless of whether the problem is codebreaking, winning a game of checkers, or locating an item in a warehouse, mastering each turns on either your or someone else’s past knowledge. Codebreaking seeks to exploit shortcuts such as knowing whether, for example, a weather forecast is the first message sent every day. Winning at checkers means knowing that you cannot bring a captured piece back onto the board. Finding an item in a warehouse means having information about how you stored things. Historians of AI tend to agree that it was the recognition of the value of experience in the 1980s that helped AI researchers to escape from a series of ‘winters’ or low points in its development (Li, Reference Li2023; Stokel-Walker, Reference Stokel-Walker2024, pp. 29–36; Charniak, Reference Charniak2024, p. 21; Walsh, Reference Walsh2025; Hao, Reference Hao2025; Swedin and Ferro, Reference Swedin and Ferro2025; Adams, Reference Adams2025). Another way of seeing this history is that AI took a turn towards learning from the past.

1.2 Learning from the Past in Artificial Intelligence

Advances in AI accelerated at the end of the twentieth century through an appreciation of the past. We can see this, for example, in the broad use of backpropagation approaches in machine learning algorithms. I am going to explain backpropagation in simple terms by returning to our example of history rubrics. If you have never made a history before, you might look through examples of history rubrics and notice that many of them include the expectation that you acknowledge your sources. If you then dutifully create a list of references without any other text and submit it to be assessed, you will fail at the task and be given the feedback that you are far from being successful. If you go back and check the rubric, you may notice that you also need to summarise the key points in those references. You do that and add the summary to your list of references. You still fail, but you might be less unsuccessful this time. You keep refining what you do based on the feedback you are given on what you did wrong and whether you are getting closer to solving the problem. This involves adjusting the emphasis or weight you place upon each of the criteria. This may be a familiar story to you if you have studied or taught history. We learn how to make histories via help in identifying what we have got right and what we have got wrong. Another way of saying this is that we look to our past and propose actions for the future. Without feedback, our chances of being able to make a history are so infinitely small that we might as well give up. It is also the case, though, that we might get there faster if we are given more feedback and examples of histories to consult. This, very simply, is how neural networks function in machine learning. They use backpropagation – recursive adjustments from feedback and examples – to iteratively close the gap between a problem and proposed solutions.

In history classrooms, teachers give help and feedback to improve their students’ history-making abilities. In AI, help is given and iterative improvement is achieved through examples in training data. If you wanted to solve the problem of making a history, you could use examples of history rubrics as your teacher. We would call this a training set of history rubrics. You could also, though, use examples of histories as a training set. How many examples of histories might you need? This is an important question because we want to make a history without succumbing to overgeneralisation or overfitting. You might, for example, conclude that all 800-page books are histories (overgeneralisation), or you might think that only 800-page books on the Second World War are histories (overfitting).

One way of reducing the risk of overgeneralisation or overfitting is to use a larger training set of histories. Those histories might be licensed from publishers or scraped or extracted from websites. Our history of interacting with digital devices, though, also contributes to training algorithms to learn from the past to make histories. If you have ever typed a distorted string of letters and numbers to access a website, you helped Google Books to decode over forty million books from the past, including histories. If you have rated a historical book, film, or game, you have contributed to its recommendation as a history to others. If you have ever highlighted text in a history eBook, you have trained artificial historians to develop rules on how to make a bestselling history. More broadly, every cat video you have liked, photo of a historical site you have taken, traffic light you have selected in a CAPTCHA (Completely Automated Public Turing Test to tell Computers and Humans Apart), and every website you have visited are contributions to training sets at the local, national, and global levels. Your cursor moves, clicks, pictures, words, and sounds tag phenomena in ways that help algorithms to search for answers, choose the best moves, follow rules, and compute the probability of events. Examples include the National Institute of Standards and Technology (NIST) Special Database 19, a set of over 800,000 handwritten characters, and ImageNet, a collection of over 14 million images which have been labelled or metatagged by humans (National Institute of Standards and Technology, 1995). These datasets and others like them have been used for a broad range of purposes, from instructing machines on how to direct mail in sorting facilities, to training robot vacuum cleaners to avoid walls, furniture, and animals.

It is important to remind ourselves that datasets like these are archives – collections of information or records from the past – and that they can be useful for understanding phenomena in and across time. As Walter Benjamin argued, the way that we perceive phenomena is historical. This means that our ways of looking at archives may change over time (Benjamin, Reference Benjamin[1931] 1972; see also Crary, Reference Crary1999). Alix Paré, for example, has traversed the long history of depictions of cats in art; Pietr Tryjanowski, Michał Beim, Anna Maria Kubicka, and colleagues have undertaken a big data analysis of how animal road warning signs have varied globally over time; and Heather J. Jackson has documented the different ways in which readers have annotated texts (Paré, Reference Paré2024; Tryjanowski, Beim, Kubicka, et al., Reference Tryjanowski, Beim and Kubicka2021; Jackson, Reference Jackson1992). Moreover, the makers of ImageNet have acknowledged that the way that people labelled pictures of faces at the beginning of the twenty-first century might not be as fair as the labels that are used now. Recognising the need to improve labels, ImageNet revised the dataset in 2023 (Yang, Qinami, and Li, Reference Yang, Qinami and Li2020; Li, Reference Li2023). This last example helps us to understand why there may be many histories on ostensibly the same topic. New evidence may be found about a phenomenon from the past, but our ways of looking at, asking questions about, and dealing with the gaps in archives can also change and can reflect technical possibilities. AI has made it possible to look at data from the past in ways that were not possible or time intensive in the past. Ways of looking, asking questions, and dealing with gaps, however, also reflect ethical decisions and judgements. Histories turn on decisions and judgements about what matters and how things should be described.

Human judgement and decisions are not always at the forefront in AI. It is possible for algorithms to group and label phenomena without human annotations, or what is called ‘supervision’. An algorithm can, for example, cluster or group history rubrics by looking for common and distinctive features, in the same way it can identify patterns in reader highlights in history eBooks. Unsupervised machine learning can lead to overgeneralisation or overfitting and thus to results that are at the least trivial and at the worst unethical, but it might also identify groupings that are both novel and helpful for the makers and audiences of histories. Some of these AI-led processes identify patterns by plotting data in multiple dimensions and identifying the angles that can be used to discriminate groupings and outliers (Ananthaswamy, Reference Ananthaswamy2025). Regardless of whether AI is supervised or unsupervised, though, it keeps learning from the past as we keep learning from the past. The history of how AI has learned from the past is yet to be written and provides an outstanding opportunity for further research.

1.3 Training and Learning from the Past for Artificial Historians Such as Todai Robot

In the small space of this Element, we can make a start by looking at some examples of how AI has learned from the past and made histories over time. We begin in Tokyo in 2017, where Todai Robot passed the history component of the entrance exam to Tokyo University (Arai, Reference Arai2017). This is a good example for us to consider because it links back to our discussion on learning to make a history from rubrics. It also gives us a useful glimpse into the development of machine learning and generative AI over the past decade. Todai Robot is what we would now likely call a small language model. The model was trained with a small set of mandated history textbooks, a world history glossary, past exams, sample question and answer sets, examiner reports, and a Japanese thesaurus. Some of that training was a case of simple matching where the wording of questions and answers were the same across the materials. In other cases, though, the wording of the questions varied. Variations in wording can be addressed through human labelling of training data. Consider this example of a question from the exam:

We can re-label this as a request to find one true statement. In the same way, we can also label all the sample answers and all the sample sentences in the mandated texts, but that is laborious and unnecessary. We can approach making sense of this data in a different and less laborious way. To explain, let us look at the potential answers to the question posed above:

1. 1. Ouyang Xiu and Su Shi were the best-known writers of the Tang dynasty.

2. 2. Yen Chen-ching was the best-known chirographer in the Song dynasty.

3. 3. Wang Anshi of the Song era implemented the reform called the New Policies.

4. 4. Qin Hui fought with the war hawks on the relationship with Yuan. (Kanayama, Miyao, and Prager, Reference Kanayama, Miyao and Prager2012)

Number three is the correct answer. Rather than asking Todai Robot to work sequentially through each of the words in these four statements – which would be computationally expensive and slow – the team programmed it to give more weight to matching proper nouns from its training set, which I have marked in bold text:

The model then identified the likelihood or probability of these proper nouns appearing together. Another, more technical way of describing this process is as a search for the probability of entailment (Kanayama, Miyao, and Prager, Reference Kanayama, Miyao and Prager2012; see also Kawazoe, Miyao, Matsuzaki, et al., Reference 49Kawazoe, Miyao, Matsuzaki, Nakano, Satoh and Bekki2014). Probability is critical to AI. In this case it supports the identification of patterns in data, even when the wording of statements, questions, and answers vary. The patterns might not be 100% certain, but they are good enough for us to complete the task of identifying one true statement and for Todai Robot to do well in the entrance exam. Probability is a useful approach to dealing with uncertainty, and Aviezer Tucker has made an excellent case for considering Bayesian approaches to probability in historical research. Probability, he notes, can help us to explain the iterative inferences historians make from evidence – a kind of backpropagation – and to develop computational approaches to reasoning about the past (Tucker, Reference Tucker2025). Not everyone agrees, though, that Bayesian probability – or probability in general – is the only or the best way of thinking about certainty and uncertainty in history making. We will return to this topic in Section 3.

