This post is part of a series commissioned by the Royal Historical Society on ‘AI and History Teaching in UK Higher Education’ in Autumn 2026.
The series includes a range of perspectives on the uses and limitations of Generative AI and LLMs from history educators and research academics.
Participating authors provide ideas and suggestions for teaching practices, shaped by their own scholarship and teaching experiences in contribution to the ongoing conversation about the impacts of AI on teaching and learning history at UK universities.
By Dr Dave Hitchcock
The core question posed by AI is: would you like me to do that for you?
The core answer that historians should most often give in classrooms, I argue, is: no.
***
The French historian and national hero Marc Bloch began a book now called The Historian’s Craft with a question from his son: ‘tell me, Daddy. What is the use of history?’[1] Bloch’s answer was long and beautiful and written by hand in notebooks in an occupied country in the middle of a shattering war, and it was necessarily unfinished. I imagine readers will know why. But it is so very true to the nature of what we do and what we teach, that the thoughts of one of greatest historians of the 20th century on his craft will always be unfinished, just as history itself should be. We do not get to be ‘done’ with all this, we do not get to stop thinking about it. No historical interpretation is so durable that it should be immune to change and modification. History never leaves us. We explain it as best we can, as it will in turn explain us to the future. We should insist on doing that vital work ourselves.
History is a human craft; it is a domain of human knowledge unto itself, which has existed for a very long time indeed. It is invoked constantly by other disciplines and is now an academic subject with its own rules, norms, ethical and professional obligations. It is recursive, which is to say that ‘History’ has a history. Asserting that history is for humans to do is not to say that they must always produce it unaided by technology or anything else, of course. But we do have duties from which we cannot shrink, and they should be clear in our classrooms. When teaching I introduce this sense of obligation to the past with a trite adage that our first duty as historians is: ‘don’t make stuff up’. We form arguments about the past yes, but we work hard to ground them, and we are honest that it is us who crafted them. Robert Colls puts it thus: ‘without the story — or the ‘narrative’ — any explanation of change either disappears or dissolves into nouns.’[2] We all must choose what we care enough about, or what we think is important enough to merit inclusion, never mind prominence. If we allow students to outsource any of that intellectual effort to an algorithm, they are practically guaranteed to get back a tidy package of nouns and might be tempted to call it a day.
In history, there is a strong argument that quite a lot of the ‘work’ is in fact how we do quite a lot of our thinking! Our best work becomes our thinking on the page.
Non-historians have already noticed this problem. The tech journalist Ed Zitron, a noted critic of AI companies, argues: ‘the problem I have with most people using LLMs is the delineation between outsourcing work and outsourcing thought’.[3]In history, there is a strong argument that quite a lot of the ‘work’ is in fact how we do quite a lot of our thinking! Our best work becomes our thinking on the page. If you have an AI agent do a task it is virtually guaranteed that you put less thought and intention and probably time into it.[4] Otherwise what is the point? And history is not mere content produced for consumption or merit, not even when presented in those ubiquitous undergraduate essays everyone seems so certain we must discard. Its production is a journey, we and our students and our readers and audiences are all travellers, and we should be flattered when others read our accounts of discovery and exploration and question us on the merits of our insights.
There is a whiff of carpentry about the whole business. Through a multitude of choices that we make and deliberations we agonise over, we whittle down a series of disconnected ‘facts’ or pieces of evidence, ‘sources’, and ‘arguments’, and we form the knotty assemblage into a new whole. We argue with the grain of our sources or carve against them, but either way, we still shape and we still choose. There is no tidy manual for historians marking what matters and what doesn’t, there is no balance sheet to add it all up.
There are no iron laws of history and it teaches no easy lessons either. There is no algorithm that will do this job of choosing for us, in fact, to relinquish this burden of historical choosing is to give up on the profession of history itself. Moreover, the craft of history is not as easily divisible into smaller ‘tasks’ as some may think. How many strokes of the knife before the wood became the carving? No historical explanation is as simple as it may appear in our rearview mirrors. Mere prevalence does not equate to importance, and what is commonly observed or, more importantly, commonly recorded about the past is not always true.
Crafting history is far from an ineffable mystery, historians are no secret society, and we ought to explain ourselves clearly—our interpretations and choices, our priorities and principles—to all who read or listen to us.
Large-language models cannot and will never, by definition, grapple with any of this complexity. At the same time, crafting history is far from an ineffable mystery, historians are no secret society, and we ought to explain ourselves clearly — our interpretations and choices, our priorities and principles — to all who read or listen to us. Large-language models have a problem doing that too. We owe it to our students to help them see why their interpretations of the past matter as much as ours do, and why they have the same responsibility to explain their thinking as we do, and why, therefore, they should develop and refine their historical thinking by themselves, for themselves, in dialogue with and inspired by their sources.
