The AI tell isn't in the sentence — it's in the set

I wrote a batch of explainers for this site and reviewed each one the usual way — no hype, no filler, no buzzwords. Individually they passed.

Read as a set, they gave themselves away. The same closing move on every piece. The same "the interesting part isn't X, it's Y" turn. The same em-dash rhythm, over and over. Nothing in any single paragraph was wrong; the pattern was in the repetition across all of them.

That's the part a per-piece edit can't catch. You're checking one document against a quality bar, and the bar is met. The tell lives one level up — in the sameness between documents that no single document can show you.

So I added a pass that reads the pieces side by side and looks only for what repeats: shared structure, recycled transitions, identical cadence. De-duplicate at the set level, not just the sentence level. Restraint within a piece isn't enough when the whole corpus moves in lockstep.