While I was working on the manuscript for More Than Words: How to Think About Writing in the Age of AI, I did a significant amount of experimenting with large language models, spending the most time with ChatGPT (and its various successors) and Claude (in its different flavors).
I anticipated that over time this experimenting would reveal some genuinely useful application of this technology to my work as a writer.
In truth, it’s been the opposite, and I think it’s interesting to explore why.
One factor is that I have become more concerned about what I see as a largely uncritical embrace of generative AI in educational contexts. I am not merely talking about egregiously wrongheaded moves like introducing an AI-powered Anne Frank emulator that has only gracious thoughts toward Nazis, but other examples of instructors and institutions assuming that because the technology is something of a wonder, it must have a positive effect on teaching and learning.
This has pushed me closer to a resistance mindset, if for no other reason than to provide a counterbalance to those who see AI as an inevitability without considering what’s on the other side. In truth, however, rather than being a full-on resister I’m more in line with Marc Watkins, who believes that we should be seeing AI as “unavoidable” but not “inevitable.” While I think throwing a bear hug around generative AI is beyond foolish, I also do not dismiss the technology’s potential utility in helping students learn.
(Though, a big open question is what and how we want them to learn these things.)
Another factor has been that the more I worked with the LLMs, the less I trusted them. Part of this was because I was trying to deploy their capabilities to support me on writing in areas where I have significant background knowledge and I found them consistently steering me wrong in subtle yet meaningful ways. This in turn made me fearful of using them in areas where I do not have the necessary knowledge to police their hallucinations.
Mostly, though, just about every time I tried to use them in the interests of giving myself a shortcut to a faster outcome, I realized by taking the shortcut I’d missed some important experience along the way.
As one example, in a section where I argue for the importance of cultivating one’s own taste and sense of aesthetic quality, I intended to use some material from New Yorker staff writer Kyle Chayka’s book Filterworld: How Algorithms Flattened Culture. I’d read and even reviewed the book several months before, so I thought I had a good handle on it, but still, I needed a refresher on what Chayka calls “algorithmic anxiety” and prompted ChatGPT to remind me what Chayka meant by this.
The summary delivered by ChatGPT was perfectly fine, accurate and nonhallucinatory, but I couldn’t manage to go from the notion I had in my head about Chayka’s idea to something useful on the page via that summary of Chayka’s idea. In the end, I had to go back and reread the material in the book surrounding the concept to kick my brain into gear in a way that allowed me to articulate a thought of my own.
Something similar happened several other times, and I began to wonder exactly what was up. It’s possible that my writing process is idiosyncratic, but I discovered that to continue to work the problem of saying (hopefully) interesting and insightful things in the book was not a summary of the ideas of others, but the original expression of others as fuel for my thoughts.
This phenomenon might be related to the nature of how I view writing, which is that writing is a continual process of discovery where I have initial thoughts that bring me to the page, but the act of bringing the idea to the page alters those initial thoughts.
I tend to think all writing, or all good writing, anyway, operates this way because it is how you will know that you are getting the output of a unique intelligence on the page. The goal is to uncover something I didn’t know for myself, operating under the theory that this will also deliver something fresh for the audience. If the writer hasn’t discovered something for themselves in the process, what’s the point of the whole exercise?
When I turned to an LLM for a summary and could find no use for it, I came to recognize that I was interacting not with an intelligence, but a probability. Without an interesting human feature to latch onto, I couldn’t find a way to engage my own humanity.
I accept that others are having different experiences in working alongside large language models, that they find them truly generative (pardon the pun). Still, I wonder what it means to find a spark in generalized probabilities, rather than the singular intelligence.
I believe I say a lot of interesting and insightful things in More Than Words. I’m also confident I may have some things wrong and, over time, my beliefs will be changed by exposing myself to the responses of others. This is the process of communication and conversation, processes that are not a capacity of large language models given they have no intention working underneath the hood of their algorithm.
Believing otherwise is to indulge in a delusion. Maybe it’s a helpful delusion, but a delusion nonetheless.
The capacities of this technology are amazing and increasing all the time, but to me, for my work, they don’t offer all that much of meaning.