Removing AI Tells

A detection and rewrite system

A practical guide to the rhetorical signatures that mark AI-generated text, and the detection and rewrite pipeline that removes them before they reach a reader. Drawn from a production system built to gate real manuscripts.

If your AI-assisted writing passes a grammar check but still reads as machine-made, this is the system that fixes that.

Five areas, one detection system

• echo-devices — The repetition patterns Claude defaults to: anaphora (same sentence opener ×3+), epizeuxis (immediate word repetition), diacope (repetition with intervening phrase), epistrophe (same ending ×2+). Detection is regex-based, flagged by paragraph, with the matching evidence string quoted in the output.
• parallel-structures — Why AI writing over-uses balanced sentence pairs and the three-item list with the Oxford comma. Detection: consecutive sentences of near-equal length, lists of exactly three abstract nouns, repeated subject–verb–object structure across adjacent sentences. Rewrite target: vary length, merge two items into one richer clause.
• banned-constructions — The X-not-Y framing (“not a bug, but a feature”), the negation pivot (“not X. Instead, Y.”), and the false-contrast frame (“while X, Y” without a real contrast). All three appear in nearly every AI draft and read as machine-generated on sight. Detection is single-pass regex with zero false-positive tolerance.
• detection-pipeline — A two-pass scanner: LIKE Engine (rhetorical tells, pattern-match, exits nonzero on any hard-fail) followed by ARC Engine (statistical tells: hedge density, elaboration-chain depth, passive-construction ratio). Both return structured JSON with evidence, both gate on exit code. Zero model calls in either pass.
• rewrite-discipline — The editing pass that removes tells without collapsing the sentence. Rules: replace anaphora by varying the syntactic subject. Replace banned constructions by cutting the contrast and stating the positive directly. Minimum rewrite principle: change one word if one word is enough.
• parallel-structures — Why AI writing over-uses balanced sentence pairs and how to break the symmetry without losing the point.
• banned-constructions — The X-not-Y and “not X, but Y” framings that appear in nearly every AI draft and read as machine-written on sight.
• detection-pipeline — How to build a scanner that flags these patterns automatically before content ships.
• rewrite-discipline — The editing pass that removes tells without collapsing the sentence structure into something worse.

A preview

AI writing doesn’t sound artificial because it’s wrong. It sounds artificial because it’s too smooth — every sentence balanced, every transition marked, every point restated. The tells are in the symmetry.

The goal isn’t to make the writing worse. It’s to make it stop sounding like it was produced by something that had read too much and experienced nothing.

Anyone shipping AI-assisted content

Anyone using Claude to produce written content — books, articles, documentation, copy — who wants it to pass the reader’s instinct that something is off. Standalone; no prerequisites.

A longer excerpt is available to newsletter subscribers.

Buy on Amazon