The author coins "expansion artifacts" as the AI-era counterpart to compression artifacts. While JPEG and MP3 discard imperceptible data to shrink files, large language models do the opposite—they expand sparse prompts into full outputs by extrapolating from blurry training data. The result: text stuffed with hedging verbs like "delve" and "intricate," essay-structured paragraphs with signposted takeaways, and code that over-comments obvious operations.

Stanford researchers detected these tells by tracking word frequency spikes after ChatGPT's release, estimating 17.5% of recent computer science papers contain AI-drafted content. Image generators produce six-fingered hands and suspicious jewelry; video models let limbs disappear and physics break down. These flaws function as forensic markers—breadcrumbs revealing AI involvement just as compression artifacts expose edited photos.

The danger compounds when generations feed into generations: a CEO voice memo expands into a strategy doc, which becomes product specs, which vibe-codes a prototype, which generates launch copy. Each stage interpolates from the previous, pulling from the "blurry JPEG" of training data. The author even takes credit for indigo becoming an AI aesthetic default, apologizing for making every Tailwind UI button that color five years ago.