A Review of Literary Theory for Robots

Come for the entertaining history, but take the conclusions with a grain of salt.

Dennis Yi Tenen’s book Literary Theory for Robots landed in my mailbox a month or so ago (a free copy, thanks to Norton), and I finally found time to pick it up while traveling to Iowa State University to consult with their first-year composition program on AI. The book is not without its faults—more on those below—but overall, I found it to be both informative and entertaining as a read.

Tenen’s primary goal in the book is to historicize automated writing technologies in order to 1) help readers understand where generative AI came from both technologically and conceptually, 2) break down barriers between the arts/humanities and science, and 3) alleviate some of the unwarranted panic surrounding AI in favor of more nuanced understanding and critique.

Before delving into those specific points, I want to share Tenen’s definition of artificial intelligence because I find it useful for understanding the relationships among writing, thinking, and technologies. Early on, Tenen takes issue with the notion that thinking occurs solely inside our brains, arguing instead that it “moves by mental powers, alongside the physical, the instrumental, and the social.” Therefore, pointedly, he argues, “Intelligence demands artifice” (p. 4) There is no separation between our minds and objects of creation, whether the stylus, the pencil, the keyboard, the touchscreen, or the chatbot. It’s all “artificial intelligence” for Tenen. Automated writing is just a subset of this expanded definition of artificial intelligence.

This expanded definition also allows Tenen to trace automated writing technologies back much earlier than the chatbots of the 1960s. For instance, Tenen describes a fascinating medieval-era “Arabic Q&A bot,” as he calls it (p. 22), developed by Ramon Llull, that used a table structure allowing a scholar to input a question, convert it into a series of variables, and retrieve an answer. I found his descriptions of the specific mechanisms by which this table worked hard to follow, perhaps because I was reading on a darkened airplane, or perhaps because they’re so esoteric. Regardless, they depended fundamentally on taxonomies, which are always rhetorical and ideological objects—and it is no coincidence that modern AI likewise depends in part on human-generated taxonomies in training data.

Despite taxonomies’ situated, interested nature, those seeking to automate knowledge production fantasized about a kind of universal language, often mathematical, that could describe the world in all its complexity. Here, Tenen writes, early experiments failed:

Intelligence struggles with universal application . . . because the world cannot be the same everywhere at once. The word itself attains universality only from a great distance, described in broad strokes by physics or theology. Its particulars often differ depending on the local context. (p. 46-47)

To me, this struggle with the universal is precisely the problem with current generative AI models: although they can be responsive to well-tuned prompts, they nevertheless default or drift towards the most common, supposedly universal, ideas and discourses in the world. Tenen’s analysis leads me to be skeptical of any AI model’s claims to universality or generality, at least insofar as it means there may be a tradeoff in specificity and situational relevance. Large-scale AI models may provide waypoints on a rhetorical journey, but we must fill in the local details ourselves with our ideas and experiences.

This was the case in the past as much as now. For example, Tenen details what he calls “Template Culture” in the late 19th and early 20th centuries, during which a cottage industry arose to describe the commonplace features of various archetypal stories so that authors could take them up in their own writing. This fascination with templates arose amid a massive wave of literary production and alongside the industrial revolution and Fordist production. Surprisingly, these tools for literary production had a direct influence on computer scientists developing early chatbots:

Earliest documented examples of AI text generators implemented rudimentary versions of these paper-and-ink systems, using similar techniques such as story grammars, event databases, multipass expanders, random tree walks, randomized-event engines, network traversals, and background-world generators. (p. 78)

Here we can start to see why Tenen is so interested in breaking down silos between the arts/humanities and the sciences: they were never so separate as they are now, certainly not in the world of automated text generation. Their supposed separation has consequences:

Viewing them apart has impoverished both communities: poets, in terms of financial and cultural capital, and programmers, in terms of belonging to a deep intellectual tradition. Worse yet, the division gave both communities a kind of a shallow myopia, where AI seems to entail either the death of us all or a cure for all ills. (p. 121)

I agree with Tenen that the arts/humanities and sciences could certainly stand to enrich one another more, and his history goes a long way toward demonstrating what their mutual influence has made possible in both literary production and automated text generation. However, his final point feels dismissive of a great many commentators (including, if I may, me) who are attempting to be clear-eyed about the potential benefits and harms AI does and might entail. Some have already written about the topics in his conclusion, including collective labor, distributed cognition, human responsibility, and the politics of technology.

And what’s more, I don’t know that Tenen’s ideas do much to alleviate what my friend

Ira Allen

call’s “judicious panic” at AI’s worse-case impacts. In broad strokes, just because automated writing technologies have always been with us, and just because all intelligence entails artifice, doesn’t mean we should not be concerned with the exponential growth of AI technologies. Most of his examples were so much esoterica; now they’re becoming ubiquitous, with potential wide-scale implications not witnessed before.

AI at scale is and will have tangible impacts that Tenen appears to dismiss. For example, he contrasts the notion of “society” with “AI” on the basis of material difference, writing:

In real terms, the word “society” identifies a cluster of related ideas supported by a dense tissue of tangible material connections: work, trade, play, dinner with friends. The word “AI” identifies a cluster of related ideas, but without the underlying material support. (p. 129)

I think his point here is to distinguish human community from automation technologies, which have no community of their own. But by dismissing AI’s underlying material support, he forgets about the palpable, real-world impacts of AI (understood as an interrelated cluster of technologies and humans) on the environment via the mining of rare earth minerals, the energy needed to power servers, and the water needed to cool them. “AI” isn’t just a metaphor—although it is one—but also a set of material existences in the world, ones that are currently doing their fair share of harm.

Likewise, Tenen introduces, but then seems to wave away, concerns that AI is poised to replace many knowledge workers. He writes, “It is true that in the future markets may require far fewer doctors or software engineers. But those that remain will also find their work enriched” (p. 135). Good for them? This passage strikes me as tone-deaf, in that it pays little heed to the lives of knowledge workers who may need to be replaced.

To be fair, I am being a little nit-picky here. In broad strokes, Tenen’s book is a great read. It includes valuable ideas about the relationships among arts, humanities, and sciences, and it usefully historicizes the development of the current wave of AI technologies. It’s also downright funny at times. I especially recommend it as a focal text in college classrooms, such as a first-year seminar or a composition class, and perhaps for interdisciplinary faculty reading groups. But if you come to it looking for well-reasoned, pragmatic takeaways, buyer beware: you may not get them.