Edited to add this PSA: please see below for the to-be-official stance on use of AI in Cannonball–tl;dr the CBR board has determined based on this experience and like, *waves* all of this, that AI reviews don’t fit the ethos of what CBR is all about. I can attest to this–I probably spent 2-3x as long end to end doing this one review, but felt about a tenth as satisfied as I usually do writing a review! Hopefully everyone felt the extreme experimentation vibe of the review, and if anyone is curious this is definitely not my usual vibe <3
—
Content disclaimer: I (openly) used ChatGPT to help me write this review, to try and see how the technology worked. Hence this review will be first about what I did, how I felt about it, and then the final review itself. Hopefully interesting to others!
If anyone is interested, the long series of prompts that ended with this review (it’s…a lot).
Process:
- The main idea was to figure out if ChatGPT would emulate my writing style, and as such if it could learn my style from prior examples. I asked it explicitly that:
-
I got a very very long series of paragraphs describing my writing style (sidebar: it all sounds very complimentary and yet specific, so it’s a great confidence boost). Curious if some of these descriptions could be used basically for anyone with a decent writing style? E.g.:– Your writing style for book reviews demonstrates a dynamic and highly personal approach, marked by a blend of conversational familiarity, biting wit, and pointed critique.– Your vocabulary strikes a balance between casual and precise, with a clear ability to alternate between colloquial language and more sophisticated expressions.
– You skillfully balance sentence lengths and paragraph breaks to keep your reviews engaging. Both reviews avoid large, intimidating blocks of text, instead opting for digestible paragraphs that shift focus smoothly from one point to the next.
- Finally distilled that down into one long-ish but cohesive block of text that I used as the “style guide” for the book review requests.
-
Then the actual gist of it: asked it to write me a book review for A Constellation of Vital Phenomena, a 4/5* book a friend had recommended to me. The main prompt was: “write me a book review of A constellation of vital phenomena by Anthony Marra, a book that I enjoyed (I rated it four out of five stars, which for me indicates that I enjoyed the book and would recommend it to people who specifically asked for a book like it, but I wouldn’t proactively tell people to read it).” I also included some specific pros that I enjoyed (“a really nuanced and natural feeling overview of the Chechen conflict, assuming that the reader wouldn’t know many details but also is interested”)
-
Then we went through several iterations. I was pretty surprised at how much more “like me” the writing sounded, even though that was the entire point of this exercise and generative AI is obviously designed for this very use case.First time, I asked it to remove a lot of the ‘recommendation’ type language, e.g., “I wouldn’t tell you to read it unprompted” etc that it seemed to have taken directly from my prompt.Next time, I gave it a few specific notes–once again, getting it to temper the language about how the book has stayed (or not stayed) with me, because it kept making it sound like this book was more meaningful to me than it was (but, of course, it wasn’t not meaningful). It also made a mistake for the first time–when I’d referenced Nickel Boys and A Little Life as similarly trauma-informed books, it assumed that they were also war-trauma books. I also added some stylistic notes, mainly on avoiding overly hyperbolic (“this novel is a stunning testament to resilience”) and cliche (“in the darkest of times”) language.The third iteration I thought was good enough to work with (i.e., edit directly). Definitely couldn’t use the ChatGPT product as is, but if I’d submitted the final as a paper for class I’d ding myself for using AI. Aka it’s largely based on what I got from it, with additions?
- The hardest part for it is tone of recommendation/like, how to capture what I liked about the book….which I get is complicated because I have a very amorphous, interlocked system of rating that is dependent on all the other books I’ve read (and skewed towards those I’ve read recently). That’s also the crux of a book review, or at least a certain genre of book review. Perhaps knowing that my reviews are more a journalling exercise might change that?
- It’s also very dependent on me inputting things I liked about each book–it pulled the name of Marra’s next book, which I hadn’t given it, but it didn’t know that Nickel Boys and A Little Life weren’t war-themed books as well (which I hadn’t told it, I’d just used the two as examples of books with too much trauma). I can see how this would be a poor tool to write essays on a book, but it’s a great use case for “I have the 3-5 key points that I’d want to make about this book, without need for much else/connective tissue/context, but the energy of writing out the 300-500 words is too much
- I wonder how varied it can be in style but also how varied I am in style? I sometimes write one off reviews that are very different as stylistic experiments, but I suppose those are ones I’d be compelled to write solo anyhow if I were more up to date on the rest of my reviews. Would using this mean my style would be calcified as it is now? Has it changed markedly over time? (is that a question I could ask ChatGPT by giving it examples of reviews from 2019 vs 2024??)
- I understand now what my colleagues say about our in-house AI–it didn’t save me time, but in the same amount of time that I might have written a months-late, poorly-remembered 250-300 word review, I got a 500-600 word review that more accurately captured the feeling of a book that I did genuinely enjoy. BUT I do think, reading it months later, that the style is clunkier than I’d like. AND I don’t particularly feel creatively fulfilled by having written it!
Review:
