Pranav Anand is a Professor of Linguistics at UC Santa Cruz, specializing in semantics and pragmatics, particularly in the study of context-dependence, perspectival expressions, and subjectivity. His work, which leverages linguistic fieldwork, logical analysis, philosophy of language, and computational linguistics, has examined affect and sentiment, debate and persuasion, narrative, ellipsis and fragments, and modality and knowledge. On March 11, 2024, Professor Pranav gave the March Slugs and Steins lecture on ChatGPT: A Selective History, and Notes on the Future of AI/ML Language Models.
The Ends of Writing
As we contemplate perhaps being on the cusp of a fundamental change in our relation to writing, it’s important to think critically about its practice, what we might be losing, and what we may gain.
In the aftermath of ChatGPT’s introduction in late November 2022, a raft of articles arose in the popular press auguring the end of writing, e.g., in The Atlantic’s The College Essay Is Dead or Forbes’ more full-throated AI is the end of writing. One and a half years in, students continue to complete essay assignments and people do seem to be pursuing writing-based careers, though there is a growing stream of reports on how people have turned to large language models (LLMs) like ChatGPT to replace the work people have historically done. As we contemplate perhaps being on the cusp of a fundamental change in our relation to writing, it’s important to think critically about its practice, what we might be losing, and what we may gain.
Writing is our earliest technology invented for the recording and transmission of language. So useful is writing that it developed in four distinct times and locations — three thousand years ago in Mesoamerica, five thousand years ago in China, and six thousand years ago in both Mesopotamia and Egypt. For some temporal perspective on other technologies, this is a thousand years after the invention of plumbing and the wheel, and half as long as humanity has had agriculture. Widespread literacy is far more recent. Writing and reading were the province of bureaucrats, religious institutions, academics, and the financially elite until the 1500s in Europe, and literacy remained relatively circumscribed until the 1800s, when movements toward public schooling led to an astonishing democratization: whereas around 10% of people are estimated to have been literate by the late 1400s, global literacy rates are now approaching 90%, though large gaps persist by country, age, and gender (you can explore data compiled by UNESCO and the World Bank here).
At the same time, like all technologies, reading and writing have inherent structural barriers created by their affordances (the capacities that enable the technology to function). Unlike spoken language, which can be acquired by the vast majority of people within the first few years of life and without any explicit training, written language seems less universal: Whole word/whole language methods, which emphasize readers deciphering text on their own alongside images (like picture books), have been estimated to work for at most 40% of English-reading children. Most need something like phonics training, explicit instruction in the mapping between symbols and sounds. And writing is a more complex endeavor, requiring a person to juggle in memory their intended point, the rhetorical patterns of their genre, and the subjective state of a simulated reader. It is thus unsurprising that effective writing is even rarer than effective reading. It’s also why writing instruction invokes so much sequentialization and offloading of cognitive labor, recruiting mind maps, graphic organizers, outlines, and peer critique.
These technologies thus enable us to peer into the link between linguistic expression and learning with paradigms hitherto unrealistic, a reason for excitement, not dread.
Beyond its ability to communicate across time and space and preserve knowledge, early scholars of the history of the book, like Jack Goody and Walter Ong, emphasized the cognitive advantages that written language was argued to bring, including the capacity to reflect deeply on a complex argument via visualization of a text and rereading. Writing instruction continues to favor argumentative essays, which are believed to foster rhetorical command, deep analytical thinking, and knowledge consolidation. Chat-style LLMs like ChatGPT, which can produce an essay directly from a user’s prompt, threaten this form of learning through writing. And yet, experimental investigation of writing’s downstream effect on learning has suggested that while students do, on average, learn from writing, the results are quite variable and the gain from writing is modest at best (see recent meta-analyses by Bangert-Drowns et al. 2004; Graham et al. 2020; van Dijk et al. 2022). Most relevantly, students seem to learn best not necessarily from argumentative writing, but writing that encourages them to recruit their meta-cognition, their ability to reflect on what they know and what they don’t.
