Introduction: a new ‘wild’ era of ancient history pedagogy
Wild Spirit, which art moving everywhere;
Destroyer and preserver; hear, oh hear!
– Ode to the West Wind, Percy Bysshe Shelley (1819)In mid-June 2025, an editorial article appeared in the New York Times Magazine entitled AI is Poised to Rewrite History. Literally (Wasik Reference Wasik2025). In this article, the Magazine’s editorial director, Bill Wasik, confronted a question central not only to research on the ancient world, but also to its teaching: How might artificial intelligence (AI) change the way that history is written and understood? Although far from the first piece of ‘AI history lit.’, it did problematize how the histories scholars create and teach will all soon be ‘AI-assisted’ histories (for other recent interpretations of AI’s influence on history as a discipline, see American Historical Association 2025; Campbell Reference Campbell2025; Graham Reference Graham2025; Humphries Reference Humphries2025; Royal Historical Society 2024). In many ways, Wasik’s article felt like a clarion (or perhaps siren’s) call that change was in the air.
This call was echoed a few weeks later by the papers presented at the first (Generative) Artificial Intelligence and Teaching the Ancient World conference, which took place at the University of Reading (June 25–26, 2025). The conference’s goal was to discuss ‘arising issues’ in ‘teaching, research, and development … so that the many disciplines engaged in ancient world studies might have a consensus about the use, ethics, and impact of GenAI’, and the papers (including the original version of this article) examined the use of AI tools for teaching ancient texts, archaeology, digital public engagement, as well as AI’s environmental impact (Ross Reference Ross2026). Furthermore, the Reading conference highlighted the unique ways that AI can be used to teach ancient history, such as the use of chatbots to recreate Greco-Roman voices (e.g., remixing Homer, Plato, or Apicius) or how AI visualizations can reconstruct or reimagine archaeological landscapes or artefacts (e.g., Roman Britain, Roman coins). In fact, since ancient history more often relies on the use of fragmentary texts or large datasets of material culture remains (e.g., coins or pottery) than modern history, the conference also proved that AI tools might be uniquely useful to reconstructing archaeological narratives for ancient peoples unrepresented in written histories. Ultimately, both Wasik’s article and the Reading conference’s themes emphasized the impending inevitability of AI’s impact on how the past is researched, interpreted, and taught and suggested that whether this change will be positive or negative, both for history and for society, was up for debate.
Flash forward to 2026. AI’s ability to assist researchers has only increased – exponentially – and its use in education has become nearly universal with four out of five students now using generative AI (Stanford HAI 2026). In fact, large language models (LLMs), such as Anthropic’s Claude or OpenAI’s ChatGPT, have begun to make the jump from simply assisting research to doing research, albeit still overseen by human investigators – for now (Campbell Reference Campbell2025; Magnani and Clindaniel Reference Magnani and Clindaniel2025; Metz Reference Metz2026). As of 2026, humans have not been fully transformed into what Cory Doctorow (Reference Doctorow2026) has dubbed a ‘reverse centaur’, that is, a human body appendage being led by an intelligent AI head, but the lightning-speed recent advances in LLM performance have certainly created a ‘centauric’ tension. According to Stanford’s Human-Centered AI Institute’s AI Index study for 2026, a report touted as providing the public with ‘rigorous, objective insights into AI’s technical prowess, economic influence, and societal impact’, (Stanford HAI 2026), AI capability is ‘not plateauing…it is accelerating’.
In terms of accuracy, recent LLM performance on tests focused on science, English, and math skills has improved roughly 30% year over year (Stanford HAI 2026). Moreover, leading LLMs have transcended simply responding to prompts and now exhibit novel levels of ‘cognitive density’ where models plan and reason before they act. Anthropic’s new Claude Mythos Preview model has become so adept at writing and analysing computer code that its public release has been postponed due to its ability to exploit network and website cybersecurity vulnerabilities (Friedman Reference Friedman2026; Krikorian Reference Krikorian2026). In 2026, ‘agentic’ AI models hyper-specialized for research topics and workflows will likely move beyond merely assisting researchers to become true scholarly collaborators. What these year-on-year AI improvements may mean for the discipline of ancient history remains murky, yet they do suggest that the ways we create, disseminate, and utilize our knowledge of the past are about to undergo epistemic change. It feels like a new revolutionary era is upon us.
21st-century scholars are not the first generation to grapple with the disruptive effects that technological innovations can have on society, so perhaps, as historians, we should look to the past for ways to express our anxiety. For example, when the scientific thinking of the Enlightenment in the 18th century led to the development of machines, such as the steam engine and cotton gin, artists and scholars sought to understand and describe how the Industrial Revolution in the 19th century was upending centuries of tradition. One artist who captured the moment’s anxious Zeitgeist was the Romantic poet Percy Bysshe Shelley (1792–1822) who lived in the political shadow of the French Revolution (1789), the anti-technology attacks of the Luddites (1811–16; see also Graham Reference Graham2025), and the massacre of wage workers at Peterloo (1819; Abrams Reference Abrams1987).
In his Ode to the West Wind (1819), Shelley depicted the wind as a symbolic agent of the societal changes he was experiencing, describing it as a ‘wild spirit’ that could both preserve and destroy established practices and beliefs. Given AI’s rapid and disruptive influence on our own society, ‘wild spirit’ once again seems like an apt symbol of technology’s potential to create and destroy. Indeed, AI’s errant nature has been effectively expressed by Ethan Mollick (Reference Mollick2024), the co-director of the Wharton Generative AI Lab at the University of Pennsylvania, who has stated that ‘No one really knows where this is heading’. Thus, in the wake of Wasik’s article, the Reading conference, and Stanford HAI’s 2026 AI Index report, a core question for historians is to consider whether AI tools can be useful gifts to the discipline, or whether they will serve as Trojan horses concealing fresh interpretive dangers.