Weighting and probable entailment are components of the transformer architecture (2017–) that is now used broadly in generative AI tools. These tools offer a significant advance on the already impressive achievements of the Todai Robot development team. They can be used to produce short or extended responses to history queries very quickly. Transformer architecture breaks questions and statements into tokens. These tokens may be words, parts of words, or multiple words. The tokens in a text are analysed in relation to one another, and weighting or attention is given to those tokens most important for understanding a sentence’s meaning. In our example above, relationships between proper nouns are helpful for identifying the answer to a question. The tokens of a text are then compared to very large training sets of word relationships or vectors over time. This data from the past is used to predict the probability of words appearing near or next to each other in questions and responses (Vaswani, Shazeer, Parmar, et al., Reference Vaswani, Shazeer and Parmar2017). Guiding this process are parameters – the weights or biases – learned by the model over time. Generative AI tools use very large numbers of parameters. A temperature parameter, for example, can be tuned to select more or less probable words from vectors for responses. A top_k parameter can then be applied to limit the number of probable words that are considered.

Transformers decode and generate text more quickly than Todai Robot because they look at words in texts simultaneously, rather than word by word in sequential order. Moreover, they look at the relational vectors between words across sentences and across paragraphs, as well as words that appear next to one another in sentences. We can think of transformer architecture therefore as enabling dynamic scaling between smaller and larger word relationships. They may also save time by looking at a subset of data if a prompt is on a popular topic. GPU chips mean that large volumes of vectors can be looked at simultaneously and that a very large number of parameters can be developed, applied, and revised. ‘Large’ in some cases means tens of billions of parameters. Transformer text technologies are developing so quickly that descriptions of how they work are often out of date. It is therefore worth you looking to see what the latest developments are.

So far, my examples have focused on written histories. As surveys by Roy Rosenzweig and David Thelen, and Pete Burkholder and Dana Schaffer have reminded us, most people connect with histories in visual formats (Rosenzweig and Thelen, Reference Rosenzweig and Thelen1998; Burkholder and Schaffer, Reference 45Burkholder and Schaffer2021). Transformer architecture can also be used to generate still- and moving-image histories, as well as games and sounds. Indeed, large multimodal models (LMMs) can generate many or all these formats. Large vision models work by breaking training set pictures down into patches, giving weight or attention to each patch, and then predicting a response from patch vectors (Dosovitskiy, Beyer, Kolesnikov, et al., Reference Dosovitskiy, Beyer and Kolesnikov2021). AI Time Machine (MyHeritage, 2023–), for example, can generate pictures and avatars of you in different historical settings and animate your family photos. Paul Trillo’s short film Thank You for Not Answering (2023), to look at another example in more depth, shows us what is possible for AI-supported moving-image histories. The film explores a man’s memories as he leaves a voicemail to someone in the past. Trillo wrote the script and then generated still images using the tool Stable Diffusion (2022–). Using the still images as storyboards, he then described the animation – including camera angles – for another tool, Runway Gen-2 (2023) and connected the animation sequences together. The film is an impressive achievement. It includes some distortion of human faces and hands, and continuity across scenes is clearly challenging, but it is a matter of time before feature-length AI films set in the past appear. As Zhang, Yu, Min, and colleagues have emphasised, AI can be used now to generate everything from scripts and story boards to AI voiceovers and moving images (Zhang, Yu, Min, et al., Reference Zhang, Yu and Min2025). So too, AI has been used to support everything from script and map generation and route finding to artificial opponents in historical games such as Ara: History Untold (Oxide Games, 2024), Crusader Kings III: Road to Power (Paradox Studio Development, 2024), Thaumaturge (2024), and Ghost of Yōtei (Sucker Punch Productions, 2025).

It is important to remember, though, that there is more to AI than generative AI. Recommendation systems for streaming or online shopping platforms work with your data – as an individual and as a part of groups – to make predictions via machine learning about a broad range of activities, including what you might watch or buy. Recommendation systems sometimes place people in groups (which they do not know about) to mine their collective and individual data. If I have seen most of the historical films that you have seen, for example, then it makes sense for AI to recommend the ones that you have seen but I have not, and vice versa. Historians also look for patterns and anomalies in the activities and ideas of groups, and they may also group individuals without their awareness (if they are living) to advance that work. Recommendation systems draw upon conditional logic (e.g. If you have seen the movie, then I might like to see the movie) – which has a long history of discussion in philosophy – and, as Daniel Woolf points out, approaches to lessons from the past (Woolf, Reference 54Woolf2026). Meg Jones’ history of internet cookies, like the work of Schneier (Reference Schneier2015), highlights just how much data is collected about us as we connect to the digital world (Jones, Reference Jones2024). A closer look at cookies shows that they are both blocks of data that enable websites to load quickly or to keep you logged on and also records of your digital tracks. Cookies are used by a range of organisations to personalise advertisements and information for you. Depending on where you live, you can get a reasonably complete picture of your cookies by accessing your internet browser settings. That browsing archive may – depending on whether the cookies have expiry dates – extend back over many years. Jones has looked at the legal and privacy issues that cookies raise, but it is also possible to analyse them as histories and to make recommendations about your rights to know, shape, and forget your experiences from the past. Extending her thinking, it is worth remembering that your digital tracks may be used to identify anomalies that suggest a potential cybersecurity threat. Your internet browser may, for example, list whether any of your passwords may have been accessed in an unauthorised way. This is an important reminder that deciding whether AI technologies are good or fair can be complicated. You will likely want your information to be protected, but you might also want websites to keep you logged on. Moreover, philosophers such as Onora O’Neill have highlighted that your right to privacy does not also give you the right to use anonymity to harm others through hateful or violent speech. Her point is that our valuing of privacy ought not to come at the cost of harming others (O’Neill, Reference O’Neill2022).

1.4 Designing Your Own Artificial Historian

Even this small list of examples suggests that thinking about histories is helpful for AI development. If you wanted to make an artificial historian, you must make a few decisions. It is also worth noting that history making can be supported by AI to varying degrees. Transformer architecture, for example, is now embedded in the spelling and grammar checking functions of word processing software, and it is not unusual for bots to check the spelling, formatting, and references of histories that are being published or graded. Bots can also check to see whether text has been copied without permission, or images or data have been manipulated or reused repeatedly without good reason. You can use a wide range of free generative AI platforms to create your own document checking and formatting bot and to plug it in to your word processing software. This may be particularly helpful if you need to use different referencing formats for journals, books, or classes. AI in this minimal sense supports us in following customs and rules for making histories. It is hard to avoid these minimal AI supports for history making now, and we might not even be aware of the level of AI support in software and applications that we use every day.

AI can also, though, provide more extensive support. It can help us to search for patterns and anomalies in data from the past; translate and transform data into different formats (e.g. text to moving images or data to text), generate, summarise, edit, or extend questions, text, images, sounds, and code; and adapt histories for different audiences. You can, as Paul Trillo has done, create a short moving-image history by generating text, converting that text into storyboards, interpolating those storyboards into moving images, and compiling and editing those moving images into a sequence. AI can even generate a voiceover for you and provide a transcript in many of the globe’s major languages.

Moreover, histories can be AI directed and generated. Artificial historians work pervasively around the globe. They collect, cluster, and label datasets generally without human input, and search for patterns and anomalies using rules which may derive in part from non-humans. Search functions, for example, may operate in ways inspired by the activities of ants, locusts, or bees collecting food. These activities support recommendations, predictions, and even decisions that can affect your opportunities for shopping, viewing, education, healthcare, financial support, and even justice. Da Silva, Silva, De Arruda, and colleagues, for example, trained a recommendation system to predict bestselling books published from 1895 to 1923 with 75% accuracy (Da Silva, Silva, De Arruda, et al., Reference da Silva, Silva and de Arruda2024). Grobys, Kolari, and Niang have shown that machine learning from past hedge or investment fund data generates better predictions and thus financial returns than funds managed with more human involvement (Grobys, Kolari, and Niang, Reference Grobys, Kolari and Niang2022). In yet another example, Mahawar, Rattan, Jalamneh, and colleagues have used machine, ant colony, and artificial bee colony optimisation techniques to predict the academic performance of computing students (Mahawar, Rattan, Jalamneh, et al., Reference Mahawar, Rattan and Jalamneh2026). Journals and the internet are now brimming with examples of how data from the past can be processed with little or no human supervision to guide human activities at personal, local, regional, and global levels.

