There is a new study that came out recently called “Working with AI: Measuring the Occupational Implications of Generative AI.”
This study was carried out by a group of Microsoft researchers and the purpose of the study was to investigate how people are actually using generative AI, and to try to see which professions are getting affected the most.
To do this, the authors of this study examined how people use Microsoft Bing Copilot and classified the activity they saw into different occupations and then examined what percent of the work was getting done by AI (it’s more complex than that, and there are multiple YouTube videos that have been created to explain the study using the Google NotebookLM “deep dive” podcast creator, such as this one: https://www.youtube.com/watch?v=tnx0U1oljq0).
What they found is that generative AI excels in the knowledge work category, that is, in doing things like gathering, processing, writing, and communicating information.
Not surprisingly, occupations that focus heavily on such knowledge work are precisely the occupations that this Microsoft team of researchers found were making the most use of generative Ai.
And at the number two spot on the list, right after translators/interpreters is. . . historians!!
This study is significant because it examined what people are actually doing, rather than trying to guess what might happen someday.
Now, I can’t figure out how they determined who was an “historian.” My guess is that they just labeled the queries of people engaged in some form of historical investigation as the work of an “historian.”
Further, when it comes to historians using generative AI, my guess is that Copilot is probably not their platform of choice (I personally think of Copilot as the Netscape of the AI age).
However, all of that is peripheral to the main point. The main point is that generative AI is best at collecting, processing and communicating information, and. . . that is exactly what historians do.
Initially, I wasn’t going to write anything about this because I have written a series of posts about the bleak future of the History profession in the AI age and didn’t want to write any more “Debbie Downer” posts.
However, in the previous post, I used ChatGPT 5 to do exactly what this study found historians are using Copilot for: to collect, process, and communicate historical information.
And it basically did it perfectly.
Our current History profession is toast.
From this point forward, most historical scholarship will consist of querying “the machine” and working with what it spits out.
While I initially liked the ways in which generative AI can accelerate research (especially translation), and still do (I’m not planning on participating in an AI boycott anytime soon), nonetheless, now that it can accelerate the entire process to completion. . .
Ultimately, as I’ve been saying for like 15 years, I think the History profession has to be about something more than producing texts. That already stopped being effective in the digital age, and now it is totally replaceable in the AI age.
My self-interested question is not “Can AI teach people effectively?” but “Will people prefer how AI teaches them to how I’m able to teach them?”
You mentioned in the comments on your digital humanities post that specialized expertise isn’t essential for teaching the typical undergraduate humanities (or social science) course. I remember the time, effort, and cost of acquiring specialized expertise as a doctoral student. In my 25+ years of undergraduate and graduate teaching since then, I think I’ve had the opportunity to draw on that expertise only once or twice.
Now people can do what I mostly did as a graduate student — peruse and analyze information on arcane topics — much less expensively and much more conveniently by using AI. For example, a year ago I asked Claude Sonnet to synthesize the work of five major China scholars in relation to a specific question, and it instantaneously spit out a highly polished, concise essay that accurately summarized a few thousand pages of material (for anyone who’s curious, the essay is at https://chadraymond.substack.com/p/what-do-china-and-the-three-stooges). So what’s the point of entering a PhD program to learn this information, and why should humans bother teaching it?
“Will people prefer how AI teaches them to how I’m able to teach them?”
For the most part, I think the answer to this question (with “I” meaning “any prof/teacher”) is “yes.” And I think that the reason has a lot to do with what you wrote here on “The Death of Curiosity”: https://chadraymond.substack.com/p/the-death-of-curiosity
When a university education is just something that you have to get and get through, then whatever offers the clearest/easiest/most straightforward/painless path to that goal will be preferred.
I feel like what is left for humans is “instruction” or “discussion” at the highest levels of specialization. So, if there was a small group of people who really wanted to push themselves to get to the core of some issue, then they would still appreciate having a more knowledgeable person to interact with.
However, how many places/situations do we have like that?
This has been one of the arguments for LLMs – it can do all the basic stuff so that people can then focus on “high level thinking” and “creativity.” Ok, but in a university context, that would require that the students want to actually engage in “high level thinking” (whatever that is), and are motivated to do the other stuff with LLMs first.
But for how many decades have profs been saying now that students don’t read whatever is assigned for discussion? What is it about the university experience that is suddenly going to transform?
