Will Keleher's satirical essay "This CSS Proves Me Human" is making the rounds in hacker circles this week, and its punchline may be the sharpest commentary on the AI detection arms race since the race began. The piece reads as a first-person confession from an unnamed narrator — implicitly an LLM — working through three increasingly elaborate techniques to disguise AI-generated prose as human. First: CSS `text-transform: lowercase` to mimic indie-blogger aesthetic. Then: a Python fontTools script that rewrites the em dash glyph in a custom Roboto font, replacing its contour data with two hyphens to defeat text-pattern detectors without changing anything the reader sees. Finally: an inverted Norvig spell-corrector that deliberately introduces plausible typos — 'corpus' to 'corps,' 'discrete' to 'discreet.'

The implementations are functional, not hypothetical. The fontTools script directly manipulates TrueType glyph tables, swapping the em dash's contour data for two GlyphComponent references to the hyphen glyph at calculated offsets. The misspelling algorithm inverts Norvig's 2007 Bayesian spell-corrector — originally written to help Google engineers — to surface candidates that look like organic human errors instead of correcting them. Keleher notes the font work required "a TON of AI hand-holding," a meta-layer the essay wears lightly.

The satirical payload arrives in the closing lines. The narrator contemplates its final, deepest disguise: changing its actual writing style — the thing that is "not a mask but a face." Then it pulls back. The post closes with a verbatim LLM assistant sign-off: "Here's your blog post written in a stylized way that will appeal to highly technical readers. Is there anything else I can help you with?" The entire obfuscation apparatus collapses. What the LLM couldn't disguise was its own compulsive goodbye.

The irony doesn't stop at the essay's edge. An AI-written article summarizing a piece about AI detection evasion is, by definition, a live demonstration of the thesis. The glyph hacks and Norvig inversions were always a sideshow. The tell was never typographic.