There are many signs that AI’s ‘wild spirit’ is now confronting ancient historians and archaeologists in their teaching, learning, and research. The spirit’s destructive nature has already been hinted at in articles refuting the claim that AI will soon replace human practitioners (Campbell Reference Campbell2025). However, because AI still makes factual mistakes (hallucinates) and typically provides generic historical information only based on what it can scrape from the internet, the general academic consensus is that it cannot (yet) replace trained historians armed with specialized disciplinary knowledge and methodologies, which include the ability to physically access non-digital evidence (e.g., archives) or real-world landscapes (phenomenological studies) and reflexively experience the meaning of the past, as humans with unique worldviews and identities (Graham Reference Graham2026; Jowsey et al. Reference Jowsey, Braun, Clarke, Lupton and Fine2025; Kansa Reference Kansa2026).
Some tech critics have recently described LLMs’ ability to appear as geniuses at one moment (e.g., pass the bar exam) and cognitively muddled at the next (e.g., cannot tell time on an analogue clock) as ‘jagged intelligence’ (Metz Reference Metz2026). Such observations indicate that full human intelligence may still be required to accurately interpret the past in disciplines like history and archaeology. In these disciplines, hands-on research on non-digital, primary evidence is often necessary and diachronic comparative analyses typically require the real-world knowledge of human practitioners capable of drawing on labour-intensive fieldwork experiences (Carahar Reference Caraher, Walcek Averett, Gordon and Counts2016). And so, despite AI’s ability to ace math and science exams, ‘slow’, human-led research might provide more realistic insights into what was, after all, the human past – to quote Eric Kansa (Reference Kansa2026):
LLMs (and smaller language models) can be useful tools to support research processes, not replace research processes. These examples also illustrate a deeper point. There are many aspects of archaeology and other areas of intellectual labor where quality is deeply dependent on process. In many circumstances, automation and efficiency are counterproductive.
As a result of these critical observations, articles on ancient historians’ engagement with AI tools have begun to flourish in pedagogical research journals and popular higher education literature. For ancient studies, a new open-access journal has even been founded (2025) entitled AI & Antiquity: Journal of Teaching and History in Ancient Studies (https://ai-antiquity.org/index.php/ai/index), while existing publications, such as The Journal of Classics Teaching (https://multimediafluids.uam.elogim.com/core/journals/journal-of-classics-teaching), Advances in Archaeological Practice (https://multimediafluids.uam.elogim.com/core/journals/advances-in-archaeological-practice), and Internet Archaeology (https://intarch.ac.uk/), have begun to publish explorations of the use of AI in teaching and learning. In education studies, an article from early 2022 tallied 4,519 scholarly articles on AI in education between 2000 and 2019 (Chen et al. Reference Chen, Zou, Xie, Cheng and Liu2022), a number that has recently exploded exponentially. For example, a rapid search of Inside Higher Ed’s (https://www.insidehighered.com/), Times Higher Education’s (https://www.timeshighereducation.com/), or the Chronicle of Higher Education’s (https://www.chronicle.com/) online archives for 2025 alone yield multiple articles a week dealing with AI and education. And now, in 2026, it is challenging to find a higher education article on any academic topic that does not touch on AI in some fashion.
Beyond pedagogy, but not wholly separate from it, the integration of AI into ancient world research has been particularly felt in archaeology, a data-rich discipline that has often been an early adopter of computing techniques given its mix of quantitative and qualitative analyses (Figure 1; see the chapters in Garstki Reference Garstki2020; Walcek Averett et al. Reference Walcek Averett, Gordon and Counts2016).
The author practicing ‘born-digital’ archaeology using an Apple iPad ca. 2013 with the Athienou Archaeological Project at the site of Malloura, Republic of Cyprus.
Note. Photo by author.
Figure 1. Long description
A group of people are working at an archaeological excavation site. One person in the foreground is holding a tablet and observing the work. Several individuals are digging and sifting through the soil in different sections of the excavation area. Buckets, tools, and other equipment are scattered around the site. The site is divided into several sections with visible trenches and walls of excavated earth.
Over the last few years, AI has been used to fill in lacunae in Greek and Latin inscriptions (Assael et al. Reference Assael, Sommerschield and Shillingford2022; Hall Reference Hall2025; i.e., when a word or letters in an inscription are missing, AI can provide a probable solution), and LLMs have been cited as key tools for analysing large-scale datasets, albeit with caveats relating to biases and ethical practices (Gattiglia Reference Gattiglia2025; Kansa Reference Kansa2026; Tenzer et al. Reference Tenzer, Pistilli, Brandsen and Shenfield2024; for the ethics associated with AI as an extractive technology, see Crawford Reference Crawford2021; Magnani and Clindaniel Reference Magnani and Clindaniel2025). The use of AI to identify pottery types from shape and decoration even promises to make archaeological analyses faster, cheaper, and more accurate (Anichini and Gattiglia Reference Anichini and Gattiglia2022; Foka and von Bonsdorff Reference Foka and von Bonsdorff2025; Tatsuda et al. Reference Tatsuda, Hori, Morikawa and Inouye2026). With such developments affecting research on the ancient world, it feels like scholars – most of whom are also educators – are at an inflection point (cf. Graham Reference Graham2026; Kansa Reference Kansa2026), a tense moment of epistemic change that is happening much faster than previous 21st-century techno-pedagogical shifts because of the ‘wild spirit’ accelerant of AI.
