I’ve just complete a project using AI, namely Claude. It involved linguistics.
I’m new to using AI seriously; I first used Claude only a month ago. AI has been mostly irrelevant to what I do. When I write, generally it’s to offer a new take. AI doesn’t seem good at novelty. For one thing, it’s too contaminated by contracts conventional wisdom. (See this blog post about how Claude invoked the fallacy of “tested” contract language.) And I assume AI would never have made the connections on display in, for example, my article on agreement and plan of merger. (I had no interest in seeing whether it could, because I was able to make those connections for myself.)
That said, I’ve found AI of some use. For example, it offered suggestions for a bit of polemic I wrote recently, leading me to make some worthwhile changes.
But this new project called for a different approach, because I’m no linguist. So I enlisted Claude. My goal was to overhaul a 10,000-word analysis. The project ended up occupying several hours a day over eight days.
The process was iterative: Claude would offer specific comments on a small part of my analysis. That started a back-and-forth that could be protracted. I recall noting on one occasion that finalizing a paragraph of seven or eight lines had taken three hours. (Admittedly, I spent some of those three hours gazing into space—something Claude doesn’t do.)
I’d previously found Claude prone to responding in a waffly, wordy way when presented with more open-ended or value-judgment kinds of questions. On this project, Claude’s responses were clear, concise, and almost always to the point.
The project worked. Thanks to Claude’s assistance, I was able to dramatically reorganize my analysis and make it more accurate and more detailed.
I suggest this project succeeded because of the following factors:
- The original version of my analysis made for a good starting point. Rodney Huddleston, co-author of The Cambridge Grammar of the English Language, had helped me with parts of it. (Go here for my 2016 blog post about my visit with Rodney.) And over the years I had been able to add to it. So although it needed serious work, it contained all the components I needed.
- I wasn’t looking for big ideas. Instead, my needs were technical: I needed a linguistics gloss on constructions noted in my original analysis and those I’d subsequently encountered. The goal was to create a taxonomy for my analysis, to make it more orderly and easier to digest.
- Thanks to my decent semantic acuity and my having been exposed to Rodney’s work, I was able to understand the linguistics concepts Claude deployed and determine whether they were relevant.
Did Claude make mistakes? Yes. But given the detailed nature of the work and given our constant back-and-forth, I spotted them easily enough. Claude made the necessary adjustments, and we moved on. And occasionally, I would consult other resources, including The Cambridge Grammar of the English Language, to check Claude’s take against what they had to say.
Claude also kept track of what we had discussed and noticed if some component was missing or in the wrong place. As I was immersed in detail and prone to losing track of some of the moving pieces, that function proved valuable. And I was happy to take advantage of Claude’s spotting stray typos.
The alternatives to consulting Claude would have been unrewarding. I know it, because I’ve gone through that process several times. I would have had to look around haphazardly for one or more linguists willing to lend a hand. I’m not confident I would have succeeded. I once had to resort to paying a linguist for what turned out to be their ho-hum take. A couple of years ago I sought authorities on vagueness; I couldn’t find any able to offer the sort of insight I was looking for.
Even if I had found someone to help with this project, they wouldn’t have been able to devote to it anything like the attention that Claude brought to bear. Claude was always willing to tackle whatever I threw at it. And Claude responded almost instantly.
So without Claude, my project would have taken months, not eight days. And I expect the result wouldn’t have been as good.
I have no qualms about the role Claude played. It’s not as if I plan to submit my work for a prize in creative writing. All that matters is that Claude helped make my work better, and those who consult my work will benefit. If you’re concerned about my using Claude, perhaps you should also be concerned that Rodney Huddleston helped me out. When I publish my analysis, I expect I’ll comply with any conventions used by my peers to disclose that they’ve used AI in preparing all or part of a given work.
And I didn’t surrender agency to Claude: I scrutinized everything it offered.
But don’t assume from this I’m now “pro AI,” whatever that might mean. I found this adventure with Claude to be edifying, and I now understand slightly better the implications of AI. But deciding what we do about AI and coming up with suitable policies would require we take into account how we all experience AI and what would further the common good. Consider this account a drop in that ocean.
(Do you have a comment you’re itching to unleash on the world? Consider posting it on this blog. I’ve reinstalled Disqus to make that a more appealing prospect.)