Thinking about what AI can do is helpful because it may extend our thinking about the possibilities for making histories of AI, human uses of AI to make histories, and histories made by AI. We must remember, though, that using existing AI technologies to make histories is not necessarily the same as designing AI to make histories. History makers need not be on the receiving end of technological change; they can help to drive it. Assael, Sommerschield, Cooley, and colleagues, for example, have used AI to date and place Roman inscriptions, even when they are damaged (Assael, Sommerschield, Cooley, et al., Reference Assael, Sommerschield and Cooley2025), and Guldi’s lab has created a web-based AI-supported application to help people to search and analyse records of the US Congress, among other sources (Guldi, Reference Guldi2023b). Writing in broader terms, Esamagu, Wazhi, and Adeyinka (Reference Esamagu, Wazhi and Adeyinka2024) have highlighted ways in which multiple AI tools can support research in African history and bring different humanities and social sciences disciplines together.

If we shift from using to designing AI to make histories, we need to spell out what we want to do through design requirements. Requirements both enable and constrain the technical solutions we might want to use or to make. Conversely, current technical solutions can constrain what we want to do. Drawing on the list of activities I have described so far in this Element, here is a start at design requirements – which we might also think about as principles – for an artificial historian:

• Histories can examine the past for multiple purposes.

• Histories can be made with multiple sources from the past.

• Histories can be shaped by different forms of logic or methodologies focused on the past.

• Histories can be made in multiple formats or media.

• Histories can persist for varying lengths of time.

• Histories can be updated, extended or transformed.

This list reminds us of a salient point about histories; they are made all over the globe for different purposes and from different sources. They may have different formats, and they may be treated as perennial or as ephemeral. An artificial historian designed to meet these design requirements may be described as an artificial general intelligence (AGI). It would be able to demonstrate learning from the past across AI technologies. Togelius (Reference Togelius2024) has looked at AGI from the perspectives of psychology, animal behaviour, and computer science, but there is also an opportunity to think about it through the key lens of history making. Research from robotics suggests that an AGI might be trained effectively if it encounters a broad range of phenomena over time in its testing phases (Brooks, Reference Brooks1999).

One of the challenges in designing an AGI, though, is that each of the requirements – and examples of histories which may demonstrate them – may be subject to dispute and even conflict. People might not agree that history making has multiple purposes, for example, and argue for a particular purpose. This is not different from debates about older forms of history making. What may be different in this case, though, is that technical decisions in favour of one view of learning from the past over another in designing an AI tool might not always be explained to users. This may generate questions of trust, as well as ethical and technical questions about whether other options are preferable. We will return to these matters later in this Element. For the moment, we can note that there is another option available to us: to narrow the requirements to meet a particular goal.

We can begin by thinking about who might use our artificial historian, and why. Who might it help, and how? We might build on our insights on history rubrics, for example, and design a chatbot which helps students with their questions about their assessment tasks. In a very different example, we might want to use AI to help people in palliative care to make histories about their lives. You may have your own views about where AI might be useful in making histories. Our thoughts about what we might do, though, are not without limits. There may be technological limits to what can be achieved at this time, or what you want to do may be expensive – in terms of computing time, money, or environmental impact – or out of the reach of people in your community and across the globe. Moreover, we might not agree that we ought to use AI to make histories in particular ways, or from specific sources, or for specific purposes. Communities do not always agree for good historical, cultural, and economic reasons that their knowledge should be used in ways that they do not know about or have a say in. The point is that not everything is possible or desirable in designing an artificial historian.

2.1 Scaling Histories Up and Down with Artificial Intelligence

AI may help history makers to realise the ambition of creating works that scale dynamically. Histories explore the past through different spatiotemporal scales. They can be made about moments in tiny places or range across billions of years in the universe. Up until now, though, we have seen only glimpses of how history makers may be able to shift between small and large scales and produce works that advance our understanding of the strengths of each scales and of both in combination (Magnússon and Szijártó, Reference Magnússon and Szijártó2013; Hughes-Warrington and Martin, Reference Hughes-Warrington and Martin2022; Margócsy, Reference Margócsy, Gänger and Osterhammel2024). AI is strengthening our sense of how this might work on account of advances in processing and analysing large volumes of data from the past. Kuzey and Weikum, for instance, have developed a backpropagating AI neural network to make on-demand timelines from English-language Wikipedia articles (Kuzey and Weikum, Reference Kuzey and Weikam2012). Their tool helps readers of Wikipedia to switch between articles and large-scale visual timelines, and vice versa. It provides users with the opportunity to place articles in context and to check for conflicting chronologies. Making the tool was a significant computational task, as English Wikipedia now boasts just over seven million dynamic articles made by public contributions, editors, and artificial historians that check, format, and protect entries from attack (Wikipedia, n.d.; Apostolopoulos, Reference Apostolopoulos2024).

Dynamic scaling is particularly useful because it can help you to detect relationships between historical phenomena and changes in those relationships over time. A good example of this can be seen in the research of Silcock, Arora, and Dell (Reference Silcock, Arora and Dell2023). They developed a backpropagating neural network to detect and visualise variations in the headlines written for the same underlying newswire articles in local US newspapers from 1920 to 1989. They were also able to track these variations over time by newspaper and by geographical area, and to make predictions about spatiotemporal information when it was missing from articles. They enabled users to look at the level of individual articles and at national patterns over time. In other words, their approach to AI supports the consideration of articles on specific historical phenomena in depth and in broad spatiotemporal context. It also gives us insight into the possibilities of research which looks for shifts in language to better understand the dynamics between individuals and groups. Historians have long been interested, for example, in detecting when group discussions might presage acts of violence or turning points in social and political change. Large-scale datasets and the AI tools to analyse them provide further support for this work. So too, AI-supported research may help to explain turning points in environmental, economic, and cultural change at both micro and macro scales. Advances in AI-supported historical research also prompt thinking about how scale shifting might be best understood by the audiences of histories. Imagine a future – likely not too distant from now – in which readers can move between words and visualisations in histories in ways which support their interests in understanding of events from different spatiotemporal scales. Examples like these remind us that histories have been made in different media and formats over time and that we should expect them to keep on changing in the future.

2.2 Acknowledging, Interacting with, and Grieving the Dead

One of the arguments made against scaling histories is that larger-scale views may obscure the contributions of individuals. People fear that informative details and ethical insights might be lost (Hughes-Warrington and Martin, Reference Hughes-Warrington and Martin2022). Findings from the moral machine project suggest, however, that we may need to give further thought to our association of smaller scales with ethics. The project collected and analysed 40 million decisions in 233 languages in response to protection from harm scenarios. Users were asked, for example, to decide whether they would prioritise saving individuals over groups in life-threatening emergencies. They found that prioritising individuals over groups is not universal and that some participants saw the age and identity of individuals as important in guiding their decisions (Awad, Dsouza, Kim, et al., Reference Awad, Dsouza and Kim2018). History makers engage in ethical decisions of this kind when they research phenomena in which people were harmed. They may choose to write about a tragedy at scale, to focus on a group, or to prioritise telling the story of harms on older people or on children or babies. In practical terms, though, they often work with multiple scales to advance understanding of a wide variety of phenomena from the past. They may focus their work on a small number of phenomena, but their writing tends to scale up to provide audiences with a sense of the context for those phenomena (Hughes-Warrington with O’Brien and Martin, Reference Hughes-Warrington2025). Their work can inspire the makers of the moral machine, as well as those who develop and test ethical thought experiments that include scenarios which include large-scale harms and harms at multiple scales.

There are good examples where AI-supported research has supported broad understanding across multiple scales. Attention to individuals is evident in work to create AI historical avatars, and attention to larger scales is seen in their contextualisation in exhibitions. The ‘Hello Vincent’ exhibition at the Musée d’Orsay (2023–2024), for example, provided people with the chance to chat with Vincent van Gogh as he painted Wheatfield with Crows (1890) and to find out more about the context for his work, and the Pradhanmantri Sangrahalaya’s 2025 debut of an avatar of Prime Minister Sardar Patel (1875–1950) is part of a broader exhibition on India’s leadership legacies (Pradhanmantri Sangrahalaya, n.d.)