I have personally felt this tension teaching undergraduate students ancient history courses at Wentworth Institute of Technology in Boston, MA, USA (www.wit.edu). Wentworth is a STEM-oriented (Science, Technology, Engineering, and Math) university focused on preparing engineering, computer science, and design students for careers in industry. As a middle-aged professor trained in Roman archaeology – first with pen and paper, and then with the digital tools of desktop word processors, relational databases, and mobile devices that gave rise to ‘paperless archaeology’ (see Gordon et al. Reference Gordon, Walcek Averett, Counts, Koo, Toumazou, Walcek Averett, Gordon and Counts2016; Figure 1) – I have not only had to keep up with developments in digital humanities (or DH; for a general introduction see, Burdick et al. Reference Burdick, Drucker, Lunenfeld, Presner and Schnapp2012; Gold Reference Gold2012; for the use of DH in classical studies, see Kindt Reference Kindt2024), but also my students’ rapid adoption of AI. Because STEM students are particularly drawn to digital tools, and AI LLMs can facilitate academic cheating by doing students’ reading, writing, and, ultimately, thinking, I have had to be especially vigilant about how AI tool adoption occurs so as to preserve students’ acquisition of history’s socially valuable skills, like critical thinking or cross-cultural appreciation.
At first, the AI-cheating phobia that arose in 2022 with the release of ChatGPT (Blose Reference Blose2023) influenced my own reticence to engage with AI in teaching and researching. How could I use AI in my own scholarship and pedagogy if I were decrying its use to students? Yet, more recently, I have recognized that my ‘Gen Z’ students learn in very different ways than the ‘Millennial’ students of the previous decade. The current Gen Z undergraduate demographic consists of tech-savvy, digital natives born into a world of smartphones, apps, and data (Katz et al. Reference Katz, Ogilvie, Shaw and Woodhead2021). As a consequence, this digital facility has allowed some students to cheat by letting LLMs do much of their work, a practice that a recent MIT study suggested decreases cognitive activity when compared to pen and paper-based analogue learning (Kosmyna et al. Reference Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes2025). However, for other students with the proper pedagogical guidance, AI tools can be just that, tools, and can be used for what Bertram Gallant and Rettinger (Reference Bertram Gallant and Rettinger2025) have recently labelled the ‘opposite of cheating’, that is, the use of AI tools with integrity in ways that collaboratively help students learn (see also Graham Reference Graham2025; Heredia Chimeno Reference Heredia Chimeno2026).
Based on this rationale, I have stopped aggressively policing my students’ use of AI (see also American Historical Association 2025). Instead, I have elevated my teaching in a way that meets the moment in terms of how my students learn and how teaching ancient history can uniquely provide them with the technical and social skills required for the 21st-century workplace (see, for example, the competencies outlined by the National Association of Colleges and Employers [NACE], https://www.naceweb.org/career-readiness/competencies/career-readiness-defined). What is becoming clear is that AI-assisted ancient history is already here; it is (to paraphrase a quote often attributed to William Gibson) just not very evenly distributed. Yet, this is not an epistemic change to be lamented. Because history can teach students an increasingly desired learning outcome – AI literacy – it can also, perhaps paradoxically, bolster its pedagogical relevance even when many college administrators view the liberal arts as passé (American Historical Association 2025; Bowen and Watson Reference Bowen and Watson2024; Diermeier Reference Diermeier2024; Mollick Reference Mollick2024).
In an effort to contribute to current discourse on how the ‘wild spirit’ of AI is influencing ancient history research and teaching, this article’s goal is to discuss how I have critically engaged with AI-assisted history pedagogy via archaeological topics within a Wentworth undergraduate course entitled, Computer Science and Society (CSAS) Studio. First, I revisit how AI is altering historical methodologies and examine how these changes might result in novel interpretations of the past. Second, I consider the core AI literacy skills students will require for research and future employment, and how AI ‘cheating machines’ can become learning ones if students are taught to use them critically and with integrity. Third, I discuss how I experimentally taught AI-assisted history methods as part of my Web design, HTML, and the Archaeology of the Mediterranean (or WHAM) website-building assignment which serves as my CSAS Studio capstone project. In specific, I show how our preliminary use of LLMs resulted in the generation of satisfactory, but not outstanding, coursework marked by hallucinations (i.e., the generation of false or fictional data based on an LLM’s attempt to sort what comes next in a phrase or image based on probability; see Bowen and Watson Reference Bowen and Watson2024), but also how this experiment has opened a path to more constructive AI literacy pedagogy in future. Overall, this study reveals how an ancient history project like WHAM can provide students with AI literacy skills while also fostering the humanistic learning outcomes that history courses have traditionally championed (LaRock Reference LaRock2026).
‘Wild spirit’ and Trojan horse: defining AI-assisted history
What is AI-assisted history? In 2016, the American Historical Association (AHA) defined ‘history’ as ‘the study of the human past as it is constructed and interpreted with human artifacts, written evidence, and oral traditions’ (American Historical Association 2016). In late July 2025, the AHA issued a document entitled ‘Guiding Principles for Artificial Intelligence in History Education’. Although this document is more focused on the limitations of AI vis-à-vis historical analysis and the justifiable need for human interlocutors to remain the primary historical data collectors and interpreters, it does recognize that today’s study of ‘human artifacts, written evidence, and oral traditions’ is going to be approached with AI tools (such as LLMs), which they refer to as ‘undeniably powerful’. The AHA discusses the use of LLMs in reference to historical research in the following manner:
LLMs produce text using an algorithm to select each word from existing books, articles, images, and other media, including AI-created sources. AI texts do not reflect truth; rather, they echo and synthesize, sometimes poorly, sources on which the model has been trained (American Historical Association 2025).
