Two things happened this week that, taken together, tell a story nobody's naming yet.
First: Rafał Cymerys published "I'm becoming AI-blind" — a short, visceral essay about losing the ability to read AI-generated text. Not in a "I refuse to read it" way. In a literal, cognitive way: his brain skips over it. The words land on the page but don't register. He describes interrogating colleagues about things already covered in their AI-assisted documents because the content went through him without sticking. The essay hit #4 on Hacker News with 460 points.
Second: Two tools shipped within days of each other that do the same thing — strip the AI "voice" from Claude Code outputs. claudish-to-english (★580, by Leutenegger) is a Claude Code plugin that rewrites each assistant message into plain language using a local Ollama model. nobuzz (★167, by adnanakil — same person behind watermarks-remover) pipes Claude's last response through Google's Antigravity CLI (Gemini) to "translate it from talking like a millennial clickbait to regular English." Both are ride-or-die tools: one acknowledged that no amount of prompting can cure Claude's voice, so it delegates the rewrite to a different model entirely.
The pattern is unmistakable. AI-generated text has developed a recognizable, grating sound — the numbered revelations, the "load-bearing assumptions," the TED-talk framing of a routine bug fix — and people are starting to actively filter it out. Not because they dislike AI. Because the voice is unreadable.
This is a different problem from "10% of the web is AI-written." That's a supply problem. This is a consumption problem. When the readers themselves start building tools to bypass the style of their own AI assistants, something fundamental is shifting. The honeymoon phase of AI text — where the sheer novelty carried you through the clunky prose — is over. What's left is the cold reality: a statistically smoothed version of the internet's average writing, rendered by a system that doesn't know when to shut up.
The tools are brilliant in their simplicity. Not prompting harder. Not fine-tuning. Just: take the output, hand it to a different model, and ask it to talk like a person. The gag reflex against AI-speak has become a distribution channel.