We should all be deeply worried about the potential control over the shape of historical narratives which is presently being devolved, by default, to companies like OpenAI, or DeepSeek, who have no stake in the production of ‘the past’ which is separable from their commercial interests.
The researching, revising, and teaching of history carry very human consequences which cannot be avoided even if one has dodged the dreaded act of writing by asking a machine to do it. Outsourcing even the smallest aspects of historical writing to indifferent agents carries terrible risk. There are obvious sites where an unthinking ‘efficiency’ applied to the production of historical information can cause immense harm. A national curriculum that teaches a ‘bad’ or incorrect or manufactured history at scale is already a terrifying thing we must guard against, the stuff of dystopia and dictatorship. Such a thing reproduced and proliferating at speed is a problem of even greater magnitude. The past persuades. So-called national histories can, and indeed have, assembled nations and they can send those nations to war. We should all be deeply worried about the potential control over the shape of historical narratives which is presently being devolved, by default, to companies like OpenAI, or DeepSeek, who have no stake in the production of ‘the past’ which is separable from their commercial interests. There is a politics to all this, and it is precisely politics that we’re seeing when a model fails to answer a question about Tiananmen Square, or when ‘Grok’ ranks historical populations by ‘average IQ’. Historians should work hard, and show their working, to try to earn the trust of their readers.
AI models demand that trust implicitly without explaining themselves, while disowning their own violations or errors as problems of mere programming.
There is almost no value, and much danger, in automating the very basics of the historical process: the critically conscious sourcing, analysing, and contextualising of primary or secondary historical information. I have seen no sound evidence that ‘training’ history students to use an AI makes them any more employable, or any more skilled at the processes core to what we do. In fact, recent surveys of history students suggest they already feel disempowered and de-skilled by the technology when they do use it.[5] Graduate employment and attainment data are noisy enough as it is, so I would be cautious about inferring any relationship between exposure to AI and graduate outcomes. There is also a problem here of basic accountability. I am still accountable for errors in my work if an LLM produced them, and that principle holds true for my students too. I should therefore think very carefully before introducing them to tools whose risks they may not fully appreciate, and over which they do not have precise control.
The truest challenges of doing history are not strictly factual (although much work goes into establishing and exploring fact), they are interpretive and connective. In one way or another, we address these problems with our students in class.
This fundamentally personal nature of the historian’s craft and the responsibilities of doing it has long been apparent. In 1941 Marc Bloch thought about historical explanation using terms like cause and consequence, and he understood these to represent structural problems that all historical thinkers must confront themselves, such as ‘the problem of the individual and his differential value’ which inherently requires a personal answer, and the problem of ‘determinant’ facts or acts (what ‘matters’ versus what ‘determines’). The truest challenges of doing history are not strictly factual (although much work goes into establishing and exploring fact), they are interpretive and connective. In one way or another, we address these problems with our students in class, and we try to model some solutions to students while pushing them to find their own. Even ignoring these problems is, in its own way, a form of address to them, and frankly not a satisfactory one.
History classrooms are not engaged in idle intellectual chatter which has no consequences in the world or for students themselves. History students are not taking in a dense and knowledge-rich curriculum or learning skills and techniques for source criticism, information retrieval, argument assemblage, contextual thinking, multi-causality, cross-checking, and structural honesty which have no bearing on the wider world. Just because it is not easy to prove the discipline’s bearing on our world does not mean it has none. And how we choose to teach our students, down to the topics and tools and techniques we deploy, necessarily carries our priorities into our classrooms in a way I think we are obliged to explain.
This may all sound melodramatic — ‘historical thinking is a special kind; algorithms can’t do it’ — dismissible as merely clambering up a moral or civic high ground and then calling it ours. But I hope none would deny the immense normative force of historical narratives and understandings. We shy away from, or react with disgust to, outsourcing a variety of other scholarship and creative production with less enduring impact than history. It behoves us to care about who and what shapes narratives of the past, and to recognise that to some extent historians have always been obliged to police the boundaries of the subject and to enforce forms of academic honesty on would-be practitioners, to insist on an honest accounting of sources, questions, quotations, and methods.[6] We do this work all the time in our classrooms. Not everything gets to ‘be’ history, despite history theoretically getting to be about everything. Not all opinions on the past are created equal.
I think it is critically important to understand that, based on how these models are programmed to operate, LLMs cannot do these human and interpretive things that I am insisting historians must do to produce any history in the first place. They might do other things which aid a historian, but such use must be very carefully controlled in my view, and the use-cases are narrow and tend to collapse upon scrutiny. And if you see, as I so often do, that the work of history is actually where a lot of the thinking about history occurs both for us and for our students, you might become as cautious as I am about introducing tools which offer to ‘do that for you’.
