The Ghostwriter in the Machine

Machine learning has been around for decades as a technical research field, but public interest was galvanized by the release, in November 2022, of ChatGPT, which provided an easy interface to a series of powerful large language models. From this has followed a seismic shift in public attention, with countless think pieces and long reads probing the ethics, costs, implications, perils (up to and including human extinction), and promise of these technologies, to say nothing of the high-profile lawsuits, congressional hearings, labor disputes and settlements, and an absolute torrent of venture capital investments. Artificial intelligence is in your email and your search engine, and it’s also in your word processor and favorite photo-editing widget. Some argue we are living through another Gutenberg revolution.

This is the milieu into which Dennis Yi Tenen’s book Literary Theory for Robots: How Computers Learned to Write (part of the new Norton Shorts series marketed at general readers) arrives. Tenen, a professor of comparative literature at Columbia University and a former software engineer at Microsoft, is ideally situated to escort us back through the centuries to foreshadow and perhaps even find some genuine forebears to the computational technologies that so compel us today. By looking at where we’ve been, he argues, we can gain some insight into where we might be headed.

The opening chapters of Literary Theory for Robots walk us through a range of historical figures and their associated writings, thought experiments, and sometimes actual devices. Some, like the philosopher Gottfried Leibniz or the mathematicians and inventors Charles Babbage and Ada Lovelace, are familiar as recurring characters in popular histories of computing; others, like the medieval Tunisian scholar Ibn Khaldun, the German poet Quirinus Kuhlmann, and the 17th-century Jesuit polymath Athanasius Kircher, are likely less so.

Tenen develops two converging themes. The first is that the quest for a universal symbolic language was also accompanied by a quest for a universal machine, what ultimately became the modern computer. The second is the idea that “intelligence” is always an artificial construct, which is to say a function of external conditioning and tools as much as it is innate ability or aptitude. Thus we are treated to the example of Kircher’s Mathematical Organ, a contrivance that took the form of a large wooden box filled with rows and columns of removable slates, themselves inscribed with strings of letters and numbers. By referring to set formulas and instructions, the letters and numbers could be assembled into coherent sentences forming an answer to a query.

In other words: a chatbot, avant la lettre. Or was it? Certainly aspects of the interactions Tenen describes resemble the way we might play with a tool like ChatGPT. But the underlying operating principles of Kircher’s wooden box and the computer science behind large language models have nothing in common in any material sense. It is not obvious how our newfound knowledge of this device might help us better grasp the environmental harms of contemporary AI, or its propensity for bias and racial slurs, or the illicit use of copyrighted material in training data.

The full shape of Tenen’s argument only emerges by the midpoint of the book, where we are introduced to what he calls “template culture.” This is his name for the late 19th century’s proliferation of style guides, how-tos, and manuals for fiction writing that appeared in response to the public’s growing demand for leisure reading material. (The phenomenon was essentially the Victorian equivalent of “Novel Writing for Dummies.”) Fascinatingly, some of these contained actual formulas and tables whereby an author could outfit themselves with characters (their names, backgrounds, appearance, motives), a setting (much the same), and a plot (revenge, inheritance, thievery, romance, and combinations thereof). This trend reached its apogee (or nadir) in a 1931 publication by one Wycliffe A. Hill titled The Plot Genie, which featured a spinnable cardboard wheel labeled the “Plot Robot.” The analogy to the kind of AI-generated genre fiction one now finds littering platforms like Amazon Kindle is tantalizing.

From there, Tenen sidesteps to the rise of what came to be known as structuralism in academic circles. Just as the template enthusiasts were concerned to discover universal methods for putting stories together, so structuralists sought to reveal universal ways of taking them apart. The smoking gun for this connection turns out to be some conjectural but well-considered evidence that Vladimir Propp, a Russian folklorist, read a book on the archetypal forms of dramatic situations (intended for new playwrights) and based his influential typology for folk tales on that same work. Tenen then introduces us to Noam Chomsky’s theory of universal grammar, which became the basis for the work of Victor Yngve, who was also a linguist at the Massachusetts Institute of Technology, on an early computer program for generating sentences and, eventually (by the 1970s), for even more sophisticated kinds of programs for generating not only sentences but whole stories. These programs had names like TALE-SPIN and QUALM and MALAPROP. Boeing would eventually use TALE-SPIN to generate fictitious “incident reports,” forming the basis of instructional scenarios for pilots.

