Fussing with the word level of AI co-authored text feels a bit like pursuing photorealistic art after the invention of the camera. I mean, impressive that you can paint photorealistically, but I’d rather go look at impressionist, cubist art, or watch camera-based media like movies. If I want photorealism I’ll take or commission a photo.

Increasingly I don’t touch the word level, and rarely even suggest spot edits to ChatGPT unless it is truly strategic (like a word choice in a title or lede phrase). Litmus test — if you touch a word, it should be strategic enough to trigger a refinement edit pass (often it will suggest this, especially with titles…it seems to realize that titles are true names). I keep the prompting process firmly at the idea/tone/argument/style level. And rarely touch below a subsection-sized chunk. I’d like to do full-gestalt level modulation of entire essays, but that tends to be fragile. Mid-level structure offers the most control over the medium. Interfere too much at word/sentence level and it will thrash. Do full rewrites too much and it will lose intent in unstable ways and circle the intent like an amnesiac unreliable narrator. Subsection to paragraph is the strategically grippable abstraction range for current ai models. I iterate at that level and assemble. Not quite vibe-writing but maybe contraption-vibing.

So far I’m still aiming for “AI-realistic” writing that “photographs” an idea in latent space in a very conventional form like an essay, story, briefing etc, just as early photography favored portraiture style posing. But I’m starting to see what both AI-native kinds of writing (equivalent of movies/tv for cameras) might look like, and what “anti-AI-realistic” handcrafted writing should look like (the equivalents of Impressionism/cubism etc). The key is to look through the lens of AI writing even when you’re not using AI to write. Impressionism takes a cue from the camera’s ability to capture fleeting moments and lighting. Cubism takes a cue from multiple point-of-view ways of seeing, which is easy for cameras. I’m sensing there are similar moves to be made in relation to AI. For example, at the AI fiction workshop we did in Austin, one participant wrote something where the rest of us noted a kind of syncopated style that felt AI-like. He used AI in the reverse way from me — he used it for ideation and wrote the actual words himself, which I think will be a great direction (though too much effort for me). It won’t be AI-realistic writing, since there is AI upstream there. It will be more like post-camera painting styles.

The more I experiment the more addicted I get. I think 4o was the threshold for me personally, when the medium became genuinely interesting. Previously it was janky and frustrating to work with.