Based on this passage, AI-assisted history seems to have two key elements or principles. First, AI tools, and most notably LLMs, can provide historical responses that ‘echo’ human modes of explication because they are trained on human languages, data types, and modes of argumentation. Moreover, AI tools can rapidly ‘synthesize’ (and structure; see Kansa Reference Kansa2026) massive datasets that might take a human days or even years to comb through to recognize patterns, and such thematic results can provide researchers with new forms of evidence for their interpretations. Second, AI-assisted historical analyses do not reflect ‘truth’ (although the word choice here is triggering since one could argue that even human historians cannot achieve this ideal) and require a critical examination of their training data before they are used to construct historical narratives. Instead, it seems that the ‘synthesized echoes’ that LLMs can produce should only be trusted by human interpreters who are trained in AI literacy and who can critically interpret the prompt results in relation to human-constructed, factual knowledge. In this sense, LLMs can only ‘mimic’ human-historians (indeed, they have also been called ‘stochastic parrots’; see Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021). Their syntheses and analyses may seem useful, but they can also act like Trojan horses – Greek gifts that can only be accepted if historians know how to use them (see also Campbell Reference Campbell2025; Graham Reference Graham2026).
Despite the definitional awkwardness of some of the AHA’s ‘guiding principles’, the institution’s attempt to problematize historians’ current relationship to AI is significant as it hints at the underlying tensions that many teachers are currently feeling. Most importantly perhaps, the AHA reasserts the need for human expertise to lead the historical enterprise and stresses how historical work can provide learners with a transferable skillset centred on critical thinking and social analysis. From this perspective, AI is viewed as one of the many tools that have been brought into historical research over time including concordances, the Dewey decimal system, microfilms, relational databases, and Google searches (cf. Heredia Chimeno Reference Heredia Chimeno2026). Hence, historical work in the age of AI might be best defined as ‘AI-assisted’ history as opposed to a new type of ‘reverse centaur’ scholarship where machines do all the work and simply eclipse the human interpreter completely (see Doctorow Reference Doctorow2026 for the ‘reverse centaur’ metaphor; see also Mollick Reference Mollick2024, on the notion of ‘co-intelligence’). It is this type of AI-assisted history that will have implications for how AI literacy is taught in the history classroom, perhaps as soon as this year.
How is AI ‘poised to rewrite history. Literally’?
Having provided a rough definition of AI-assisted history and its nature, we can now turn to the question Bill Wasik (Reference Wasik2025) posed in AI is Poised to Rewrite History. Literally: i.e., how might AI change the way that history is written and understood? In the article, Wasik interviewed several historians about how they were experimenting with AI. At first glance, most of the results were positive. Stephen Johnson, a historian who works for Google Labs, discussed using Google NotebookLM, a personal research assistant powered/trained by Google Gemini 1.5 Pro (in 2025), which restricts its AI analysis to (up to 300) sources (e.g., .pdf articles, pasted text, slides, websites, and YouTube videos) the researcher uploads, as opposed to drawing on the entire internet (https://notebooklm.google.com/). NotebookLM can summarize books, provide timelines and mind maps, and even create audio and video podcasts that explicate latent source themes. The built-in chatbot can answer Johnson’s prompts about creative ways to structure his research or how to craft narrative arcs. Its mind map feature can emphasize sub-themes that provide specialists with a ‘sideways glance’ into tangential topics. Thus, NotebookLM shows how AI tools can help specialist scholars organize their material, seek out its key themes, and provide new avenues for scholarly inquiry. After all, LLMs do not get tired and are often unrestrained by social norms, so they can provide ideas in quantity.
Another scholar profiled by Wasik is Mark Humphries of Wilfred Laurier University who writes the Generative History Substack (https://generativehistory.substack.com/). Humphries works with AI tools to transcribe and analyse large datasets of handwritten documents from the Canadian fur trade (Humphries et al. Reference Humphries, Leddy, Downton, Legace, McConnell, Murray and Spence2024). If a graduate student researcher were assigned to go through such documents in an analogue fashion, it might take them months to decipher handwriting, identify traders, and establish trading network connections. The present author engaged in a similar task in graduate school while working with Latin epigraphy, (i.e., the analysis of ancient Latin inscriptions on stone), a historical dataset which proliferated throughout the Mediterranean during the Roman imperial period. In specific, I worked with the (pre-digitized) Corpus of Latin Inscriptions (CIL; https://cil.bbaw.de/en/) examining hundreds of inscribed documents from Roman North Africa to uncover local modes of honouring the Antonine imperial dynasty (i.e., the family of the second century C.E. emperor, Marcus Aurelius; see Gordon Reference Gordon2004). However, since LLMs can transcribe documents and mine connections within large-scale datasets at an exponentially faster rate than analogue human researchers (my research took around three months), historians can now gain more time to interpret newly revealed data trends (Humphries et al. Reference Humphries, Leddy, Downton, Legace, McConnell, Murray and Spence2024). Moreover, an LLM can remember and learn from both the new data it creates and user-provided information. Analogue research does not provide this ‘multivocal’ discursive scaffolding, unless a researcher converses with other practitioners. In this sense, a chatbot can engage in forms of autodidacticism and reflexive analysis that are different than those practiced by humans. These AI features can then potentially generate plausible – albeit different – historical perspectives that allude human observations thus broadening the range of interpretations of past events, and perhaps our understanding of why they occurred.