Some small-scale AI histories may have mixed ethical outcomes. We cannot say, though, that they are harmful in a blanket sense. We gain a good sense of the ethical complexities of AI-supported history making by looking at the example of digital ghosts, which are also called ‘thanabots’, ‘deathbots’, and ‘griefbots’. It is worth thinking about their benefits, even if we decide that they should not be made on ethical grounds. Multiple organisations now offer services to capture and animate histories before and after people die. HereAfter AI (2019–), for example, captures oral histories which are then used to make avatars constrained to that material, whereas You, Only Virtual (Reference You2020) ‘versonas’ generate both archivally based and novel responses to prompts. These two approaches to the creation of digital ghosts can also be applied to those who died long ago. Ethicists have raised important questions about whether digital ghosts violate the remains of the dead and whether they have a pathological effect on users (Lindemann, Reference Lindemann2022). Öhman and Floridi, for example, have argued that the digital remains of the dead deserve the respect accorded their bodies, which may mean deletion (Öhman and Floridi, Reference Öhman and Floridi2017). Palermos has also raised questions about whether the use of another person’s data without their consent may be considered a form of assault (Palermos, Reference Palermos2026). Stokes, on the other hand, argues that the deletion of the digital remains of the dead erases them from the consideration of the living. He notes that social media sites contain profiles of the dead that are used as memorials and can be managed without causing harm (Stokes, Reference Stokes2021). These insights on uses of the digital remains of the dead, of course, are part of a much longer history in which archivists and history makers have given much thought to the obligations we have to the dead as well as to the living in looking to the past. Derrida’s and Kleinberg’s works on how we may be haunted by our ethical responsibilities to the dead – particularly in cases of trauma, lost social or political movements, or injustices – spring to mind (Derrida, Reference Derrida1993; Kleinberg, Reference Kleinberg2017). So too, we may consider Ovenden’s insightful recognition of the fragility of records in theatres of war (Ovenden, Reference Ovenden2020).

Respecting the digital dead may be part of addressing traumatic pasts. As Todd Presner and colleagues emphasise in their careful – and often painful – examination of the relationship between Holocaust testimonies and algorithms, technologies can humanise and dehumanise. Their work reflects on the differences between testimonies you might hear directly, testimonies you might read, hear, or see, and testimonies that are used to create avatars. Analysing their experiences of interviewing Fritzie Fritzshall and then engaging with an avatar of her after her death, they were repeatedly struck by what the avatar did not say and how it did not say it. The avatar left out important details in Fritzie’s testimony – even when they prompted the avatar to offer those details – and recounted events without anger, disbelief, humour, and reflection. Some parts of testimonies will be told repeatedly, they remind us, and other parts will never be told. Are we comfortable with that? History makers select, they remind us, but what ethical principles guide those selections? Avatars may be new, but the ethical questions Presner and colleagues prompt are not, and they provide a sharp reminder of the need to consider both what humans and algorithms should when they make histories (Presner, Bonazzi, Deblinger, et al., Reference Presner, Bonazzi and Deblinger2024).

First Nations approaches to respecting the dead – particularly in Australia, where I am from – also provide innovative insights into viewing history more holistically. First-Nations history makers, as I will argue in the next section, see history as contributing to holistic understandings of the past, present, and future of the planet. They are in a good position to advise on why and how data from the past should be retained and how it may be used and misused.

It is also important that we scrutinise the assumptions about those who interact with avatars. There are plenty of examples from histories of technology which can assist us in thinking about audiences. It was long assumed, for example, that an early movie showing a train made viewers flinch and flee. History makers such as Gunning questioned that assumption by asking whether we might have underestimated the experiences of audiences with theatrical and technological attractions (Gunning, Reference Gunning2025). We must ask ourselves, again, whether we have good evidence to conclude that all those who engage with digital ghosts need our protection. Lee-Talbot (Reference Lee-Talbot2025) has provided an excellent example where users who engaged with Charlie – a fictional avatar of the First World War soldier made from archival and media records – raised critical questions about whether it was respectful to the dead and advanced understandings of diverse contributions to the war. Digital ghosts may, for instance, help some people to grieve, to avoid loneliness, and to speak up about those who may have caused them harm (Cambell, Liu, and Nyholm, Reference Campbell, Liu and Nyholm2025; Elder, Reference Elder2026). More broadly, digital ghosts may have a positive impact on learner engagement and success (Zhang and Wu, Reference Zhang and Wu2024). These examples remind us that those who engage with avatars are agents and that they may exercise authority over the dead. More broadly, the examples of histories of technologies remind us that the users of AI bring a host of experiences to their engagements with platforms and that they may use them in ways that are not anticipated.

2.3 Seeing the Past with Non-Human History Makers

Indeed, one of the important points made by You, Only Virtual (Reference You2020) is that their ‘versonas’ reflect the perspective of the person who makes them. This point is also true of records and data from the past more broadly. Archival records are made for a purpose, and sometimes they record harmful acts. The virtuosity of history makers is often seen in their ability to look differently at archives of this kind to illuminate other stories. Microhistorians, to take just one example, often work with judicial records to find the voices of individuals and groups who were imprisoned, tried, and even lost their lives (Magnússon and Szijártó, Reference Magnússon and Szijártó2013). Larivière, for instance, has used judicial records to illuminate the beliefs, innovations, and experiences of ordinary people in Murano under Venetian rule in the sixteenth century (Larivière, Reference Larivière2021). AI may foster more of this innovation, particularly in the consideration of the non-human. Consider Scalzi’s fictional invitation for us to consider a world in which a digital ghost of a person is made from the perspective of a pet that outlives its owner (Scalzi, 2019, in Elder, Reference Elder2026, 144–145). In this case, a digital ghost is made from the perspective of their non-human interactions. If we flip the scenario around, we can also imagine a pet owner making an avatar of, say, a beloved dog or cat. Animated digital ghosts of pets are not currently available, and I confess that this surprises me given the ways humans have memorialised animals over time. Things get ethically interesting when we give thought to pets being comforted by digital ghosts of their owners. Artificial historians may make histories for non-human audiences. Moreover, they might tip over to the confronting when we consider examples of museums creating avatars of extinct and service animals. Can digital ghosts of passenger pigeons or Tasmanian tigers, or animals employed in theatres of war, haunt us in the ways that Derrida and Kleinberg highlighted for dead persons? The answer appears to be yes when we consider the now burgeoning body of historical research on how human activities have shaped planetary change. Chakrabarty guides us in seeing human and non-human phenomena as reminders of our entanglement in different spatiotemporal scales and relations, and that they can incline us to act for the good in different ways. They can be ethically empowering, rather than harmful or disempowering (Chakrabarty, Reference Chakrabarty2021).

Avatars, it might be argued, are a new form of attraction in which their audiences are in on the act. Focusing on avatars, though, may lead us to miss other more fundamental ways in which non-human actors shape AI. Recall the examples of rapid search, summary, and translation tools. As I noted earlier, many of these AI-supported methods of learning from the past exist thanks to human admiration for the social activities of insects, as well as geological or biological processes. Ant colony and bee colony optimisation algorithms, for example, offer highly effective and efficient approaches to the location and retrieval of data from the past. Imagine watching a group of ants finding the most efficient way to find and retrieve food to their nest, and you gain a good picture of how algorithms may search vast volumes of information to respond to your prompts (Hughes-Warrington and Martin, Reference Hughes-Warrington and Martin2022). If you have engaged in any searches for historical materials lately, you may have insects to thank for the results. We may also have insects to thank for our opportunities to shop, to watch movies, and to gain access to healthcare, education, finance, or justice. The movement of insects thus may have an ethical impact on human opportunities via AI. The makers of the moral machine assumed that humans make ethical decisions, but we have already reached an age when insects may also play a role in preventing harm.

2.4 AI-Supported Approaches to History Education

History is often connected with AI via the topic of academic integrity in schools and universities. Some teachers worry that students may generate class notes and assessment tasks without the effort needed to analyse materials; check citations and claims for accuracy, bias, and context; and form an argument, let alone a novel argument. Students may even instruct AI to respond to tasks in alignment with a rubric. AI is seen as helping too much and providing outcomes that would not stand up to scrutiny by a history maker. Other teachers can see the benefits of helping students to advance their understandings of AI to ensure that they can contribute to the social and economic changes ushered in by these technologies. AI in this case becomes part of activities in which students both create and scrutinise outputs. Along the way, they learn more about the theories and methods history makers use to generate novel insights that rest upon a solid foundation of evidence (American Historical Association, 2025; Arrow, Reference Arrow2025).

Some of these differences of view may be explained by people’s experiences with AI. AI may be new to some teachers of history, and it may be the norm for others. Many are also trying to work with contradictory or unclear institutional, publishing, and grant submission policies on AI (Arrow, Reference Arrow2025; Evans, Reference 47Evans2025). There is little question, though, that everyone shares a love of history and that they want to see it flourish in their local community and around the globe. There are ways of supporting an enhanced approach to academic integrity with AI without relegating its use to spelling and grammar checking or a return to in-person exams or oral assessment tasks. Furze, Perkins, Roe, and colleagues, to take just one example, have developed an AI assessment scale that teachers and students can use to set the parameters of AI usage, from ‘none at all’ to ‘extensive’ (Furze, Perkins, Roe, et al., Reference Furze, Perkins and Roe2024). The scale can be used multiple ways. Teachers can set expectations with the scale, students can self-declare AI usage with it, and both can work together to see what outcomes might come from using AI to differing degrees. While a helpful framework for students and teachers, this Element will have shown you that there is more to AI than generative tools. Students may, for example, work with transcription tools to advance their consideration of primary, handwritten materials. In a similar vein, there is more to education than assessment tasks, and there is more to AI than being a user of it. Teachers and students can be the makers of artificial historians, and they can use a broad range of AI tools to advance their understanding. If scales can be made to codify AI usage, they can also be used to codify AI ingenuity. History makers can also be part of AI research and developments designed to identify the appropriation and manipulation of materials.

