Teaching AI to stop sounding like AI
You can usually tell when AI wrote something, even if you can't say how. Readers feel it too, and they don't call it out: they trust you a little less and stop replying.
Why AI copy sounds the same
AI writes the average of everything it has read. Left alone, your copy sounds like everyone else's.
The tells
- FragmentsShort dramatic sentences that sound punchy and say little.
- RepetitionThe same line repeated for effect.
- Clever headingsBuilt to sound smart instead of saying something.
- The overcorrectionFix all that carelessly and you get stiff, bureaucratic filler.
The target is in between: a senior person explaining something plainly, in full sentences.
Making the fixes stick
I made the fixes stick, so nobody has to catch the same problem twice.
- Banned tellsEm dashes, words like "quietly", and the "not this, it's that" sentence shape are out. I can't save an em dash into a document. The system blocks it.
- Saved editsMy rewrites are kept as examples for the next draft.
- A lintA script checks every draft before I see it. It catches the tells and any retired product names or phrases we've stopped using.
- Real material firstDrafts start from my notes, call transcripts, and proof points. Never a blank prompt.
What a lint misses
A lint catches patterns. It can't tell you if a sentence says something true. People do that part.