This is the paper I gave to historians about “AI” last week.
I was at Queen Mary University in London last week at a charming conference that turned out to be mainly about modern British history. I went there explicitly to talk about a huge looming problem with “historical bullshit” and populations getting much, or most, of their “history-esque content” straight from an AI-prompted answer to a search query. Below you can find the paper I wrote up verbatim.
***
My title was: “British History, Historical Bullshit, and the Crisis of ‘Information Hygiene’.”
Paper in full:
Thank you to the organisers for having me along. I’m very excited for this conference, and today I’m looking to do something a little different and rather polemical. My abstract mentions twin crises engulfing our sector. One I’ve written about elsewhere recently for History Workshop, that is the perverse “free-market” system and resulting catastrophic financial circumstances and job losses we’re seeing across UK HE. Today I’m talking about the other crisis, a crisis of “information hygiene”. I want to organise my talk using three questions, one definition, one example. I want to end up with one ethical imperative that I think we, as academic historians, now collectively must confront.
I think history is the professional craft of producing a multivalent, collective, and interpretive truth about the past. It’s multivalent because rarely is there a singular answer, it’s collective because our individual labours, whether agreeably or not, productively add up to a larger whole, and it’s interpretative in all the ways we’re familiar with here, while, I hope, retaining fidelity to our source materials and honesty with our readers. I think we historians do this using a remarkably stable, and robust, set of methods, core questions (think the 5Cs), scholarly priorities, and obligations to our source materials.[1] I think one of the things that the profession of history ought to be implacably against is bullshit. This is not a given. There are entire disciplines are quite comfortable with bullshit, and some which use it as a method, and I’m not proposing to stand here today taking free swings at them.
While our opposition to bullshit immediately sets up the huge problem of determining what gets to be bullshit, what is outright lying whether by omission, misrepresentation, or direct statement, and what gets to be ‘true’; I think the profession’s healthy embrace of multivalence in pursuit of historical truths in the aftermath of debates about postmodernism provides a workable shared understanding of what we’re on about. I’m not that worried about us. I’m really worried about, and for, people without our training as they try to navigate an ever-more hostile information space. I suspect many people here can guess why.
Question 1: what is bullshit and what, by extension, is historical bullshit?
The Princeton philosopher Henry Frankfurt’s exploration of the concept of bullshit has become famous. Today in 2025, I think it is fair to say that the trend toward speech acts and writing acts indifferent to truth has only accelerated. Opinion articles increasingly refer to our present information ecosystem as one of “total collapse”. Very recently and very rapidly, we’ve had to come to grips with the ability of language machines, so-called “AI”, to bullshit without any intent whatsoever, through the production of potentially infinite and customisable “content” through probabilistic exercises (more on this later). Behind this automated bullshit though, remains a very human set of bullshitters feeding prompts and content and priorities into the machine.
Frankfurt’s criteria for what distinguishes bullshit from other forms of dishonesty has been updated as recently as 2018:
(1) The bullshitter is indifferent toward whether what she says is true or false.
(2) The bullshitter is indifferent toward her audience’s beliefs.
(3) The bullshitter intends to deceive her audience into thinking that she is not bullshitting.
(4) Bullshitting and lying are incompatible.
If bullshit is to be “historical” then, what distinguishes it?
I would contend historical bullshit can have two sufficient characteristics, or modes of operation. It comprises statements indifferent to multi-valent historical truth as evidenced in historical sources, generally preferring singular frameworks which admit the legitimacy of no counterproposals. And/or, it deploys counterfactual reasoning (rather than exploratory claims) to make the case for a singular interpretation. I think there are problems right away with what I’m suggesting, not least that no matter how much we might pretend otherwise, historians are making interpretive judgements about the past all the freaking time. The best of our work does this honestly and transparently, based on huge, or at least understandably sufficient, bodies of evidence we’ve carefully assembled and worked our way through. We establish our positions and debate them inside a shared framework of peer review and historiographical disagreement. Historical bullshitters do none of these things, or, they do them entirely indifferent to the outcome and immune to any suggested changes in their interpretation. It’s not debate so much as a display.
Question 2: “Do LLMs dream of statistically probable numbers of sheep?”