To gain a sense of the wider opportunities that come from connecting history education and AI, we can look at the example of adaptive education. Adaptive education draws upon the pattern identification and recommendation capabilities of AI tools to adjust the format, language complexity, types of feedback, and pace of essential learning materials to support student success. Imagine, for example, a digital history text which unfurls materials as students demonstrate mastery of tasks and which accommodates some of their interests (Strielkowski, Grebennikova, Lisovskiy, et al., Reference Strielkowski, Grebennikova and Lisovskiy2025). The text, in turn, is connected to a dynamic learner profile which helps a student’s teacher to see strengths and opportunities for development. Ethical questions may be raised about the retention and protection of student and teacher data, whether these technologies are globally affordable and whether forms of data and algorithmic bias undercut the opportunity for people around the world to see themselves as a part of history making. We can acknowledge these concerns but also note that these approaches may help more students to experience success with history at different educational levels.

Learn Your Way Google (2025) currently offers two sample adaptive texts for history – one on Roman history and the other on human migration – to give us a sense of this approach. Each of these texts can be adapted for early adolescent and university student readers. The texts have the appearance of print textbooks with quizzes that have been digitised, and that means a lost opportunity to test out innovations in dynamic spatiotemporal scaling and in the combination of textual and visual formats to advance learning. Providing students with the opportunity to adapt language complexity and to select tests as you go or final test formats, though, provides something of a glimpse of what might come. To be clear, adaptive texts are not intended to provide every student with different content and methods development; they simply adapt the order, pace, and format of essential materials. They are used to ensure that all students master the materials and tasks before they progress and that they are provided with timely and practical feedback. Adaptive texts can be used to support students who might have previously struggled to engage with a text-rich subject, as well as students who wish to connect their studies of history with other curriculum subjects. Adaptive texts may, to take just one example, help dyslexic students to experience greater success with textual comprehension. Adaptive texts are not a recipe for teacher replacement, and they may indeed help teachers to support student progress and success in large classes or as students change teachers, terms, and study years (García-Martínez, Fernández-Batanero, Fernández-Cerero, et al., Reference Chapman, Gómez-Carrasco and Monteagudo-Fernández2023).

2.5 Seeing the Value of History across the AI Ecosystem

Enhanced student success in history can contribute to the strengthening of expertise in AI. This is because, as I have argued throughout this Element, understanding how AI works with data from the past can be the key to developing more effective approaches. The need for expanded student success in history is timely because people can be fearful of AI. They might not have had the opportunity to learn how AI technologies work in ways that make sense to them. They may worry that if they are not a part of developments, then AI might take their job, their ideas, and the planetary future of generations to come. They also worry about bias, misinformation and disinformation. ‘How can I help you?’ – the question which opened this Element – can be understood as an invitation to contribute to AI not only as a user but also as a designer. This is not asking history makers to be coders or computer scientists or for computer science to absorb the discipline of history. Rather, it recognises the value that more inclusive reasoning about the past may have for a broad range of AI tools.

A current barrier to this more inclusive approach to the development of AI is that people might not appreciate how helpful history education and research can be. They might not have associated history with their aspirations and job preferences. They might not be able to connect what they have learned about history with a broad range of activities, where the past is explored and recommendations are made about the present and the future. ‘How can I help you?’ is not therefore just an invitation to contribute to AI as a designer; it is also an invitation to appreciate the value of history making more broadly. One of the important insights Roy Rosenzweig and David Thelen, and Pete Burkholder and Dana Schaffer offer is that there are opportunities to enhance trust in teachers of history at all levels of education (Rosenzweig and Thelen, Reference Rosenzweig and Thelen1998; Burkholder and Schaffer, Reference 45Burkholder and Schaffer2021).

A small tilt in teaching and communication approaches can make the value of history to AI more apparent to students and the broader community. These approaches can highlight the critical role of history making in supporting trust in information as well as in the development of ingenious approaches to reasoning about the past. Consider the goal of teaching students and the broader community about the interpretation of historical phenomena. History makers interpret and explain phenomena through a range of decisions focused on scale, methods, sources, audience, and so on. Interpretations can be both made and dissected in history education to help students to see that histories are not unmediated mirrors of the past. These activities can help to illuminate the assumptions that students and their communities might hold about phenomena from the past – such as museum objects, eyewitness accounts, or data from the past – as well as the teachers and makers of histories. Chapman offers an incisive approach to advancing understandings of historical interpretation. He starts with everyday prompts such as different book titles and films on the same topic. He also explains how we can help students to reason about different perspectives on the same phenomenon and to figure out how to build a narrative sequence that is strongly grounded in reasoned interpretation. His examples reinforce how interpretation runs through history making and contributes to research rigour and factual authority (Chapman, Reference Chapman, Gómez-Carrasco and Monteagudo-Fernández2023; see also Chapman, Reference Chapman2024).

We can extend Chapman’s insights to include generative AI. Burzlaff’s study of generative AI commands our attention with an uncomfortable exploration of what happens when Holocaust testimonies become machine made. This work reinforces and extends many of the insights of Presner and colleagues (Reference Presner, Bonazzi and Deblinger2024). Comparing testimonies and artificial accounts, Burzlaff notices the moment in which an artificial historian fails to acknowledge survivor Samuel W’s recorded testimony of being praised by a German guard for accepting a beating in silence. With that divergence, he argues, we confront our assumptions about silence and the Holocaust. We see generative AI laid bare as a describer rather than an interpreter of events (Burzlaff, Reference Burzlaff2025; see also Gossard, Reference Gossard2024; Leme Lopez, Reference Leme Lopes2023). Samuel W’s testimony prompts our need to reckon with assumptions about how groups respond to persecution and genocide. Burzlaff’s research can be usefully transposed to the consideration of avatars of those who either survived or who were murdered in the Holocaust. Do they too describe what happened, rather than interpret it? Looking at the use of LLMs in another historical context, McLean and colleagues show how prejudicial stereotypes in training data can affect AI-supported work to generate narratives of Australian convicts from tabular data (McLean, Roberts, and Gibbs, Reference McLean, Roberts and Gibbs2025). In another example, Worrell highlights how the good intention of educators to include First Nations perspectives in their teaching may encourage them to create content with generative AI that reinforces patterns of theft and misappropriation. The resulting clash of good intentions and poor outcomes provides an important opportunity to think about how artificial historians and histories might be recast in culturally and ethically appropriate ways (Worrell, Reference Worrell2024). These examples show us that history making, as well as AI, is brought sharply into focus as a matter of decision, responsibility, and accountability.

We need not limit our focus to generative AI. GitHub (Microsoft, 2018–), Kaggle (Google, 2017–), to take just two examples, provide entry points for thinking about the value of history making to AI. GitHub is a site where data and software projects are shared and stored, and Kaggle is a platform for hosting both software and data projects and data-focused competitions. A search for history projects on both platforms returns a variety of results, from work to visualise archaeological work or to analyse digitised books or archival records to the design of algorithms to search large volumes of news reporting to identify potentially significant events that will make history and impact financial markets. Not all these projects are clearly explained, but a variety of topics and approaches to making sense of data from the past can be presented to students for analysis, discussion, and the identification of priorities and gaps. The same activity can be undertaken by looking at the trending topics lists of both sites: for example, current trend lists include stock and retail prices over time, football player career duration, retail sales datasets, open-source textbooks and technical solutions for storing all the data on your computer on a single timeline. Multiple applications for each of these projects may be considered, as well as what might happen if helpful or malicious approaches to reasoning about the past are used for each. The more you look at these platforms, the more you realise that the kinds of reasoning that Chapman encourages us to develop in students have broad application. Exploring current and possible approaches in AI can help to build appreciation of the value of historical reasoning across what we should think of as an AI ecosystem. I am going to expand on this point, which is the key one of this Element in the next section.

3.1 Silences, Unfairness, and Popularity

The various harms of AI often stem from an underappreciation of its historical nature. Critical commentaries on AI have repeatedly highlighted the problem of bias in data and algorithms. Buolamwini, for example, has explained the prejudicial impacts of AI through her poor experiences with facial recognition tools as a black woman. Facial recognition tools do not work well with many people around the globe because they were not developed with them in mind. Social, political, and economic inequalities are codified in platforms because of unfairness in data collection, retention, and use (Buolamwini, Reference Buolamwini2023; see also Criado Perez, Reference Criado Perez2019). Gaps are not the only issue with AI. AI tools may work with data collected in contexts that are adversarial, prejudicial, distressing, or even life-threatening for individuals and groups. That data may be used to make unfair recommendations and decisions about access to services and goods. Your financial or residential history may, for example, determine whether you are able to secure a loan or a rental property. D’Ignazio and Klein provide a useful outline of how ‘redlining’ or the risk rating of loans based on neighbourhood demographics rather than individual credit history has supported unfair financial outcomes for over a century (D’Ignazio and Klein, Reference D’Ignazio and Klein2020, ch. 2). Concerns have also been rightly raised about the use of data from the past to determine health insurance coverage or premiums or to target groups in crime prevention initiatives (Broussard, Reference Broussard2023). These are not new problems, as the redlining example shows us. History makers have long acknowledged the gaps, silences, and inequities in evidence from the past. Trouillot reminds us how decisions in the making of sources, their collection in archives, the retrieval of those sources from archives, and the composition of histories can create and sustain silences and inequities (Trouillot, Reference Trouillot1995).