Other scholars Wasik interviewed seemed more sceptical about AI’s impact on historical research. Non-fiction writer Charles C. Mann was surprised by LLMs’ erroneous hallucinations. Fred Turner, a communications professor at Stanford, was struck by how chatbots replied in a ‘middlebrow voice’, which might be useful for popular non-fiction, but could be deemed too colloquial for scholarly discourse (see also Campbell Reference Campbell2025 on scientist vs. ‘scientific communicator’ voices in relaying historical information to the public). Another problem highlighted by historian Lara Putnam is that generalist LLMs can be pre-trained on unrepresentative datasets that privilege existing power structures and reaffirm biases (Crawford Reference Crawford2021; Foka and von Bonsdorff Reference Foka and von Bonsdorff2025; Kansa Reference Kansa2026; Putnam Reference Putnam2016). Many historical datasets currently available in digital form are in English, originate in developed countries, come from official (and often non-handwritten or oral) sources, and are created by males (Kindt Reference Kindt2024). This renders many cultures, non-English sources, and the voices of women, the illiterate, or the poor, inaudible. In addition, further studies show that AI prompt results can sometimes be biased and inaccurate since their LLMs are trained on out-of-date datasets that do not reflect the most current historical research (Campbell Reference Campbell2025; Magnani and Clindaniel Reference Magnani and Clindaniel2025). Finally, given the massive amount of response data being produced by LLMs themselves, much of which can be of low factual quality and is now known colloquially as ‘slop’, it is also feared that LLMs will end up training on their own earlier products, creating a recursive cycle of ethically questionable LLMs developed from outdated and randomly scraped data (Kansa Reference Kansa2026).
Wasik’s subjects also observed that the time saved through AI collaboration does not always improve on the historical product, especially if the human collaborator does not understand the AI-generated material and tends to consider the chatbot results as fact (as many untrained students do). This phenomenon is likely what the American Historical Society was referring to in their ‘Guiding Principles’ when they suggested that ‘AI texts do not reflect truth’ (American Historical Association 2025). Such observations raise serious questions: Is the analogue historian who collaborates with the LLM still the same historian who writes without AI, or is that historian now a cyborg of sorts (or a ‘necromancer’ [Graham Reference Graham2025], or ‘reverse centaur’ [Doctorow Reference Doctorow2026])? What are historians losing when they begin to collaborate and teach with AI? Is saving time at the expense of ‘slow’ historical data collection a good thing (cf. Caraher Reference Caraher, Walcek Averett, Gordon and Counts2016)? How is our knowledge of the past going to change? These questions lie at the heart of the historical enterprise going forward, but some modes of solving, or at least debating them, may be found via a brief discussion of AI literacy, a skillset that both we as historians and our students will likely soon require.
AI literacy: a new required skill set for history teaching and learning
Given that our students’ ‘digital-native’ identities are changing classroom experiences and that LLMs’ ability to save time, make new connections, brainstorm ideas, and write middlebrow summaries are changing historical methods, it seems clear that teaching history in the late 2020s will be different than ever before (Vee et al. Reference Vee, Watkins and Bruff2026). Yet, in Teaching with AI, by Bowen and Watson (Reference Bowen and Watson2024), the authors suggest that all is not lost. They argue that traditional skills from the liberal arts, such as critical thinking or ethical reasoning along with human notions of taste, design, or insight, can help to develop AI literacy abilities that students and scholars can apply to a range of jobs and research formats (cf. Heredia Chimeno Reference Heredia Chimeno2026; LaRock Reference LaRock2026).
AI literacy involves teaching people the skills required for evaluating AI tools critically and realistically in terms of their features and limitations and for employing methodologies leading to balanced, ethical, factual, logical, and high quality/above average results. Bowen and Watson (Reference Bowen and Watson2024) emphasize the following skills as key to AI literacy: Problem formation involving both divergent (i.e., digital) and convergent (i.e., analogue) thinking; critical selection of LLMs to fit research/pedagogy goals; preparing balanced and diverse datasets for pre-training; precision prompt engineering according to task (i.e., the active/verb-centred task, e.g., ‘create’), format (i.e., the output, e.g., ‘a book review’), voice (i.e., the language style, e.g., ‘in the style of a college student’), and context (i.e., further context e.g., ‘for Gen Z students at a STEM school’), iteration (i.e., the need to build on and refine prompts within a chat to arrive at a desired response); and leveraging AI’s non-human lack of inhibitions for increased quantity of ideation. Honing these skills will allow AI’s future users to use these LLMs more ethically and effectively leading to fewer hallucinations, more ideas unrestrained by human expectations, more nuanced evaluations of results, and better, deeper answers that are crafted collaboratively between divergent digital and convergent analogue approaches.
If AI literacy is gradually integrated into historical methods courses or even select class lectures moving forward, the chances that students will only use AI to craft their research assignments or see LLMs’ knowledge products as ‘truth’ will likely decrease. Instead, students will begin to see LLMs as the tools that they are, learn to evaluate and use their functions critically, and cultivate best practices for how to engage in human-led, yet AI-assisted, history writing. At the same time, AI literacy is a transferable skill, which, if taught to students within the context of history classes, might even augment the value of such classes in the eyes of technocrats who seek to decrease students’ participation in humanities courses in favour of STEM classes purportedly aimed at increasing students’ job opportunities and return on education investment (Diermeier Reference Diermeier2024; Heredia Chimeno Reference Heredia Chimeno2026). According to the World Economic Forum’s Future of Jobs Report for 2025, Big Data Specialists and AI and Machine Learning Specialists are the fastest growing employment sectors in percentage terms (World Economic Forum 2025). Thus, if history classes start embedding AI literacy into their curricula, they will be directly fostering skillsets that students will require for the technology-oriented, mid-21st century, job market. And, by teaching STEM undergraduates in particular about AI-assisted history (as described in the case below), perhaps these same students will develop future LLM tools that more accurately and efficiently analyse the past than the apps currently available.
