Why AI writing is so exhausting

I rage-deleted Claude from my phone last week. I had to, to save the phone. The tics, man. I would open the app, ask a question, see "earns its keep" or "the honest answer" and have to suppress the urge to shatter the phone on the ground. Deep breaths.

I'm getting to the point where if I start to even get the sense that what I'm reading is AI-generated, I just... close the tab.

The problem isn't the tics, though the tics are a problem. The best description I've read of what feels "off" to me about AI writing comes from "Honey, We Bought an AI Story," a post by Bona Books about how they determined that two stories in a short story collection they were putting together appear to have been generated, rather than written.

The closer we examined that first draft, the more the story’s centre seemed absent. The protagonist was active, but lacked interiority. The worldbuilding had colourful details that somehow failed to build consistently. Certain hallmark sentences recurred over and over (“Not X, but Y…”). And the pacing was off, with important story beats given oddly little narrative weight.

It's this uncanny absence of a center, a thesis, an animating theory behind the words on the page that gets to me.

I find reading AI generated text exhausting. This is a big professional problem for me right now. I don't mind, or I don't think I mind, reading AI generated code, but to use an agent I have to talk to it. I have to make sense of these walls of yap. I've been moving away from expansive and free-wheeling ways of using agents towards something much more locked-down and scalpel like, where I spend a lot more time working out what I want it to do before I even open the chat window, just because I find talking to the thing so tedious.

I think what's going on here is that the lack of intent makes the texts harder to decode. When I read text that was written by a human being, I get a lot of help comprehending it from the fact that there's something there to comprehend. A human being wrote it and they wrote it for a reason. The author was trying to tell me something. If I can grasp the thesis of a piece I can use that as a structure to slot the rest of the details into the correct place. Even just trying to build a hypothesis of the thesis helps me grab hold of those details.

Human-written work has natural detail. The structure is fractal, like a tree. If you zoom out, you can see the trunk of the argument, and then the branches. If I lose my place, I can go back up a level of detail and reorient myself. If I need to understand something more precisely, I can zoom in and read word-by-word.

AI-written text doesn't have any of this. It doesn't have a thesis. It has, at best, a ghost of a thesis, the residue of the original prompt, smashing into whatever whisps of meaning it can find nearby in the model's conceptual matrix. It's an artificial pattern, with a continuous level of detail all the way through. The words have relationships to each other but it doesn't have that fractal structure. Get confused? Too bad. There are no branches to hold on to.