History makers have also devised ingenious ways to work with the silences and inequalities of the past. Indeed, there are so many examples of histories which acknowledge and counter silences and inequalities that it would be difficult to do justice to them in this short work. What we can do is note the diversity of recent examples. For instance, Chatelain’s connection of black ownership of McDonald’s franchises with US civil rights history, while Barraclough’s turn to material evidence to expand our understanding of everyday life in the Viking world. Similarly, Fullager’s employment of reverse chronology to peel back stereotypical depictions of Australian Aboriginal man Bennelong, and both Callaci’s and Witt’s use of oral histories to explain twentieth-century wages for housework campaigns, and the rise of chip company Nvidia, respectively (Chatelain, Reference Chatelain2020; Barraclough, Reference Barraclough2025; Fullager, Reference Fullager2023; Callaci, Reference Callaci2025; Witt, Reference Witt2025). Their deft work with evidence, methods, and theories highlights the possibilities that result from thinking about the past differently.

Drawing upon the expertise of history makers in seeing the past differently may be particularly fruitful for efforts to manage and to mitigate popularity bias in recommendation systems. For just as AI may codify social and political inequalities, it may also reinforce overly narrow outcomes that are not beneficial for AI businesses and organisations either. Let me explain with a simple example. Imagine you want to make an AI system to recommend moving-image histories to viewers around the globe. If you based your recommendations on the number of times that viewers have watched a film, your system may be skewed towards the oldest historical films. That’s wonderful if a viewer has not seen Dreyer’s La Passion de Jeanne D’Arc (1928), to take just one example, but they may prefer a more recent film or one that suits their interests rather than the viewing habits of the globe. In this example, the viewer does not get a recommendation they want, and thus they might not use the system again. The creators and curators of histories that are not recommended may also be frustrated with systems that do not mitigate popularity bias. It is also worth noting that in some instances, bots may be deployed to generate fake views or reviews and even to steer users to malicious materials (see, for example, Chiang, Chen, Song, et al., Reference Chiang, Chen and Song2023). I have focused on moving-image histories in my example, but we can also imagine users being frustrated with recommendations for history books, music, games, curriculum, and historical sites.

A variety of approaches have been advanced to mitigate popularity bias in recommendation systems. These focus on data handling, reasoning methods (algorithm design and performance), and the communication and management of outputs (Klimashevskaia, Jannach, and Elahi, Reference Klimashevskaia, Jannach and Elahi2024). In the case of our recommendation system designed for moving-image histories, for instance, filters, adjusted weightings and the provision of more than one suggestion might be contemplated. History makers can also advance solutions designed to improve the novelty and diversity of recommendation system outputs. They too can draw upon their expertise in data handling, reasoning about the past and communication to ensure that multiple perspectives on the past are advanced and maintained. Importantly, history makers reason about data or evidence from the past intentionally. This means that they can analyse and explain the same phenomenon from the past in multiple ways that have credence (Hughes-Warrington with O’Brien and Martin, Reference Hughes-Warrington2025). Their ability to advance novel and diverse analyses of limited and even prejudicial datasets may be particularly helpful for the development of better approaches to the early training and development of AI models.

3.2 Probability, Uncertainty, and Inadequate Service Standards for History

History makers signal their responsibility for analyses in a variety of ways. Their use of notes and reference lists, for instance, makes it possible for readers to check information for themselves. They may place their data in open repositories so that their claims can be verified. They may also use a range of approaches to reasoning and communication to signal their confidence in claims (Hughes-Warrington with O’Brien and Martin, Reference Hughes-Warrington2025). When you read a history, you may notice the words ‘if’, ‘possible’, or ‘might be’, or questions. Moving-image histories also commonly include opening titles such as ‘based on a true story’ or ‘inspired by true events’. These kinds of claims remind audiences that knowledge of the past is not certain. There may be gaps in evidence, and the discovery of new evidence about the past is always possible too. Historians also present their claims in the form of stories, and those stories have different narrative structures. It is also common for historians to cast new light on evidence and thus to generate new interpretations of phenomena. This does not mean that the claims of history makers lack authority or credence; they are simply careful to note the limitations of their claims and that new interpretations are possible. History makers also contextualise their claims. They do this by locating phenomena from the past in larger or smaller scales of analysis. They commonly cite the work of other history makers, and they consider whether phenomena share commonalities with other phenomena from the same or other times.

AI technologies also generate credible solutions to problems without certainty. Online shopping or streaming platforms, for example, may offer you recommendations with a percentage rating. A film may be recommended to you by a streaming service, for example, as a 95% match for your interests. Some navigational systems or generative AI platforms will offer you more than one option in response to each of your queries. More commonly, though solutions and recommendations are communicated with assurance and without references needed to check them. Users might not know therefore that AI technologies are probability and statistical pattern matching systems. These systems turn on algorithms that offer pragmatic, rules-based instructions which produce near-certain or near-precise outcomes. They generate outputs which, like histories, can have credence without certainty. They are often designed to generate solutions to problems, recommendations, or responses to users in short time frames and in computationally efficient ways. We should not assume that the companies which are behind AI technologies promise to deliver results with the same level of care or in the same formats as human history makers.

Users may not view the ways AI technologies present solutions and recommendations as a problem. Indeed, they may have confidence in them. Steyvers and colleagues, for example, have argued that users prefer it when generative AI agents respond to prompts in a confident tone. Use of conditionals which historians tend to use such as ‘if’ and ‘might’, as well as expressions of uncertainty are less favoured (Steyvers, Tejeda, Kumar, et al., Reference Steyvers, Tejeda and Kumar2025). Generative AI tools have also been associated with more positive, ethical, and empathetic responses than human ethicists, and high user satisfaction with responses to history prompts has been noted (Dillion, Mondal, Tandon, et al., Reference Dillion, Mondal and Tandon2025; Tomlinson, Jaffe, Wang, et al., Reference Tomlinson, Jaffe and Wang2025). This does not mean that historians are poor communicators or susceptible to job loss due to automation. Qiu and colleagues have argued that AI technologies do not reason well with materials and data from the past and that more expert systems and evaluation frameworks are needed (Qiu, Xiao, Wang, et al., Reference Qiu, Xiao and Wang2025). The opportunity exists for historians to continue to develop new ways to signal the gaps, silences, and prejudices of the past, as well as the possibilities of reading sources in new ways. We need to remember that histories have not always been made with reference lists, notes, or conditionals. Historians may, for example, develop more visual ways of indicating uncertainty and certainty, and this may be helpful for the users of all kinds of AI technologies that process data from the past. As the use of AI agents accelerates – you may interact with multiple agents every day – the urgency and importance of being able to question and to understand the limitations of artificially generated answers grows.

3.3 Hate Histories and Safety

Improving the ways AI technologies process materials and data from the past is not just desirable; it is also important for the safety of users. Historical prompts can be used to generate illegal, harmful, or hateful information in generative AI technologies. Artificial histories that give instructions on how to, for example, make weapons or promote antisemitic or Islamophobic ideas undermine both social cohesion and confidence in technologies. Moreover, the processes these technologies use to solve problems or to make recommendations might not be robust, particularly over time (Russinovich, Salem, Eldan, et al., Reference Russinovich, Salem and Eldan2024; Yu, Liu, and Liang, Reference Yu, Liu and Liang2024; Chen, Gong, Liu, et al., Reference Chen, Gong and Liu2024). Responsible history makers recognise that there is not always alignment in ethics between the past and the present, and they demonstrate expertise in reasoning with information that is incomplete, biased, or even prejudicial. Nor do they make histories to vilify or to threaten the safety of individuals or groups in the present. They do not make what I call hate histories (Hughes-Warrington, Reference Hughes-Warrington2013). Rather, they work to explain or to contextualise phenomena in ways that make sense for their readers and to make careful, critical judgements about acts of cruelty, harm, and hatred (Little, Reference Little2022).