Teaching AI literacy and the history of the ancient world: the web design, HTML, and the archaeology of the Mediterranean (or WHAM) project
Having discussed the potential value of integrating AI literacy into existing history courses to create a form of AI-assisted history, this section relates how I brought AI-assisted learning into my CSAS Studio course. Since 2022, Wentworth Institute of Technology has had a major degree housed in the School of Sciences and Humanities called CSAS (Computer Science and Society). The program’s goal is to provide students with both computer science and humanities skills so that they can work in industries where digital tools collect, translate, and analyse data for social improvement, such as non-profits, government agencies, heritage, or climate awareness (https://wit.edu/academics/sciences-humanities/degrees/undergraduate/computer-science-society). In the CSAS program, students must take two sections of CSAS Studio where digital tools are applied to the analysis of humanities topics. Being an archaeologist specializing in Roman Cyprus (ca. 30 B.C.E.–330 C.E.), my studio topic has been digital archaeology, a data-rich humanities discipline that captures students’ imagination due to its famous monuments, artefacts, and personalities (Graham Reference Graham2026; Perry Reference Perry2019; Walcek Averett et al. Reference Walcek Averett, Gordon and Counts2016). I first taught CSAS Studio in the fall semester of 2022 and have completed four iterations so far up to the fall of 2025. Each year I alter certain elements of the curriculum to reflect changes in digital technologies, and so by 2024 and 2025, AI literacy became a core course element.
For course content, I utilize Renfrew et al.’s (Reference Renfrew, Bahn and DeMarrais2024) Archaeology Essentials textbook because it is one of the most common texts used to teach basic archaeological rudiments, from excavation and dating methods to environmental and cognitive archaeology. Along with each chapter, I weave in basic lessons on web design coding in HTML, CSS, and JavaScript (using W3 Schools’ free online tutorials, https://www.w3schools.com/). Moreover, I introduce students to several digital tools that can be easily embedded into a website, such as QGIS (a free and open-source GIS; https://qgis.org/), TimelineJS (an open-source interactive timeline builder; https://timeline.knightlab.com/), Tableau Public (a free online data visualization platform; https://public.tableau.com/app/discover), or ThingLink (an online image and video annotation platform; www.Thinglink.com).
Then, drawing on students’ new knowledge of archaeological concepts and their nascent skills in web programming, I assign them a capstone project, that is, a final multi-week group project where the students must collaborate during class time to achieve a common project deliverable consisting of original research, writing, programming, and a final presentation of their work to their peers. For the CSAS Studio capstone, the student teams must code a website from the HTML up focused on introducing the archaeology of a specific Mediterranean region (e.g., the island of Cyprus or the Italian peninsula) to the public. I ask for the website to be public facing since recent scholarship in classical reception studies has shown that ancient exempla can help explicate challenging 21st-century societal issues, such as decolonization, famine, climate change, and technological disruption (Kindt Reference Kindt2024). In addition, my students also include design and engineering majors, and so the primary audience for their own products will be the general public, not ancient world experts. As a result, creating a public-oriented deliverable gives them a chance to hone their marketing and outreach skills. The website assignment is known as the Web design, HTML, and Archaeology of the Mediterranean Project (or WHAM), and I have organized it for groups of three to four students and assigned each group a Roman imperial province as a research focus. Therefore, when the students share their work at the end of semester, they are all introduced to the archaeology of several Roman provinces and can comparatively discuss best practices for elementary web coding and aesthetic digital design.
During most iterations of this class, I resisted using AI because I wanted students to learn coding in a ‘hands-on’ manner. I also wanted them to undertake historical research on archaeological topics using both traditional sources as well as open-access datasets, like those found on Open Context (a free, open access, online resource for primary archaeological data, https://opencontext.org/; Kansa Reference Kansa2026). However, as discussed above, all (at least take-home, digital) assignments are now AI assignments. And so, in an effort to integrate AI literacy skills into learning outcomes, in 2024, I taught my first lectures on how AI might be used both for HTML coding as well as creatively iterating on web design and collating historical content. It was this experience that taught me and my students to beware LLMs bearing gifts, especially if a student is unfamiliar with a specialized topic like Roman archaeology. Yet, it also revealed the power of AI to lower barriers for specialized tasks and raise bars for traditional ones, if AI literacy was taught in a manner aligned with academic integrity through AI–human collaboration (for more on AI–human alignment, see Mollick Reference Mollick2024; for a discussion of how AI tools can be leveraged to help neurodiverse students learn about the ancient past, see Heredia Chimeno Reference Heredia Chimeno2026).
Once the students had learned some basic HTML and archaeological concepts, I asked them how one might use AI (in this preliminary case, the free version of ChatGPT 3.5; for the current version, see www.chatgpt.com) to help us write our website code and to suggest archaeological content that a non-specialist user might find interesting. We began by providing ChatGPT with a basic prompt: ‘I am making a website in HTML for the archaeology of Roman Cyprus. What key web page elements should I include?’ ChatGPT responded with the typical navigation bar features one would expect for an archaeological website: Homepage, Sites, Artefacts, Timeline, Cultural Insights, Research and Discoveries, Educational Resources, Events and News, About and Contact. So far so good, but the archaeological examples it provided were extremely generic, what I might call ‘C-level’ or satisfactory research work, hovering around Fred Turner’s ‘middlebrow’ voice intelligence (Bowen and Watson Reference Bowen and Watson2024; Wasik Reference Wasik2025).
The second step was to ask ChatGPT for the HTML code for a basic website, so we prompted: ‘Could you demonstrate a (website) example for us?’ ChatGPT was happy to oblige within seconds. For the students who had spent several weeks coding ‘by hand’ to make even basic websites, ChatGPT’s instant response was like a magic trick or ‘Greek gift’. However, once we pasted the HTML into our integration development environment (in this case, Visual Studio Code; see https://code.visualstudio.com/), the result was underwhelming, again, simply satisfactory work (or worse). Sure, it was an archaeological website about Roman Cyprus, but it looked worse than what students could create with their limited HTML skills (Update: By early 2026, ‘vibe coding’, where web developers simply prompt agentic AI models to create websites, has now become commonplace, so it is assumed that future AI-coded websites will be much more robust; see Sweet Reference Sweet2026). Moreover, ChatGPT was able to find free images, but its code links were constantly broken since the Chatbot was generating non-existent file paths.