From that standpoint, it’s not hard to imagine the potential benefits of an LLM-aided approach to writing — let’s call it co-writing — one where the LLM helps to organize rhetorical structure, framing, and meta-cognitive thinking, and where the writer is more curator than crafter, selecting from choices the LLM offers up and editing the result. While it may seem that such an approach is destined to lead to less learning than writing from scratch, existing evidence reminds us that the act of writing is not alone sufficient to trigger appreciable learning. Indeed, for many people, being able to offload much of the cognitive juggling act of writing could enable them to more fully attend to one component, thereby increasing both learning and willingness to write. These technologies thus enable us to peer into the link between linguistic expression and learning with paradigms hitherto unrealistic, a reason for excitement, not dread, and may spread writing further than our existing methods.
Selfie of a female macaque, Naruto, in North Sulawesi, Indonesia, who triggered photographer David Slater’s camera.
At the same time, the products of co-writing as curation are not the result of the writer’s individual effort, which I acknowledge does not feel like writing to us. At the core of this issue is the idea of writing as property, an idea so central to our contemporary world of copyright and plagiarism as to be self-evident, but which is the result of the lengthy historical development of the cultural concept of authorship. For most of written history, authors were artisans, supplying a manuscript to a patron, who gained complete control over that object, including the right to take credit for the content and to declare who could copy it (i.e., copyright). Our modern conception, where an author claims both intellectual credit and legal ownership, developed over four centuries, starting actually from from censors’ need to hold someone responsible for publications that violated content guidelines. Michel Foucault called the development of this modern authorial figure “the privileged moment of individualization in the history of ideas, knowledge, literature, philosophy, and the sciences,” and this individualization reaches its zenith in the U.S. Copyright Office’s understanding that only human beings may serve as authors of copyrightable work. On the basis of this policy, the selfies of Naruto, a Celebes crested macaque, were famously found in 2018 to be uncopyrightable in the U.S.
The crux of the problem here is what counts as sufficient agentive, intentional involvement to constitute the minimal standard for authorship, and this is ultimately a factor of social conception, which can shift.
It remains to be seen how this policy applies not to macaques but to machines, though already we can see some tension. In 2022, The Copyright Office determined that the generative AI-created images of Kristina Kashtanova’s graphic novel Zarya of the Dawn are not copyrightable. The Copyright Office argued that Kashtanova’s text prompting method, which the Office likened to suggestions to an artist, had insufficient control to constitute authorship. Kashtanova’s lawyer, Van Lindberg, disagreed, pointing out that mistakes are copyrightable, as are the aletoric compositions of Jackson Pollack and John Cage. The crux of the problem here is what counts as sufficient agentive, intentional involvement to constitute the minimal standard for authorship, and this is ultimately a factor of social conception, which can shift. The case of photography, introduced in 1839, provides a ready analogy: Early photographers emphasized the new technology’s chemical, automatic, naturalistic capture of real objects, contrasting with the subjectivity of painting, but also suggesting that photographs were not authored. But forty-five years later, the Supreme Court ruled that photography is copyrightable, arguing in the process that the intentional hand of the creator was visible in choices of posing and scene. It’s hard not to see the mainstreaming of point-and-shoot photography as the cause of this liberalization. And it is similarly hard not to see the same path being followed as AI-imagemaking spreads; indeed, Lindberg closes his response to the Copyright Office with this line: “AI-assisted art is going to need to be treated like photography. It is just a matter of time.”
The question is whether LLMs will make AI-assisted writing like photography as well, and that depends again on how social conceptions of textual authorship shift with adoption. It may be that choice among the bevy of LLM options (e.g., ChatGPT over Claude) or which LLM output is selected will be enough to constitute sufficient involvement to grant someone authorship. Already, this year, Japan’s Akutagawa Prize for early career fiction writers was awarded to Rie Kudan for a piece partially co-written with an LLM. If this kind of view of authorship comes to pass more generally, it will come with a change of what it means to author a text. A change, but maybe not a cheapening, if we consider longer arcs of authorship, of learning, and of writing. All technologies have life cycles, and we are likely soon entering the next one for writing.
Banner Image: An excerpt from the comic book Zarya of the Dawn, whose AI-assisted images are not subject to copyright.
The Humanities Institute’s 2024 Technology Series features contributions from a range of faculty and emeriti engaged in humanities scholarship at UC Santa Cruz. The statements, views, and data contained in these pieces belong to the individual contributors and draw on their academic expertise and insight. This series showcases the ways in which scholars from diverse disciplinary perspectives contend with the issues connected with our annual theme. Sign up for our newsletter to receive the latest piece in the series every week!