June 12, 2026
i hate AI
a marker, written down so i can check back later
i hate AI. and in a few minutes i'm going to use Claude Code to automate posting this exact rant to neocities and instagram.
i thought about that for a while before writing this. the contradiction is the whole post.
it started small for me. summarizing instead of reading. drafting instead of writing. asking instead of thinking. each thing on its own felt fine, barely worth noticing. but a hundred small things like that and suddenly the tool is the first move, the head is the second.
what gets me is how little friction there is. there used to be a small wall between "i wonder" and "i know" - looking things up, sitting with a question, working through it. that wall is gone now, and i didn't vote on it, i just walked through it like everyone else.
people argue about jobs and automation and all that, fine, that's its own thing. the part that bothers me more is quieter - it's people slowly giving away the bits of themselves that used to do the work. the muscle doesn't go all at once. it just gets a little weaker every time it sits unused.
and it's not just people getting weaker. the models are too, and there's a name for it now: model collapse. it comes out of a paper called "The Curse of Recursion: Training on Generated Data Makes Models Forget" - work by Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot and Ross Anderson, out of oxford, cambridge and a few other places, that later landed in nature. the idea is simple once you hear it. AI is trained on whatever text is sitting on the internet. the internet is filling up with AI-written text. so the next model trains on the last model's output, and the one after that trains on that, and so on.
every time that happens, something gets lost. not randomly either - the weird stuff goes first. the unusual phrasing, the odd opinions, the rare words, the things that don't sound like everyone else. the researchers call this the "tails" of the data - the low-probability edges of the distribution. those tails get cut off generation after generation, and what's left is the middle - the average, the expected, the safe. they define collapse plainly: training on model-generated content causes "irreversible defects" where the tails of the original distribution disappear. irreversible is the word that should scare you.
and they didn't just see it in language. they ran it on variational autoencoders, on gaussian mixture models, on large language models - different machines entirely, same result every time. by a handful of generations the output narrows down to one flat voice. there's a now-famous example from their runs: a model fine-tuned on text about medieval english church architecture, asked the same kind of question generation after generation, eventually stops talking about architecture at all and starts babbling about jackrabbits. the meaning rots from the tails inward until it's noise wearing the shape of an answer.
one of the researchers, ross anderson, framed it bluntly - large language models are a bit like fire: useful, but something that pollutes the environment around it. and the pollution doesn't announce itself. nobody's tagging which paragraph online was written by a person and which was written by a model. the next model can't tell either. it just absorbs all of it, a little more diluted each time, and there's no way to undo that once the original variety is gone.
that's the bigger version of the thing i was already feeling on a personal level. if i lean on AI for everything, my own writing flattens toward whatever's average. if everyone does that, the internet flattens too. and then the models train on that flatter internet and flatten further, and feed that back to people, who flatten more. it's the same loop at two different sizes.
so why am i still doing this then.
because the choice in front of me isn't "do it properly yourself" vs "use AI." it's "post something" vs "post nothing," because i'm tired or busy or it's late and there's a job in the morning. i'm picking the smaller compromise, and i know it's still a compromise.
the part i want to keep is the thinking - sitting with an idea long enough that it's mine before it goes anywhere. that part i'm protecting. the posting and formatting and scheduling is plumbing, and i don't need to do plumbing by hand to feel like a person.
but that's also exactly the kind of line that moves on you without asking. it's posting now. it could be drafting later. and if it ever becomes thinking, it'll probably still feel like mine, which is the scary part.
so this is mostly a marker. dated, written down, so i can check back later and see whether i held the line or just stopped noticing where it was.