History makers are not currently shaping AI research and development on safe and robust reasoning with data from the past. This is a problem, because current approaches to risk mitigation fail in at least three ways. First, universal filters and blocks for hateful, harmful, and illegal materials have led to legitimate historical material being prohibited or to the generation of anachronistic results. In 2019, for example, YouTube filters designed to block hate speech led to the removal of history teachers’ resources on the Second World War. More recently, Google Gemini’s image generation function was suspended after the generation of images of Asian women in Nazi uniforms and black Popes (Waterson, Reference Waterson2019; Milmo, Reference Milmo2024). Second, safety and reasoning test approaches predate generative multimodal AI models and AI agents and fail to examine how they process limited datasets, reason through tasks from start to finish, and perform over time. Historical reasoning is, rather, conflated with short responses to closed questions or the translation of sources, and politically or ethically sensitive questions are not included (e.g. Qiu, et al., Reference Qiu, Xiao and Wang2025; Tomlinson, et al., Reference Tomlinson, Jaffe and Wang2025). This means that complex, sensitive, or persistent historical enquiries may challenge both the safety and performance of models over time. Third, safety treatments for histories are assumed to be the same as for fictional prompts (e.g. Lin, Shen, Yang, et al., Reference Lin, Shen and Zihao2025; Sharma, Tong, Mu, et al., Reference Sharma, Tong and Mu2025). This again leads either to problems with histories being ignored or to ill-suited blocks and filters. People in the past engaged in activities that are now – and even were then – considered illegal, hateful, and harmful, but we do not want to enable their replication today. Safety issues with AI-generated histories can be improved and even mitigated via the expert input of historians and AI-reasoning simulation models that test how models work over different periods of time. There remains a considerable opportunity to benchmark and to support the improvement of how a broad range of AI technologies process data from the past safely and well.

The expertise of historians is invaluable for improving and designing robust and safe AI technologies across a broad range of applications. LMMs make sense as a focus given their use in home, community, and business settings around the world. It is important to remember, though, that improvements in reasoning with data from the past can also be of benefit to recommendation systems and in less obvious applications such as cybersecurity. Cybersecurity systems that look to predict and mitigate identity and financial theft, as well as industrial, electoral, political, and social attacks, benefit from the identification of patterns and anomalies in data that has been spatiotemporally stamped (we know its time and place of origin). They also benefit from techniques that help analysts to find relationships and anomalies in large volumes of spatiotemporally dynamic data. Historical reasoning can help to prevent attacks on social cohesion, as well as the protection of your passwords, personal details, and financial resources (e.g. Mahboubi, Luong, Aboutorab, et al., Reference 50Mahboubi, Luong and Aboutorab2024).

3.4 Missing First Nations Innovations in AI and the Environmental Impact of AI

First Nations history makers have a particularly important role to play in improving the safety and performance of AI technologies. This is the case for multiple reasons, and I am going to highlight two. First, their holistic approaches to history making – which treat the past, present, and future, and space, place, and time as interconnected – remind us of the importance of the past for human and environmental health and well-being in the present and in the future (e.g. McGrath and Huggins, Reference McGrath and Huggins2025; McGrath and Russell, Reference McGrath and Russell2022). The ethics of First Nations histories encompass human relations with the living environment and the other way around. Seeing history making in these relational and holistic ways reminds us to see AI in the round. Contemporary AI platforms process large volumes of data quickly, and this means that their hardware, locational, and human underpinnings can be costly. Cables, chips, towers, and data centres require critical minerals, land, power, and water to meet user demand. AI can be a resource-heavy activity and a contributor to global carbon emissions. It is worth remembering that every time we ask AI to help us to make sense of data from the past, we consume natural resources or contribute to the global expansion of data centres. As Hajiesmaili and colleagues remind us too, the impact of our queries can be uneven in their global impacts. The number and energy efficiency of data centres, for example, varies worldwide (Hajiesmaili, Ren, Sitaraman, et al., Reference Hajiesmaili, Ren, Sitaraman and Wierman2025). First Nations Elders invite us to consider how AI can contribute to – rather than extract from – the health of environments (Hughes-Warrington, Reference Hughes-Warrington2026).

Second, First Nations approaches to history making remind us to consider the contribution of AI to the health of peoples all over the world. Adams is one of several researchers who have provided insight into the hidden cost of what is sometimes called ‘artificial artificial intelligence’; poorly paid workers who sit behind the edifices of AI technologies and screen illegal, hateful, and harmful materials. This can come at a cost to their well-being, as well as to that of their families (Adams, Reference Adams2025; see also Hao, Reference Hao2025). Yet First Nations historians do not simply observe this global problem; they have drawn upon their deep histories to address it. Researchers who argue for Indigenous data sovereignty, for example, highlight questions about the contexts for, and cost of, data from the past. They insist that we need to ask where data comes from, who makes it, who should have access to it, and how it should be used before it is used, not afterwards. They aim to avoid the repeat of past activities in which data was used to stereotype, humiliate, and even murder First Nations Peoples (Walter, Kukutai, Russo Carroll, et al., Reference Walter, Kukutai, Russo Carroll and Rodriguez-Lonebear2021). Their holistic approaches show us pathways for connecting all too-often separate endeavours to advance respect for persons through privacy and data protection politics, with environmental protection measures (e.g. Ebers and Sein, Reference Ebers and Sein2025; Satpathy, Mahapatra, Agarwal, et al., Reference Satpathy, Mahapatra, Agarwal and Mohanty2025).

3.5 Appropriating Histories and Intellectual Property

Questions about data sovereignty might not be questions about data ownership. This is because there is not a global view on whether AI technologies – or even people – can own data. Patents and copyright have their origins in the fourteenth- and eighteenth-century Italy and Britain, respectively, and legal policy and practical approaches to the assertion and management of ownership of ideas have continued to change over time (e.g. Drahos, Reference Drahos2010; Bellos and Montagu, Reference Bellos and Montagu2024). Further changes are needed if current approaches to training AI technologies with significant volumes of data from the past persist. When you accept the terms to use AI technologies, for example, you tend to grant the organisations that run them free licences to use, transfer, modify, copy, display, or transform the data you generate or give them. This means that you allow them – or others they work with – the right to make histories about you, even if you are not aware of them doing so. Those histories may contribute to the shaping of recommendations and opportunities to access entertainment, education, financial services, and even healthcare. Given the far-reaching ways in which artificial histories can be used to shape opportunities, I have argued that people have the right to know the histories that are made about them and to contribute to the making of better and fairer histories (Hughes-Warrington, Reference Hughes-Warrington2025b). It is also the case, regrettably, that your information may be used without your permission. Pirate platforms, for example, may make copies of your data and works without your permission. Historians, along with other creators, have expressed concerns and taken legal action in cases where AI organisations have used their IP for training purposes without permission. This includes cases where AI technologies may have scraped materials from online libraries or repositories of pirated materials or argued that ‘fair use’ provisions cover AI training (Rademeyer and Selvadurai, Reference Rademeyer and Selvadurai2025). Current approaches to copyright and IP may also not work well in First Nations contexts in which phenomena such as words, ideas, and technologies are not seen as the possession of individuals or groups at specific times. Rather, they are respected as interconnected across place and time. On this view, the scraping of content may be seen as the continuation of colonial practices in which lands, languages, resources, objects, and people are taken without asking (Janke, Reference Janke2021; see also Mejias and Couldry, Reference Couldry and Mejias2024). Individuals and groups may also find it difficult to navigate situations where copyright and Indigenous Cultural and Intellectual Property (ICIP) principles appear to clash (e.g. Napoleon, Johnson, Overstall, et al., Reference Napoleon, Johnson, Overstall and McKenzie2024). Consider the case of an AI organisation that creates a digital ghost or retains digital remains in line with the European Union’s General Data Protection Regulation (European Union, 2016). This action, although legal in one sense, may cause distress in communities that restrict the use of the names and likenesses of people who have died. A family may be successful in having an individual’s account closed, but any data provided to third parties or used to train AI models might not be able to be retrieved. Older, archival approaches to respect for cultural protocols may not be advisable or practical if communities are uncomfortable with the colonial legacies of museums, galleries, and libraries, or if an individual’s data is unable to be separated out from training data in the way that you might close access to a section of an archive. It is akin to a situation in which you have archival materials about a person distributed across a collection and not catalogued under their name. Issues like this require further examination and advice to ensure the wishes of communities are respected.

An increased focus on inference approaches may help to reduce the human and environmental impact of AI technologies. These approaches focus on strengthening the ways in which AI technologies process and reason with data. Inference AI technologies can be fine-tuned over time, and their performance improved using smaller, technical datasets and expert insights. These approaches thus recognise what historians have known for a long time: that you can reason with authority and credence with materials that are incomplete, complex, and problematic by today’s ethical standards. To improve reasoning about data from the past, consideration of spatiotemporal, causal, conditional (if… then), and counterfactual (what if… then) relations between data points becomes important, as does consideration of how human experts solve problems (e.g. Mateos-Aparicio-Ruiz, Montealegre-Macias, Deniz, Reference Mateos-Aparicio-Ruiz, Montealegre-Macias and Deniz2025; Anlló and Palminteri, Reference Worrell2024). There is a strong body of research on the nature of historical reasoning to draw upon in developing AI reasoning (Tucker, Reference Tucker2025; Hughes-Warrington with O’Brien and Martin, Reference Hughes-Warrington2025). This lessens both the need for data and potentially the compute needed to process it. It also suggests that historians as experts are part of AI research and development rather than the users of it or on the receiving end of IP theft.