All of these colorful precursors to what we nowadays think of as artificial intelligence — the mathematical organs and plot genies and even a program like TALE-SPIN — ultimately belong, as Tenen himself acknowledges, to what he terms “an entirely separate branch of statistically minded language scholarship.” Only with the work of the Russian mathematician Andrey Andreyevich Markov, who developed the statistical “chaining” method that bears his name by using the poetry of his countryman Alexander Pushkin as his prime example, do we arrive at a clear conceptual ancestor to today’s large language models. Markov’s insight, presented as early as 1906, was that by looking at pairs (“chains”) of words as they have been used in the past, we can make predictions about which words should follow next in an open-ended sequence. ChatGPT uses vastly more complex mathematical relationships to build its language model, but a basic layperson’s explanation of what it does sounds remarkably like what Markov first proposed. Crucially for Markov, as for today’s AI developers, the concern was not with the truthfulness or accuracy of any sentences so produced but with whether they were grammatically correct and statistically plausible. Making sense was more important than getting things right.

From there, as Tenen breezily puts it, “the rest is history”: “You can Google it,” he suggests, “or ask your friendly neighborhood chatbot for assistance.” (For the record, asking ChatGPT about history is not really a good idea, for the reasons just stated.) Literary Theory for Robots, I should note, is written in that same light, jaunty style throughout. There are exclamation points, quips, and asides, and Tenen sometimes seems on the verge of apologizing to the reader for boring them. Two of his historical figures are introduced with a “so-and-so walk into a bar” joke; a third is referred to as “another one of our lovely weirdos.” Early on, the reader is advised: “Insert something clever here about forgetting the past to the detriment of the future.” One suspects the heavy hand of an editor, or indeed a style template reflecting a marketing strategy. Only at the very end, when Tenen situates his scholarship amid the war in Ukraine (where bots play a lethal role on the battlefield), does the writing momentarily take on more gravitas. While I appreciate the need for academic specialists to communicate with a wider public, I am not convinced the way to do it is with feigned jocularity cloaking our credentials and expertise.

But the main question is what to do with the historical knowledge we have gained from this book. Causality, it turns out, is hard for even the best historians to prove. And while that’s not the only motive for doing (or reading) history, it does call into question the project of using history to justify our outlook on the future.

These are especially urgent questions for humanists who may feel increasingly embattled and left behind by the pace of social and technological change. We humanists will know that it was Fredric Jameson who famously enjoined us (in 1982) to always historicize. But we also know our Ecclesiastes: “The thing that hath been, it is that which shall be; and that which is done is that which shall be done: and there is no new thing under the sun.” Everyone has indeed heard a variation of that “clever” phrase about those who forget the past being doomed to repeat it, but the humanist and historian knows that they are George Santayana’s words, and that what he actually wrote (in 1905) was “those who cannot remember the past are condemned to repeat it.” At a moment when we in the humanities may feel already a part of history ourselves, our vocation as professional rememberers of things past seems like one of the strongest cards still left to play.

In fact it’s a double play. On the one hand, we humanists can claim we have something to offer: By looking to the past, we can see a bit of the future and perhaps even help shape it for the better. On the other hand, historicizing the present allows us to maintain our distance: The humanist is not easily impressed, not susceptible to being taken in by trends and fads. We have seen it all before, and we see better than anyone (or so we like to think) the pitfalls that await the unwary.

Tenen seeks to steer both courses throughout. Literary Theory for Robots is carefully researched, despite its consistently self-effacing style, and (as I hope I have conveyed) filled with fascinating episodes. But with history comes responsibility. “AI was created specifically to make us smarter,” we read at one point. That’s one version of the story. Another — grounded in the same commercial impulses that produced Victorian template culture — is that today’s AI was created to make a small number of large corporations very large sums of money.

I myself don’t believe all of history is reducible to capital. But I do believe history isn’t always about exclusively people-centered interests, as is confirmed by the climate crisis. More than three decades ago the philosopher and theorist Manuel DeLanda, in a prescient book about intelligent machines and warfare, proposed the thought experiment of a robot historian of the future. What would such a historian pay attention to, he wondered? Surely not the clever stratagems of the great captains and generals that make up so much conventional military history. Rather, the robot historian would fixate on the ruptures and evolutionary leaps in what DeLanda termed the “machinic phylum,” the systems and mechanisms he charted from ancient siege weapons to the nuclear ICBM (and anticipations of drones and Tenen’s killer bots). What then of tweaks to some new algorithm by an anonymous engineer in a cubicle somewhere? Or a new chip architecture? Or the contract workers manually “cleaning” data sets for sub-subsistence wages? If “the rest is history,” then where does all of that history belong?

We may think we know how robots learned to write; at the very least, as Tenen shows, it makes for a great story. But is it the history that will matter in the end? And who — or what — will get to write it?