Matt Ström-Awn has a name for what's wrong with AI-generated text, code, and images: expansion artifacts. The concept flips how we think about AI output. Ted Chiang called ChatGPT a "blurry JPEG of the web," suggesting the loss happens during compression. Ström-Awn argues the real damage occurs during decompression, when models extrapolate from compressed training data and fill gaps with plausible-sounding filler. The result is text bloated with hedging language, code that over-comments the obvious, and images with too many fingers.
Stanford researchers tracked these artifacts in academic writing. Their analysis of nearly one million papers found up to 17.5% of recent computer science papers contain AI-drafted content. You can spot it by watching for words whose frequency spiked after ChatGPT's release. One Elsevier paper opened with "Certainly, here is a possible introduction for your topic." The prompt leaked straight into publication.
The real worry is compounding. When AI output feeds into another AI, which feeds into another **AI agents**, artifacts don't just persist. They amplify. Ström-Awn describes a chain: a CEO's voice memo becomes a strategy doc, which becomes product specs, which gets coded by Cursor, reviewed by Devin, and launched with ChatGPT-written copy. Each step interpolates the previous output against the model's training distribution. When that output ends up in next-generation training data, you get a feedback loop that homogenizes information across the internet.
Researcher Mike Caulfield has built a **Claude skill** called "jamesian" that tries to restore rhetorical complexity to AI prose. It's a small countermeasure against a problem that's compounding daily.