Moor Studio/DigitalVision Vectors/Getty Images
In conversations around artificial intelligence, I keep hearing a new phrase: Writing is a proxy for thinking. This phrase is often used to suggest that assessing and evaluating writing isn’t actually something we should worry about in this age of generative AI; rather, we should just worry about assessing thinking. (Nafisa Baba-Ahmed’s recent letter to The Guardian is just one example, though this idea has been everywhere in my social media feeds as well, offered with more or less nuance.) It is often countered by someone arguing that writing is thinking, and so we should not give up assigning writing (a claim predating the generative AI concerns).
I am more on the “writing is thinking” side of the debate, but offer here that “writing is a proxy” and “writing is thinking” may both be right and both lead us to the wrong conclusions, in that they each focus only on one aspect of writing. “Writing is a proxy” mistakes writing-as-object for the totality of writing. “Writing is thinking” tends to mistake writing-as-process for the totality. And both have a tendency to misrepresent the fundamental cognition involved, with potentially disastrous effects for understanding and mitigating the negative impacts that generative AI may have on both writing and thinking.
On the one hand, of course writing is a proxy: We never have direct access to someone’s thinking. We only ever have access to proxies: actions, speech, writing, art, brain scans, whatever. It’s never access to the neurons firing. Humans don’t even have full or easy access to their own cognition. That’s a large part of why people end up in therapy: We have to externalize much of our thinking and turn it into concrete words or images (a proxy) before we can understand it and manipulate it ourselves.
As a result, we can never actually evaluate a student’s critical thinking directly. We can only ever evaluate some proxy for it. While some kinds of thinking can be evaluated through nonlinguistic proxies, others cannot. But there is no alternative to evaluating proxies, and imagining otherwise is folly.
“Writing,” as I am using it here, includes what some faculty and teachers of writing call the prewriting process (and which some may not even consider writing at all). Many proponents of generative AI actually emphasize that AI can be used for these steps, rather than “writing” itself. But prewriting is not in any way cognitively distinct from other aspects of writing. Prewriting is written scaffolding in the form of recorded brainstorming, idea generation, mind-mapping, outlining and so on. It exists before formality, not before writing as a technology.
This false distinction leads us to the deeper issue with both “writing is a proxy” and “writing is thinking.” One of the common problems I see in arguments about cognitive offloading and “writing is a proxy” is that those who support the use of generative AI argue that we have always cognitively offloaded parts of our thinking. Jon Ippolito’s article in Electronic Book Review is one such example; as he writes, “It remains to be seen whether large language models will disrupt society more than earlier linguistic innovations, but they are certainly not the first to offload human cognition.” Ippolito identifies the printing press, the word processor and other technologies that offload cognition. Writing is thinking, these technologies offer different ways of writing and we’ve been completely fine with them, or so this line of argument goes. This is somewhat true, but also misleading.
Language itself is something built into our brains. It is not separable from thinking in the human mind. When I teach on this topic, I often assign students to listen to Shankar Vedantam’s 2018 interview with Lera Boroditsky, a cognitive science professor at the University of California, San Diego, who shows how the words and structures of different languages often provide different structures for thought. There are many different languages, and thus many different ways of thinking in language, but language is probably the external thing we can most closely link with thinking itself.
Writing, by contrast, is not the thinking or language itself, but a whole set of technologies we use to manage the cognitive load involved with language processing. Andy Clark’s 1998 article “Magic Words” highlights how language and writing are essential to a whole host of cognitive processes related to memory, environmental simplification, coordination, deliberation, attention, resource allocation, data manipulation, filtering concepts, etc. For many of us, writing is so crucial to the management of language and therefore thinking that we cannot think certain things without these tools. This is how the “writing is thinking” group ends up just as misleadingly right as the “writing is a proxy for thinking” group.
Individual writing technologies were refined over centuries and millennia to streamline this cognitive management. As just one example, punctuation and spaces were created to reduce the cognitive load of processing written language. Before, readers mostly had to read out loud and spend precious cognitive energy deciphering where one word or sentence ended and another began. (Compare the difficulty of reading the last few sentences with reading the next few, without punctuation or spaces:
Individualwritingtechnologieswererefinedovercenturiesandmillenniatostreamlinethiscognitivemanagementasjustoneexamplepunctuationandspaceswerecreatedtoreducethecognitiveloadofprocessingwrittenlanguagebeforereadersmostlyhadtoreadoutloudandspendpreciouscognitiveenergydecipheringwhereonewordorsentenceendedandanotherbegan
It’s not impossible, or even terribly difficult, to process this text, but scriptura continua like this does slow the vast majority of us down substantially because the spaces and punctuation are doing a lot of the work for us in terms of processing the language.
