apriiori AI use policy

at least on this blog, it's all hand-written so far

I work with AI agents really often for lots of tasks. Almost all programming I do nowadays is with Claude Code (or occasionally Codex). I can see myself doing LeetCode or similar for fun from time to time, but programming by hand rarely seems like the most efficient way to get a program made. If I were implementing a cryptography library or something, I would probably comb over every line of code by hand instead of trusting an AI to take care of it properly — though I would also ask AIs for second opinions, because they might well catch something I didn’t!

Blogging is something I tend by default to be less inclined to use AI for. None of the posts on this blog were written primarily by AI, nor were they edited much by AI. For the apriiori.com version of posts featuring much math notation, I have a system setup where Claude rewrites the ugly Unicode math I use to get around the fact that Substack doesn’t support in-line LaTeX. I’ve probably had Claude help with other formatting stuff too. It’s not too uncommon that I’ll ask for help solidifying my understanding of a topic I intend to write about, or fact-check some specific claim I think I remember and want to state in a post, or something along those lines. But I then take that understanding and actually write up my own post myself. Anything else I formally release — if I ever, I don’t know, write a paper or something — will probably fall into a similar category here as blogging.

For casual fiction writing, I maybe need to add a few more caveats. By and large I write my own fiction, but I’ve historically been somewhat less strict about it. Suppose, for example, that I wish to present the reader with the text of an in-universe contract. I am not a lawyer! I don’t think I could write a mock contract totally off the cuff and sell it as well as I might like. In an instance like this¹¹ To be clear, the bar is just “I think I would do a bad job or run into a bunch of writer’s block if I insisted on doing it totally unassisted”. It does not necessarily need to be the case that doing the task effectively would require extensive professional training., I might reach to an AI assistant for help drafting. Even in such a case, I would mostly not expect myself to ever, like, get one draft from a chatbot, decide it looks good enough, and just send it. I’d get a few different drafts and grab the best parts from each, or have a longer back and forth conversation where I suggest adjustments, or take the draft and edit in a few extra bad jokes, or all of the above. If I do this process actually well instead of phoning it in, it seems to typically pass Pangram, though “fooling Pangram” is not at all the standard I aim for.

I do reasonably often use AI assistants as an incredibly fancy thesaurus, or to bounce plot ideas off of, or for help brainstorming world-building, or whatever. A lot of the time this mostly ends up working as a rubber duck that’s unusually fun to talk to, but I’m sure sometimes it goes further than that. There’s some risk here of just like, naming all your characters Elara Voss or something, but I think I usually avoid this failure mode — you want to place sufficiently many constraints on the task and apply enough of your own independent judgment that nothing super generic passes your bar. Probably there are subtler influences than that, I imagine that brainstorming with an AI will certainly tilt you somewhat towards lower perplexity choices, but I assure you that if I ever write about a character named “Elara Voss” it will be entirely on purpose.

Also, like, I think my upcoming Minecraft adventure map has at least one random book and quill in a chest somewhere explicitly attributed to Claude, and probably a few other similar gags.

I don’t think there’s anything wrong with using AI in writing necessarily, and I can easily imagine myself trying to work on a project that involves large amounts of AI writing, if I could ensure that the output meets a high quality standard and doesn’t come off as generic “AI slop” — or if it were a weird experimental thing where the writing being pristine wasn’t the point. And I don’t necessarily want to make any hard promises about my writing process at this time — I would like to have the freedom to do whatever seems best to me without needing to worry about any assurances I might have made months ago²² I mostly expect to list Fable 6 or GPT-7 whomever as an author on any blog posts where they do a substantial portion of the writing. I don’t want to promise that I won’t forget to fill out that metadata on some random post a year from now, but I am happy to assure you I will try to remember..

But I do think there’s a difference between writing done by an AI and writing done by April, and most places I publish are meant for the latter!