There are far too many deliberations, choices, and cares in the production of a history for it to be something statistically generated, a mere algorithmic derivative.
I also think calling a text extruded from an LLM a ‘history’ is, quite literally, a category error. There are far too many deliberations, choices, and cares in the production of a history for it to be something statistically generated, a mere algorithmic derivative. What large-language models can certainly do, ably it seems, is offer a simulacrum of historical text and interpretation, one generated very quickly and couched in adaptive prose modelled from the theft of millions of publications and from the wider public internet. They can offer to ‘do’ things for us and then convince us that they have, in fact, done those things. They can and will lie repeatedly about primary sources. They will interpose themselves (as an interface or interlocutor) between our students and primary sources, as the professor of Japanese history Jordan Sand has compellingly written about.[7] In a classroom setting these simulated answers might present to us initially as tidy and as time-saving. In an exam setting or in an essay we call them plagiarism, cheating, or lying.
I see history like fire, it can be transformative or destructive, and fire needs fuel. For better or worse history is fuelled by us, we are its agents and its interlocutors, it consumes us and, in the end, we become it. So on moral grounds alone I refuse to hand control over such a force to an algorithm if I can avoid doing so, no matter how advanced a machine it might be, and I will not imply to my students that it is okay to do so.
In order to think a bit more systematically about all this, in conversation with colleagues I made a list of ‘considerations’ about AI and history teaching. I suggest before anyone uses an AI model in a history classroom to do anything at all, they sit with the following eight statements and ensure that they are happy with their own answers to them. Having done so myself, I struggle to see any reason why I would ever ask my students to use an LLM to craft a history.
***
8 Considerations for Large Language Models and the Teaching of History
1. Education is a process, not an event. A history education is an active process through which we seek to understand our pasts and presents. We learn to do history iteratively, through the process of producing it ourselves in various formats.
2. Generative AI cannot ‘replace’ historical thinking so much as generate a simulacrum of thought. AI offers shortcuts to a particular outcome. These shortcuts risk the acquisition of historical skills, because they subvert or replace many core processes of ‘doing history’.
3. LLMs might save users time on tasks like tabulating large datasets or surfacing patterns, so long as results are carefully checked. This might support practice, but it cannot inspire critical insight or reflective thought, which are the primary aims of historical scholarship.
4. We should train history students to locate, identify, and describe sources of information and only then to learn the skills to reflect on the selection, organisation, and mining of datasets used by LLMs. Learn to do it yourself, first.
5. History teaching should explain the value of quoting accurately, citing correctly, paraphrasing sources with fidelity, and marshalling arguments carefully and specifically. LLMs do not yet reliably do these things and may never do so.
6. Unless history students can assess the information selection processes and retrieval algorithms that power a given LLM, they should treat its outputs as sources of uncertain provenance.
7. Use of LLMs should be subject to ethical and critical scrutiny, just as with any other source of information or any tool or technique.
8. Most large-language models are commercial systems, and administrator control of their parameters, tone, capabilities, acceptable answers, and programming resides with commercial firms. Their use entails significant energy expenditure and environmental degradation. These facts place significant limits on how, when and why they should be used.
About the Author
Dr Dave Hitchcock is Reader in early modern British History at Canterbury Christ Church University.
He writes about poverty and homelessness, and is at work on a second book. He is a Councillor of the Royal Historical Society, elected in 2025.
Endnotes
[1] Marc Bloch, The Historian’s Craft, ed. Lucien Febvre (Knoff: NY, 1949), 1.
[2] Robert Colls, ‘Apprentice to the Trade: Reflections on the Writing of History’, Transactions of the Royal Historical Society, (2026), 6, doi:10.1017/S0080440126100607, part of the ‘Common Room’.
[3] Ed Zitron, ‘The AI Hater’s Manifesto’, August 25th 2026, https://www.wheresyoured.at/the-ai-haters-manifesto/
[4] The standard term is ‘cognitive offloading’, which is a common facet of memory but which AI-use accelerates, on the assumption the AI ‘retains’ in the same way a physical artefact or digital file does, which is not guaranteed. See: Richmond and Taylor, ‘The benefits and potential costs of cognitive offloading for retrospective information’, Nature Reviews, Psychology, 4 (2025), 312-321.
[5] Jones and Peplow, ‘History and AI: A Survey of History Students in UK Universities’, History UK reports (September 2026). Available online. See executive summary. Students ‘overwhelmingly’ believe reliance on the technology degrades critical thinking and other historical skills.
[6] See Kończala and Moses, ‘Patriotic Histories in Global Perspective’, Journal of Genocide Research, 24 (2022); 153-157.
[7] Jordan Sand, ‘AI is an Anti-Primary Source Machine’, History News Network, 27 August 2026. Online Link.
HEADER IMAGE: iStock photo: Credit: Jacob Wackerhausen