4.1 Ethics and Recognition for Artificial and Human Historians

Drawing upon the expertise of historians may help to improve the reasoning processes of AI technologies and thereby lessen their social and environmental impacts. What is technically possible, though, also needs to be ethically acceptable. When historians reason about the past, they think about how they should explain phenomena. They make critical, yet careful decisions about the past and think about the purposes and uses of histories for the present and the future. Many peak historical professional bodies have published codes of ethics that provide a strong foundation for considering how historians might make decisions about the past. The American Historical Association’s Statement on Standards of Professional Conduct (2023), for a start, highlights the values of critical dialogue, trust, respect, integrity, generosity, and transparency for the making of responsible histories. The Australian Historical Association’s Code of Ethics (2021) reflects these same principles but also highlights the importance of care for the past and for colleagues and the wider community (see also Royal Historical Society, n.d.). AI-supported – rather than led – approaches are emphasised in the Royal Australian Historical Association’s instructions on AI (Royal Australian Historical Society, 2024).

Large-scale reviews of guidelines and codes for AI ethics confirm many of these same values and note the serious consequences for individuals and communities of breaches in the protection of people’s data from the past, algorithmic bias, and IP theft (e.g. Corrêa, Galvão, Santos, et al., Reference Corrêa, Galvão and Santos2023). Floridi affirms the value of existing approaches to ethics but notes that explicability is also critical. By this, he means that technical approaches to processing data from the past ought to be explained in ways that make sense to communities. I argue that we can extend Floridi’s idea by connecting it to his broader arguments on the role of accountability and responsibility in ethics (Floridi, Reference Floridi2023; Hughes-Warrington, Reference Hughes-Warrington2025a).

Let me explain with an example. Imagine that you have been successful in developing artificial historians that use robust, helpful, and ethical reasoning processes and that you have been asked to consider whether the works they create ought to be considered for publication in history journals or as books. You may in the first instance look to legal decisions to respond to this question. Courts in various constituencies have ruled that AI cannot be named as an inventor or creator either for patent applications or for copyright agreements. An inventor or creator is generally taken to be a ‘natural person’, which means that, for the moment, AI, animals, and legal entities such as rivers cannot hold copyright. In practical terms, human authorship is traced back via causal thinking about a technology’s development, but debate continues on the recognition of works made without immediate human input (Bellos and Montagu, Reference Bellos and Montagu2024). As it currently stands around the world, copyright cannot be assigned to artificial historians that generate journal papers or books. It may, rather, be assigned to the people that invented the technology that generated those works.

Looked at through this legal lens, it is understandable that ethical questions about transparency, responsibility, and accountability might be raised. Ought a person be granted the benefits of authorship for a history that they did not produce or be held responsible for how it was made and accountable for outcomes for its reception and use? As O’Neill and Floridi both note, the rise of global AI technologies has brought with it a profusion of contributors that make it difficult to hold those who perpetrate harm to account. Thousands of people may generate the code for a technology, and even more may contribute to the production of hardware and the provision of data (O’Neill, Reference O’Neill2020; Floridi, Reference Floridi2013). We have created a world of histories in which authorship, responsibility, and accountability seem unclear. Given this presentation of AI technologies as being able to make histories without intent, and likely also responsibility and accountability, it is not surprising that some journals and book publishers currently prohibit the use of AI in the creation of content.

These are, however, not new problems, as O’Neill argues in her various analyses of obligations to prevent harm across geopolitical borders (e.g. O’Neill, Reference O’Neill2016). A focus on the outcomes of specific technologies – regardless of intent – or on the rights of individuals to privacy or freedom misses states and non-state actors such as corporations, non-governmental organisations, financial peak bodies, and cybergangs. Martin and I have argued for the ethical value of a deep, global history of dynamic scale switching in history making. Historians have long-traversed spatiotemporal scales to highlight what is good, fair, and just (Hughes-Warrington and Martin, Reference Hughes-Warrington and Martin2022).

4.2 The Geopolitics of Artificial Historians

There are a burgeoning number of publications that present our world of AI technologies as either an extension of, or as a new form of, global imperialism. The global scale is presented as helping us to see the accountability and even culpability of state and non-state actors and the possibilities of using technologies for good in the future. Bradford’s Digital Empires (2023) explains our personal experiences of AI as part of a global battleground between autocracy and liberal democracy. This battleground is played out in the mining and processing of critical minerals; the global production, location and movement of hardware such as chips, cables, mobile signal towers, and data centres; design standards and consumer regulations about data; and tariffs. These activities coalesce in a bimodal tension between market- and rights-driven approaches in places such as the US and the European Union, and state-driven approaches in places such as China. Adams, Hao, Couldry, and Mejias, on the other hand, focus their analyses on global processes of data extraction and processing. They highlight the exploitation of workers in some of the poorest countries in the world – or ‘majority world’, as Adams calls it throughout her book – to provide the data and the safety parameters required to make a range of AI technologies seemingly cost free. We may think that we enjoy free access to AI technologies, but they come at the cost of our information and the food, water, work, and environmental security of communities around the world. How is this different, they ask rhetorically, to other empires in global history? (Adams, Reference Adams2025; Hao, Reference Hao2025; Couldry and Mejias, Reference Couldry and Mejias2024).

Thinking globally, Drayton and Motadel remind us, does not mean that we abandon small stories (Reference Drayton and Motadel2016). It is possible to imagine future scenarios in which we are even more aware of the costs and benefits of AI and that we take action to prevent harm against ourselves and others. We may hold that a sharpened focus on the right to privacy, for example, may help to address some of the harms of AI. The intersection of individual and global histories, however, has many futures too. As they write:

We do not get to global views by cutting and pasting familiar units of analysis together. Global collaboration can transform history making, yet it is an aspiration we are yet to realise. Let me explain through the idea of the historical imagination.

4.3 Artificial Intelligence and the Historical Imagination

Saying that the future of AI can be better is not the same as showing it. Codes of ethics and regulations may express aspirations, but if they are not enacted or actionable, little will change for the better. You could write a code of ethics for history and AI, but you need to take the further step to show how its principles might be demonstrated in action. Your efforts might have little effect, for example, if the designers of AI technologies do not involve history makers. I believe that a different approach is needed; one that turns on the historical imagination. Collingwood could not have had AI in mind when he gave a lecture on the historical imagination in 1935 (Collingwood, Reference 46Collingwood, Dray and van der Dussen[1935] 1999). Nor did I when I decided to focus my PhD on his ideas in the 1990s. What he has to say, though, provides the means for us to imagine a new relationship between history and AI, even if he did not intimate it. Making histories is not a matter of cutting and pasting, whether you use actual scissors and paper or neural networks and data. Rather, it entails having a critical eye to the past; this means interpreting incomplete, biased, and even prejudicial evidence, making decisions about what to include and to exclude and contemplating what might have happened or what could happen. More than that, though, history making reflects an obligation to see things differently; to spot the anomalies, to see the small and big stories others have missed, to read evidence in authoritative ways no one else has considered, and to give flight to new theories and methodologies.

Countless people have told me that AI is not an imaginative or creative historian. This is a fair point – and currently legally true – and it is also a captivating one. Focusing on it to the exclusion of other ways of thinking about the historical imagination may mean that we are missing a more collaborative view of AI. Let me explain by taking you back to the survey of histories of AI at the start of this Element. I noted that learning from experience in AI can also be described as learning from the past. You may have also caught that while there are many histories of AI, they tend to tell the same story. In that story, we learned approaches that prioritised learning from the past replaced the symbolic approaches that held AI in a form of winter-like stasis. In the last section, I updated the story to note that a combination of symbolic and learning approaches promises to improve the safety, robustness, and environmental and human impact of AI. It is a great story, but as a historiographer, I expect historians to disagree with one another. I am surprised, and even concerned, when they all agree with one another. Historians disagree in ways that do not undermine our trust in the past but in ways that remind us that we can look at phenomena from the past in more than one way and that those multiple views can be connected to help us understand why things happened and to think about what might be in the present and the future.

Telling the history of AI the same way has meant that its historical nature has been missed. This may have the effect of positioning historians as the end users of AI, rather than as experts who may collaborate with those from technical backgrounds to improve its safety and efficacy. Conversely, it may mean that computer scientists reinvent the wheel or struggle to find solutions to problems that historians have successfully grappled with over decades and longer. Both ways, collaboration is missed, departments of history continue to contract, and we will all wonder how we can find a way to a clearer view on responsibility, accountability, and therefore also trust. AI has the potential to transform the discipline of history: it can help us to develop and apply new methodologies and forms of logic and theory. I also hold, though, that with imagination it is possible to see a future for AI that turns on historical expertise. Historians can help to shape the future of AI. In that world, reflecting the words of Collingwood, the best possible approaches to AI come from asking how we can be the best possible historians. ‘How can I help you?’ is a question we can all ask if we imagine a broader vision of history and AI.