In other words, writing is not thinking we offload to other technologies like the printing press. Instead, writing is a set of technologies that we have historically used to manage our language and its cognitive load—it is always already offloaded. Writing is not the only way to manage the cognitive load of language, of course; oral cultures developed all sorts of linguistic and nonverbal tools, like alliteration, sentence patterns, vocal inflections and so on, that managed language and thought in different ways, as Walter Ong argued in Orality and Literacy. Writing, however, has added benefits, such as its stability in time, which allows it to extend the interplay of language and memory in a way that oral language has a harder time with (and this is, of course, one of the things that Socrates was worried about).
But thinking about writing as a technology that is always already a form of offloading gives us a better way to identify what, exactly, is being offloaded by generative AI versus other writing technologies: language itself. Ippolito says that “the most obvious cognitive burden offloaded by large language models is the mechanics of language”—but there is no human language without its mechanics. (We sometimes assign arbitrary values to some mechanics over others, which can increase the cognitive burden of managing them, but that is a separate issue.) And AI goes beyond offloading mechanics to offloading language generation. That is, after all, why a large language model is called “generative” AI.
For example: Earlier this year, I was preparing a talk for a conference. I had written an outline with a great deal of detail, but it was still in bullet-point form, and I needed a very tight script in order to make sure I was within my very short allotted time. And I was preparing to teach my students about AI prompting—so what better time to provide an illustration that I could critically examine with them?
I fed ChatGPT 5.2 (via my ChatGPT Edu account, provided to all students and faculty by my university) a few examples of scripts for talks I had given previously and prompted it to take my bullet-point outline and turn it into a script. I regretted it immediately upon seeing the generated output: It did not sound like me, but suddenly I had great difficulty seeing what I would have said, because text that was close enough was right there in front of me. Editing for my own meaning actually became harder, not easier. Having to work with generated text, even if it had been generated out of some of my own words, had fundamentally changed the thinking involved. I could not help but be reminded of my experience as a graduate student dealing with an overbearing editor of a journal article, who mangled my ideas through the editing process. I became distant from my own ideas, no longer fully owning them—and it damaged my own confidence and thinking for a good while after. I never finished that revise-and-resubmit, and the piece was never published.
When we use an LLM for writing, we are not simply proceeding down the path we created with other writing technologies. None of the other technologies of writing offloaded language generation so completely; whether oral patterns, the book, the printing press, the word processor—they all required the human generation of the majority of the language.
The technologies Ippolito identified made it easier for writing to do the very things it was intended to do as a thinking tool, but they did not, so far as I can tell, fundamentally change the kind of language processing in the brain or what was being offloaded. Generative AI, by contrast, is numbing and amputating the essential offloading circuits we created when we invented writing. It allows us, sometimes requires us, to offload the very tools we use to manage our own thinking. We put ourselves at an additional, very substantial remove from the process of thinking in a way that none of the other tools did, which offers a fundamentally different danger.
(And this is before we consider the dangers of giving a substantial portion of our language processing over to tools owned by corporations and billionaires with a history of enshittification, rather than tools we can own and can control ourselves.)
This danger is closely tied to what that study from the Massachusetts Institute of Technology last year warned about: cognitive debt. Other studies on cognitive deskilling amplify these concerns. The more distance we put between us and a cognitive process via tools, the more the process itself has the potential to degrade simply by lack of use; our brains’ plasticity means that they will change how they allocate resources if some circuits aren’t used as often (Nicholas Carr’s The Shallows offers an accessible introduction to these concepts if one is not familiar).
Such change isn’t a problem with every form of cognition and offloading, though we should be careful as we introduce new tools. But this change does happen, and consequently when we make choices in the classroom about what we assign, what we assess and evaluate, and how we allow or prevent AI use, we must not only consider what skills we are preserving, but also those we are simultaneously helping to numb and amputate.
For my own classes where I teach writing and critical thinking, I still rely on the original rule I developed with students in 2023: AI should not be used for anything that another person could not ethically do for you. We’re mostly conscious of how and where offloading to another human has its limits, and this is the best proxy I have found so far for helping us establish limits without falling into a policing game that we all lose. But as I have taught my students the best practices I know for AI use, I have grown leery, wondering if even this is enough, if we are doing them some greater harm than we or they realize. I suspect the less we practice managing thought with writing directly, the more we outsource such language management to generative AI, the more difficult we will find managing our own thinking in other contexts, in the places where AI cannot—should not—help us.