Very recently, scholars working on ethics and technology have started to argue that large-language machines can bullshit, indeed, that this is all they are ever actually able to do. Sarah Fisher made this case in 2024 and her article has prompted interesting debates.[2] Fisher makes the point that, so far, all definitions of bullshit take for granted that the bullshitter has a mental state; has intent. And of course, LLMs are not in the business of assessing the truth of the statements they statistically produce. ‘In many instances’ though, Fisher writes ‘this process will end up generating truths’, because enough training data contains enough examples of the same ‘truth’ that it becomes the most statistically likely answer. We must ask the follow up question immediately: under whose control, and from where, are the huge numbers of training “examples” coming from? Because to change the statistical balance of that dataset is to change the “truth” emitting from the language machine. In the production of history, bad inputs make for terrifying outputs, and the cardinal sin of LLMs must surely be the necessary elision of all sources they require in order to operate.
Without any desire to enshrine “the archive”, as western and textually-based as it is, as a source of unmitigated and easily accessible historical truth, I think we need to be very alarmed by the possibility that LLMs will start to bullshit widely — and for a huge range of very human purposes which are put to them — not just about historical narrative, but about historical sources themselves. I have encountered it already, newspaper articles by Martin Luther King written after his assassination, apparently published in real newspapers from 1968, entirely manufactured and utterly inaccurate quotes from my own books, confidently narrated back to me as essay evidence. I think we are about to confront a tsunami of historical bullshit, both in and beyond our classrooms. I think contrary to appearances, these models produce deeply “dirty” information where the sources are occluded entirely and the crucial interpretive acts underpinning each sentence are completely unapproachable.
To bring this home to British history now, let’s think for a second about teaching classes on the British empire and the early modern slave trade. Historians collectively deploy a wonderful range of methods and emphases to teach on this difficult and important topic. Some might follow the money, delivering a thorough exploration of the economics of the slave trade and its centrality to origin stories of racialised capitalism, imperial expansion, financial innovation, and so on.[3] They might set a reading from Berg and Hudson on the links between the slave trade and British insurance and financing innovations in the eighteenth- and nineteenth centuries and then pair that with primary sources focused on the ways that slave ships were jointly financed and the profits of sale divided amongst elite investors, or examine the professional jobs created to preserve, grow, record, and defend those profit margins. Like me, they might instead emphasise experiences of enslavement and push their students to negotiate the difficult and multiple truths about resistance and rebellion, collaboration and constraint, and active participation inside a system built brutal by its own standards, much less ours. I’ve had good results recently setting chapters from David Doddington’s book on American slavery and old age, focusing on “‘sale, abandonment and neglect’”, and how once they became unprofitable, older enslaved people could be manumitted as a further form of enslaver control. Students grapple with the idea that even in finally “escaping” the system, the brutal economics of slavery still dictated seemingly humane practises. We walk those ideas back into the British Caribbean and look at manumission rates for Virginia and Jamaica to think about whether we have evidence of a comparable commercialisation of awarding freedom there.[4]
These are examples of nested, contingent, and hardly obvious historical truths. It took a lot of people a lot of thinking and reading time to arrive at them. In Deny and Disavow, Alan Lester writes that the ‘most notable difference’ in histories of slavery in the USA and UK ‘is that of distancing’. ‘Tens of thousands of Britons up and down the social hierarchy owned enslaved people [in the colonies] without ever encountering them.’[5] That sense of distance is double-edged, it allows forms of imperial nostalgia about a ‘good’ empire, to borrow Pete Mitchell’s term, to flourish as the baseline discourse about the British imperial past. I do not think the layered nuances we use to get to deep student understanding are reliably filtering into wider discourse on histories of the slave trade, despite much excellent work on the topic from more popular historians, the sort whose books will sell from Waterstones shelves. And while it’s encouraging to see shifts in recent decades in UK public attitudes towards the historical empire, I think the change is superficial, and easily rowed back with a bit of populist pushing. Outside of a history classroom, how often will members of the public, or students in other subjects, encounter such matters at the depth needed for understanding? How considered can their thinking be about the layer upon layer of trauma and inequality which early modern transatlantic slavery produced?
