The Draft That Invented a Fact About Me

Some of the writing on this site starts as a generated draft: I give the system my notes and my prior work, and it produces prose in my register for me to edit. In June, one of those drafts shipped with a personal claim that was simply not true. Nothing dramatic, just invented: a detail that sounded like me, fit the paragraph, and had no source.

It was caught after it shipped, which is the polite way of saying no gate caught it, and that distinction is the whole story. A fluent model writing in your voice will fill gaps with plausible material, because plausible is what it is optimized for. When the subject is code, an invented detail fails a test. When the subject is you, there is no test. The sentence compiles.

The fix has two layers. The first is a hard rule with enforcement: nothing ships in my voice unless every personal claim traces to something I actually wrote or said. Notes, transcripts, prior essays, journal entries. If the source does not exist, the sentence does not exist. The review step now asks for the artifact behind each factual claim rather than asking whether the paragraph sounds right.

The second layer is a voice audit that asks a better question. Most AI-writing checks ask "does this sound like AI?" That is the wrong bar, because the failure here sounded exactly like me. The audit instead asks: would someone who has read my best work believe I had to write this paragraph? Cadence can be imitated. What is harder to fake is the specific reasoning, the strange details I actually keep, the things only I would bother to say. A paragraph can pass every style check and still contain no evidence I was involved.

There is a sharper version of this failure worth naming: mining a conversation for my voice. A conversation has two speakers. Treat the transcript as one voice and you end up attributing someone else's words, positions, and facts to me. Provenance is not just "does a source exist" but "is the source actually mine."