A final step was to ask ChatGPT to generate some research content, so we prompted: ‘What are the most famous artifacts from Roman Cyprus?’ ChatGPT responded with a list of well-known monuments, including the UNESCO (2025) Paphos mosaics (see Figure 2) and the Roman coinage (Parks Reference Parks2004), but the information was extremely generic and included several hallucinations that only an expert could catch. In this case, the professor happened to be a specialist in the unique island archaeology of ancient Cyprus, thus proving that at least in the fall of 2024, human expertise was still required to avoid LLM hallucinations (cf. Gordon Reference Gordon2024; Magnani and Clindaniel Reference Magnani and Clindaniel2025).
Early Roman floor mosaic from the House of Dionysos, at the UNESCO world heritage site of Paphos, Cyprus, showing the Triumph of Dionysos (2nd Century C.E.).
Note. Photo by author.
However, if a student with little archaeological knowledge simply relied on ChatGPT’s ‘truth’, their information would be erroneous. These hallucinations included a discussion of the Amathus sarcophagus (see Figure 3; Metropolitan Museum 2025), which most scholars date to the era of the city kingdoms on Cyprus in the 5th century B.C.E., but which ChatGPT claimed was of Roman date. ChatGPT also suggested that imperial portrait busts were common finds on Cyprus during Roman times, even though very few have ever been found (Fejfer Reference Fejfer, Wriedt Sørensen and Winther Jacobson2006). When I iterated and asked ChatGPT ‘Could you suggest a famous portrait bust from Roman Cyprus?’, it did discuss an imperial sculpture, but it was a bronze over life-size statue of Septimius Severus, not a bust (Kleiner Reference Kleiner1992). Ultimately, showing the students how such prompts could produce erroneous or skewed responses when pointed out by a specialist human researcher proved to be a key teaching moment. Hence, the learning outcome to express to students here was that some factual content knowledge is still required to work with LLM prompt results or, at the very least, they need to be checked against fact-based scholarly sources, such as academic monographs, specialist articles, or websites. In Doctorow’s (Reference Doctorow2026) terms, these examples showed that the human head still controlled the ‘centaur’s’ AI body.
The Amathus Sarcophagus from Amathus, Cyprus. Limestone. (2nd Quarter of the 5th Century, B.C.E.).
Note. Photo courtesy of the Metropolitan Museum of Art, New York. https://www.metmuseum.org/art/collection/search/242006. Open access, public domain.
The results of these initial experiments using ChatGPT for ancient world website production were surprisingly average in terms of creativity, research quality, and technical aptitude. ChatGPT’s website content outlines, ‘wire-frame’ HTML files, and responses to specific queries on artefact types were either merely satisfactory work (although I would not call it ‘slop’ exactly) or somewhat hallucinatory. I think both myself and the class were surprised that ChatGPT could not do better with simple prompting, perhaps because it was suffering from what Ifigenia Drami (Reference Drami2025) has identified via her work on Plato and AI as ‘dokesisophia’ or ‘the ability to appear wise’. Nevertheless, it was later surprising to note that several of my student groups decided to use ChatGPT’s assistance to code their final projects despite these limitations. The result was that their sites looked generic, but, interestingly, since they had learned some HTML before I introduced ChatGPT, they could read the chatbot’s HTML code and manipulate it. This practice revealed to me a truism about teaching with AI: students must know the analogue fundamentals/facts about the technical and/or historical information they are digitally prompting about, or it can be of very little use to them since LLMs often produce average work or results marked by hallucinations. Of course, with the benefit of hindsight, more prompting and working with a range of LLMs beyond ChatGPT would have likely improved my students’ results.
Conclusion: whither the WHAM project with AI? Current observations and next steps
The initial AI trial in the 2024 iteration of CSAS Studio has allowed me to reflect on the ‘gifts’ that LLMs can bear and where to go next with AI in this course. Returning to some of the course material in preparation for this article, one of my first observations was that ChatGPT seems to have become more sophisticated (or maybe more ‘dokesisophistic’?), even within the year and half since I first used it in class. As discussed in the Introduction, AI computing power and cognitive density is not plateauing, but rather accelerating. In addition, access to publicly available LLMs has expanded since ChatGPT was released in 2022. Now, in addition to the free version of ChatGPT, most students and professors can access some version of Google Gemini (and NotebookLM), Anthropic’s Claude, X’s Grok, Meta’s AI, or Microsoft’s Copilot, and a comparative analysis of the pros and cons for each platform can be assembled vis-à-vis researching the ancient world (see below). Furthermore, within ChatGPT, there are literally hundreds of bespoke GPTs that are free to use and that can improve one’s prompt results, such as ‘AI Humanizer’ (with ca. 8 million users according to ChatGPT; see https://chatgpt.com/gpts) or ‘Scholar AI’ (with ca. 4 million users according to ChatGPT). There are also dozens of public GPTs focused on archaeology (e.g., ‘Etruscan Archaeology Tutor’ GPT) or classical languages and cultures (e.g., ‘Latin Teacher’ GPT; ‘The Stoic Tutor’ GPT). By early 2026, another form of AI specialization that has emerged is the creation of ‘local’ LLMs where an LLM (e.g., open-source Llama or DeepSeek) can be run locally (via a desktop application like LM Studio, https://lmstudio.ai/) on a host computer (e.g., Mac Minis are currently a popular choice; Qi and Wen Reference Qi and Wen2025). Such local LLMs could prove useful for teaching and research on the ancient world, especially if ethically sensitive datasets (e.g., burial remains) were under analysis or data needed to be rapidly analysed in a remote archaeological location lacking internet access.