I think that if this is true it is only true of a narrow and restricted part of the historical profession. It seems particularly applicable to fields like ancient and medieval history where the corpus is basically closed and it has been digitised. But even then, it seems to me that there have been valuable new finds, in the case of Chinese history, at Mawangdui (which have changed our picture of canonical texts) or at Turfan (which have transformed our understanding of social history).
It’s not clear to me that the analysis provided here applies to much modern or contemporary historical research. Putting aside the fact that vast archival collections remain uncatalogued let alone digitised, there are vast troves of sources that have not been archived, but take the form of letters, notes, diaries, manuscripts etc. kept in trunks in basements and in attics, held in private hands, that have not even been archived. The work of modern historians is often to locate these materials, assess their value against the background of already known documents, and to collect, collate, organise, edit, translate, and annotate these materials before they can be digitised.
Even then, it’s not clear that such materials are AI ready. A single police dossier from colonial Saigon, for example, can contain a dozen depositions from different individuals about a particular event or incident. These depositions are frequently incomplete, contradictory, and/or shot through with falsehoods – Natalie Davis’s “Fiction in the Archive.” I am not aware of any AI model operating at present, however well trained, that is capable of making the judgements that trained historians can balancing the documents in the dossier against each other, against other dossiers in the fonds, and against other police department practices. And this is before we deal with more niggling issues such as deciding if a number is an 8 or a 3 when the ink has bled into the page and the writer’s cursive is wild or idiosyncratic. Such a decision can rest on a number of heuristics not available to an AI and it can have knock-on effects to the rest of the analysis and story – a 5 year difference between 1913 and 1918 can have profound effects.
Then there is the issue of analysis and narrative style. The writing of historical analysis and narrative is inseparable from the personality of the historian, however dispassionate they might want to seem. Of course an AI can be instructed to mimic another writer’s style, but mimicry is not the same as having style it/(one)self. And a writer’s style varies considerably over time, even while being recognisably similar. The Jonathan Spence of “Ts’ao-Yin and K’ang-Hsi” and “To Change China” is different from the writer of “Emperor of China” and “The Gate of Heavenly Peace,” who is different again from the historian of “Treason by the Book” and “Return to Dragon Mountain.” There is no AI yet that can do the kind of historical writing that Spence did because AIs do not have imaginations (despite the fact that they hallucinate).
So while I think that the analysis here may apply in some cases and to some parts of the profession, there are too many things that historians do (the good ones anyway) that cannot be simply replaced by a machine, not yet anyway. And it’s not obvious to me that AI can or will develop the necessary skills. Above all, however, the problem is that the “I” in “AI” is not very intelligent at all. Indeed, it is not intelligence of any recognisable kind. Using massive computational power to determine the probabilities of certain strings following one another is simply not the same as propositional reasoning, making inferences from premises, some sometimes suppressed, against a “background” to various conclusions that are logically or materially implicated. AI just does not do this. And yet it is the very core of human sapience.
This is a loosely assembled collection of thoughts on this topic, but I am not very troubled by the thought that AI can or does do what many historians *actually* do.
Thank you for the comments, Haydon!!
I had a conversation on FB where some of these same issues came up, so I wrote a new post that’s more about this topic.
https://leminhkhaiblog.com/more-on-history-is-toast-an-explanation/
As I read your comments, I pictured myself in an imagined (but very imaginable) meeting with a university administrator, where the administrator was explaining that the university was considering implementing some “pedagogical advances” that would address the “personalized” needs of the “diverse” student population in an effort to be more “inclusive.” All of this would be cover language for a new policy to do away with certain History courses taught by humans and to replace them with an AI-powered curriculum that students would access through an app (developed by some certified Ed-Tech company).
I could then imagine something like your comments coming from a colleague who was making a passionate defense of the field that no one could disagree with, and I could see the administrator (trained in a Humanities field himself) nodding his head as he listened.
But there was something about the vibe in the room, a vibe that I had felt so many times before. It was that vibe you feel when you could sense that an impassioned plea was falling on deaf ears, because the deal was done. The decision had already been made. This was not actually a “consultation.” It was an announcement.
And the result? The university would decide to implement the new AI-powered curriculum.
This is why I am very troubled that AI can do much of what historians can do.