De-AI documents

Turn machine-polished artifacts into things that read like a person made them. The work has two sides. Strip the tells that fingerprint machine authorship, and put back the residue of human process. Doing only one side fails. Clean prose over an untouched generated notebook still leaves tell-tale code, and roughed-up prose next to an execution trace that says "generated in one pass" fools nobody.

What fingerprints machine authorship

Uniformity. Humans drift, models are consistent. Any property identical across a whole artifact is a confession: one quote style across every string, one prose register across every paragraph, one comment grammar, one sentence length, an evenly spaced error taxonomy. When auditing, hunt for properties that never vary.

Impossibility. A sentence that cites a number which exists only in the output of a later computation could not have been written before that computation ran. Humans run first and write after. If the artifact has runtime outputs (notebook cells, logs, build artifacts), check every claim in the text against them.

Missing debris. Human work leaves residue: a debug print that stayed, a lazy variable name, an inspection cell nobody references. First-pass machine output is spotless.

Uniform imperfection. This is the second-order tell that catches naive scrubbing. Detectors measure error rates, not just errors. A misspelling that appears five times with zero correct uses of the same word is a system. Nine dead ends that all sit one line above their fix with a little apology are a genre. A blemish in every paragraph, evenly spread, never touching the reasoning, is the signature of a humanization pass. Any layer added to fake process becomes its own tell once its shape is uniform.

The stable position

After enough rounds against detectors the only configuration that holds up is this: