It’s been 20 years since I wrote a blog post about how the keyword revolution might impact scholarly research, for better and worse. And some of the “worse” possibilities still ring true in our current “chatbot” revolution.
I wince a bit now to see how eagerly I embraced Google Print in 2005, though it’s still indisputably true that keyword searching has made possible new historical discoveries. Keyword searches can be especially helpful at finding needles in haystacks, assuming the haystack is digitized and open … and you know the name of the needle.
But what I didn’t trust as much then, and still don’t trust now, is the impulse to use computational analysis of large corpora of digital texts to comprehend the haystack, which is what LLMs and their boosters claim that chatbots can do.
Or perhaps I should say, I wouldn’t trust someone who doesn’t understand how the corpus is built or how the computational stuff works to try such analysis; I’m more open to learning from and listening to digital humanists who have a better idea of what they are doing with big data.
Here’s the thing about chatbots and LLMs though: not even the people who made them really understand how they are working.
In 2011, I wrote another blog post cautioning about the use of proprietary databases to draw quantitative conclusions about keyword search results, partly on the grounds that the companies who built these databases are rarely transparent about how things work under the hood.
Now we are in a moment when, if anything, the tool du jour is even less transparent about its training data and its functionality. Only when something really weird and disturbing starts to happen do most users start to think about what the ghost in the machine is doing. But even when it’s operating “normally,” what does that mean? How would we know?
On those grounds alone, beware.
And not only on those grounds. I wish there were more people connecting the attacks on humanities research funding with the boosterism surrounding chatbots, given that the same people are often pushing both movements. I don’t think I’m being overly conspiratorial to see the outlines of a playbook: discredit human historians as biased revisionists who just spout ideology anyway, while crediting obsequious robots who generate BS on demand as the inevitable future that we all have to accept.
If you can get enough people to accept the output of the latter as passable, why go back to funding the former?
Likewise, if you can brute-force history professionals into only writing narratives “that promote a deeper understanding of our nation’s extraordinary heritage, including our record of advancing liberty, prosperity, and human flourishing” (as the NEH Public Scholars call now prefers) or that never “disparage Americans” (as the National Park Service reportedly now requires), then maybe you can convince the general public that all it takes, all it ever would have taken, to produce more pleasing narratives about the past is for historians to choose to write them.
In other words, we should be seeing such directives to historians as versions of what “A.I.” boosters claim knowledge workers will all be doing in the future. These are policies built by people who believe that historical work is “prompt engineering” all the way down, not by historians who understand that the past can break your heart, that evidence can surprise and frustrate and confound your fondest hopes for what you will find in the archive.
At the end of the day, above all, we also have to see the boosterism surrounding chatbots as a sales pitch. The boosters want to convince paying customers to spend money on their product, even if they still can’t quite articulate what the value proposition of the new tool is.
I had to laugh, for example, that the Google NotebookLM engineer quoted by the Times suggested his product could be valuable by enriching historical publications with accompanying multimedia sites that would contain the author’s primary sources: “What if e-books of history came enhanced with a NotebookLM-like interface?”
News flash: we could have been doing that all along with the basic tools of the web, and thoughtful digital humanists have been producing such e-books for decades. We don’t need LLM tools to publish our sources online, though I begin to wonder whether we should be doing this, if this is what LLM scraping bots are going to do with what we make available.
Similarly, Alan Jacobs has pointed out that at least some of what people claim we need chatbots for … was already accomplished by the aforementioned keyword revolution and that venerable corkscrew known as “Ctrl-F.” Unfortunately, the “A.I.” revolution is probably going to break that tool, too, in the end, rather than improve it.
Indeed, for all the problems with digitized print databases that I was thinking about in 2005 or 2011, I’m afraid that we are going to miss them when they are gone.