Here then is the kicker. Do we think that large language models are at all capable of finding and communicating even this level of complexity about the historical past? I have so far seen absolutely no evidence that they can do so. I think the way the models work is inimical to historical thinking, which is not, pace Sam Wineburg, an easy or natural act. Instead: the opposite: LLMs are probabilistic bullshit engines that “reason” very similarly to the counterfactual positioning I talked about at the beginning. I think the chatbot’s answers cannot help but be historical bullshit because it is always asking “how likely was it that this was the case?” rather than “how can we know what this was, and why it happened?” The past flattens inside questions like that, and answers tend to be singular and unthinking, not multivalent and alive to context. Where do we think millions of people are increasingly going to get their historical “Content” from in the future? Will they even realise the swap that has occurred, or, will the assumption be that we historians are present there by implication, part of the training data (whether we consented or not), the background noise.
I’ll wrap up now by thinking about my final question, a bit of a double-header.
Question 3: Do we have an obligation to “teach against” bullshit in our classrooms? Can one write history entirely indifferent to truth?
By this question I don’t mean a domineering correction of student belief or opinion, not at all. What I do mean is are we obliged to spend far more time now, teaching far more people, how to conduct robust, honest, and thorough source criticism as if they were historians? Because if I could insist on one skill that I have been trained, and practiced, to have, the skill that I think historians really specialize in, as a core requirement of humanities education today, it would be this; Ad Fontes, back (again) to the sources.
In a recent interview with PoliticsJOE, the philosopher of gender Judith Butler was asked whether they thought that there was a “broader academic trend to seek to reduce and simplify seemingly complex ideas such as gender, or in economics, or politics; a flattening, a rejection of intellectual complexity”. Butler replied that actually we see the precise opposite in academia, i.e. a real embrace of complexity, and that gender studies is often found at the forefront of those efforts, but, that in the wider world, or in public, yes we emphatically do see a widespread rejection of complicated truths, in Butler’s words an explicit “anti-academic animus”, working to position public views of intellectual labour as “all these people doing the same thing, and they have a single doctrine and they make you line up, and pledge allegiance to their doctrine and you can’t take one of those classes unless you come out believing the same thing as everybody else”. “We could make a comedy out of it, except it’s deathly serious”.
How far away are we from artificial, manufactured, perhaps even automated pasts deployed as weaponised information to encourage mythological interpretations? To marshal populations towards terrible ends? I think the answer is quite obvious: they’re already here. We know that portrayal of historians as all doing and believing the same thing, signing the history doctrine pledge card, is bunk. As Butler says though, it’s a deadly serious view of us quite alive in the world, and one strongly encouraged. I think we also know we can’t simply spend the next decade tending to our patch as normal, running our REF rat race, having our debates as we always do. British history in the next ten years must face up to this challenge and confront it directly. We suddenly face a moment where widely available historical bullshit narratives can be provided at scale, not just in a way that enables some students to read or think less than presently, but as a tool of information warfare. I need not remind this audience what is being done right now to the history and humanities infrastructure of the United States.[6] We face history again being rewritten by the winners, for the winners, and we British historians must do our part to fight back, to keep the collective stories of the past as multivalent, as responsible to the source material, as pluralistic, rigorous, and honest, as we can. Thank you.
[1] https://www.historians.org/perspectives-article/what-does-it-mean-to-think-historically-january-2007/
They are listed as: “change over time, causality, context, complexity, and contingency”. Recently an AHA piece proposed adding “continuity”. Arguably should have been there the whole time.
[2] Sarah Fisher, ‘Large language models and their big bullshit potential’, Ethics and Information Technology, 26:67 (2024), n.p. (online only).
[3] Maxine Berg and Pat Hudon, Slavery, Capitalism, and the Industrial Revolution (Polity Press, 2023).
[4] David Stefan Doddington, Old Age and American Slavery (Cambridge: Cambridge University Press, 2024). SOME manumission rates for bits of Virginia can be found: Art Budros, “Social Shocks and Slave Social Mobility: Manumission in Brunswick County, Virginia, 1782–1862”, American Journal of Sociology, Vol. 110, №3 (November 2004), pp. 544. Basically between 1782 and 1810 the vast majority of enslaved people in this county were manumitted, with virtually no one manumitted after that barring a spike (22) in 1820–9.
[5] Laster, Deny and Disavow, 30.
[6] NEH, “garden of American Heroes”, and so on.
