Don’t Expect AI to Revolutionize the Humanities

Guest Post from Eric Dane Walker

a robot in a library

Hi all,

I’m pleased to publish this guest post from Eric Dane Walker, who I discovered from his thoughtful comments on Colligo. Colligo, remember, means “bringing together,” or “gathering together,” and I’m always looking for insights, feedback, and contributions from you: the readers. Eric explains and critiques a movement called the “digital humanities,” so if you’re into fiction, history, Shakespeare, or just books and all the rest of the liberal arts, I think this is an important post. AI is affecting this aspect of our culture as well. It’s time we took notice.

Erik J. Larson

------------------------------------------------------------------------------------------------------

I’d like to thank Erik for affording me the opportunity to think out loud in front of an inquisitive and intelligent audience. His recent post, “Don’t Expect AI to Revolutionize Science,” hit upon something of resounding import. (Do read it if you haven’t yet!) As I understand it, Erik’s essay testifies to a farther-reaching claim than his intellectual modesty would let him directly articulate. My objective in this post is to bring that claim to the surface for consideration and to provide some evidence for it.  

The claim is this: the robustness of any inquiry, computationally supplemented or otherwise, depends on the robustness of human judgment.

Made explicit, the claim casts Erik’s post in a new light. Using pharmaceutical biochemistry as a representative illustration, Erik suggests that human judgment is the sine qua non of the natural sciences, computationally supplemented or otherwise. In turn, I’d like to suggest that human judgment plays an analogous role in the humanities, and I’ll use the digital humanities to illustrate. The irreplaceable authority of human judgment in inquiry will become evident. The modern conceit of data science and AI is that human judgement is on its way out. Far from it, it’s central to all progress whatsoever—whether in the sciences, or the humanities.

Defining the Digital Humanities

In “Fictionality,” a 2016 paper published in the Journal of Cultural Analytics, Andrew Piper explores whether the fictionality (the quality of being fictional) of a piece of writing is a matter of elusive, context-sensitive things — such as how a writer intends her words or how her audience takes them — or whether it’s a matter of the words themselves. Rather than picking from a few paradigmatic texts, Piper employed a computer to look at about 28,000 of them, calling upon a context-blind computerized pattern-recognition system to see if it could tell, on the basis of diction and syntax alone, whether a given work was fiction or nonfiction.

Piper and Richard Jean So also co-wrote a more popular piece for The Atlantic that same year about their joint project of measuring the impact that Masters of Fine Arts (MFA) programs make in the literary world. They describe the first part of their project this way:

We began by looking at writers’ diction: whether the words used by MFA writers are noticeably different from those of their non-MFA counterparts. Using a process known as machine learning, we first taught a computer to recognize the words that are unique to each of our groups and then asked it to guess whether a novel (that it hasn’t seen before) was written by someone with an MFA.

Unsurprisingly, given an increasingly dominant AI and data science culture, investigations like these rose in prominence. They promoted a new, supposedly revolutionary approach to humanistic inquiry, an approach called “digital humanities.” The idea is to harness computational power for the purpose of humanistic analysis, criticism, and interpretation. The authors of one of its founding documents, the book Digital Humanities (MIT Press, 2012), write that the new approach comprises the activities of “digitization, classification, description and metadata, organization, and navigation” — all enabled, or perhaps even achieved, by brute computation.  Dataism and AI had penetrated the humanities.

Now wielding algorithms, the digital humanities promised a rigor supposedly absent before. Like pies cooling on windowsills, giving off a whiff of science-like exactitude, the digital humanities lured donors, administrators, and grant makers. Even amid higher-ed's “Great Adjunctification,” tenure-track job openings and post-doc positions for scholars who work in the digital humanities began materializing with resources that had been recently declared completed.

The revived interest was not unwelcome. The digital humanities, with their scientific sheen, were attracting money and clicks, and thereby bore our age’s reigning marks of worth. Humanistic inquiry appeared refreshed. But was it revolutionized? Could it be, simply by incorporating massive computational power?

The Proper Ambit of Computation

Let’s take a closer look at Piper’s project. He conscripted a pattern-recognition algorithm — because it’s unreceptive to context, by the way — to categorize different pieces of writing to see whether the system could identify pieces of fiction based solely on the words they contained.

But how do you determine whether the pattern-recognition algorithm is correct in its identifications in the first place? A moment’s reflection will inform you that it’s only by comparison to ordinary human judgments about which pieces of writing are fiction. Such judgment provides the criterion for the algorithm’s success. Without human judgment, then, the study loses its raison d’être.

We can see human judgment playing a similarly indispensable role in Piper and So’s joint project. For their study, they had to teach a computer to recognize words unique to MFA-grads and words unique to non-MFA-grads.

Prior to the computational wizardry, humans  had to sort words according to their perceived uniqueness. In othe words, the algorithm’s input were not unprocessed "givens" but human judgments suitably encoded for algorithmic processing. Not data but capta — things taken to be as they are: in this case, words taken to be unique to each group. The picture that emerges from such computation, therefore, isn’t a representation of the differences between two kinds of novel. Rather, it’s a representation of human judgments about the differences between two kinds of novel. Whether the former differences are captured in the resulting picture depends upon whether the human judgments are sound. Is AI revolutionizing the humanities, or disguising ubiquitous and indispensable human judgement?

In fact the feature of computation that promises to strengthen humanistic inquiry — its massive, cyclopean processing power — can also become its weakness. A computational result is no more and no less authoritative than the human judgments supplying the input. Any narrowness, incautiousness, or bias informing those judgments will be unthinkingly inherited by the result.

Whither the Digital Humanities Revolution?  

While acknowledging this constraint, digital humanities scholars should recognize the opportunity to explore and debate the scope and limits of algorithmic  analyses and to suggest new directions for research. A space opens up for the exercise of those good old-fashioned interpretive skills that have always been forged in humanistic inquiry. Reasoning, interpreting, critiquing, explaining —these things don’t stop when scholars apply computational methods in their work. Indeed, they must precede and follow such methods. 

Importantly, the The digital humanities’ “revolution” isn’t data science replacing humans. It’s  reinforcing the need for, and the authority of, sound human judgment. 

The fact that inquiry — scientific or humanistic, computer-aided or not — is nourished by human judgment doesn't render human inquiry feebly subjective. It shows, to the contrary, that inquiry yields objective insight only because, and only to the extent that, human judgment is finely honed and publicly communicable. At our peril, then, do we let our power of judgment atrophy.

==========================================================================

Thanks, again, Erik, for inviting me to do this. 

Eric Dane Walker is a PhD candidate in Philosophy at University of California, Riverside. He's writing an historically grounded dissertation about what diagrammatic writing systems do, how they do it, and why they're important for mathematical inquiry. You may reach him at thetricyclefallacy@gmail.com.