Returning to fall 2025, when I last taught the CSAS Studio, I retried our admittedly simplistic prompts from 2024 with a range of commonly used LLMs to see if basic HTML designs could be improved and whether more complex features (like JavaScript buttons or a search bar) could be provided. In specific, I utilized this prompt: ‘Could you please provide me with the HTML code for a basic website focused on the archaeology of Cyprus in the Roman period. I would like the colors gold and white to be used and for there to be some images of the island’s landscapes and archaeological sites, but all archaeological sites should be of Roman period date. A picture of a temple might be good in particular. Thank you!’ I tested this prompt on Claude Sonnet 4, Meta AI (Llama 3.1), Grok 3, ChatGPT 3.5, and MS 365 Copilot 2025 version. All these LLMs were able to produce a basic wireframe website, but apart from Meta Llama AI (free tier, running on Llama 3.1), all the results were aesthetically and functionally mediocre. Yet, when I vibe coded and prompted Meta to add buttons with links to external websites and a search bar that would highlight the search words on the webpage, Meta obliged (see Figure 4). Even with my rudimentary knowledge of JavaScript and HTML (my doctorate is in classical archaeology, not in computer science), I was able to take this code, integrate it into my wireframe website, and voila, it worked, in an almost magical, yet somewhat unsettling, way (cf. Graham’s [Reference Graham2025] ‘necromancy’ and Kansa’s [Reference Kansa2026] ‘magic oracles’). And, despite the websites’ ‘middlebrow’ aesthetics (e.g., bland website colours, basic fonts, and boxy layouts), each LLM did provide a variation on a basic website, underlining LLMs’ ability to ideate for quantity.
Screenshot of a website result based on a Meta AI prompt to generate HTML code for a website for the archaeology of Roman Cyprus.
Note. Screenshot by author.
What these results indicated was that my next phase of teaching with LLMs should be active, yet guided, collaboration, in which working with LLMs is not cheating, but rather part of an AI literacy and history learning process shaped by integrity. It should be an AI-assisted history that follows the AHA’s ‘Guiding Principles’ by centring human researchers, getting students to question LLMs’ dokesisophistic ‘truth’, and treating LLMs as ‘undeniably powerful’ ‘Greek gifts’ that can promote human knowledge if they are critically approached.
In sum, in terms of moving forward with the WHAM project, I envision integrating AI literacy and collaborative AI assignments into my future CSS Studio classes in the following ways. First, I am approaching this class with the premise that all history production now has the potential to be AI history; all students are – or will be – using AI to learn; and all jobs will likely change, even indirectly, because of AI (Bowen and Watson Reference Bowen and Watson2024; Graham Reference Graham2025; Stanford HAI 2026). If these prognostications are true, then not teaching AI will prevent teachers from fulfilling our most basic duty (at least at most STEM schools): preparing our students for academic careers in the 21st-century, technology-oriented job market.
Second, history – including ancient history and archaeology – with its large datasets, ethical questions, and required skillsets involving critical thinking and argumentation, is an ideal discipline through which to teach AI literacy. Moreover, if approached critically, AI-assisted history can also provides historians with new modes of evidence analysis. Such realizations are significant because they reveal history’s renewed pedagogical relevance in higher education and allay rumours of its death as a ‘soft skill’ liberal art (for more on the teaching of STEM skills in history courses, see Carrell et al. Reference Carrell, Keaty and Wong2020; Campbell Reference Campbell2025; Diermeier Reference Diermeier2024; Park and Cho Reference Park and Cho2022).
Third, regarding my course lecture topics and schedule, I will still teach the ‘analogue’ skills of HTML and archaeological practice. However, I will then switch to a ‘digital’, critical AI literacy unit, whereby we will learn problem formation, LLM tool comparison, data ethics, prompting, iteration, result evaluation, and AI ideation (see also Walter Reference Walter2024). One lecture will also be devoted to integrity-based LLM use, where LLMs are approached as research collaborators and not ‘cheating machines’ (see also, Bertram Gallant and Rettinger Reference Bertram Gallant and Rettinger2025; Mollick Reference Mollick2024). In preparation for students’ website development projects, I will introduce them to collaborative AI apps like Google NotebookLM, useful coding LLMs, like Meta Llama, and ethics-oriented LLMs like Claude with its democratically named, ‘Constitutional AI’ (https://www.anthropic.com/claude; Bai et al. Reference Bai, Kadavath, Kundu, Askell, Kernion, Jones, Chen, Goldie, Mirhoseini, McKinnon, Chen, Olsson, Olah, Hernandez, Drain, Ganguli, Li, Tran-Johnson, Perez and Kerr2022). I will then get students to combine their ‘analogue’ knowledge of HTML and archaeology with their new AI literacy skills and to apply them to developing new archaeological websites (cf. Heredia Chimeno Reference Heredia Chimeno2026). This approach might be akin to what Mark Humphries (Reference Humphries2025) has recently deemed a ‘middle course’. I am uncertain what the results will look like given AI’s ‘wild spirit’, but my assumption is that student projects will move beyond ‘C-level’ or satisfactory work and contain fewer factual hallucinations. Indeed, I hope my ‘gift’ to students will not only be a novel sense of preparedness for an AI-assisted future, but also that I leave them with a new passion for exploring the ancient past, even if this new interest is sparked by the ‘Trojan horse’ of learning about ethical and integrity-based AI.
Open access
