Machine tells
Run this over a finished batch before delivery. Most of these survive a read of any single answer and only surface when the batch is read together, which is also how a moderator and a regular reader of the topic encounter it.
The tells are not vocabulary. Word-level scrubbing does nothing here. What gives a batch away is structure: evenness where a person would be lopsided.
Structural tells
Comprehensive, balanced coverage. The loudest one. A person answering a question they care about spends most of the words on the part that irritates them and waves at the rest. Even coverage of every sub-topic is what a summarizer produces, not what an operator writes.
Fix: pick the one thing worth being worked up about, give it most of the answer, and reduce everything else to a clause.
A tidy resolving line at the end of every answer. Real replies trail off, qualify themselves, or stop once the point is made. A closing kicker every time is a template signature.
Uniform paragraph length. Especially three-to-four-line blocks repeating down the page.
Punch-punch-punch rhythm. A sharp line ending every paragraph. No one sustains that. Spike where it is earned and let the other endings fall flat.
Uniform answer length. Length across a batch should be deliberately unequal, roughly 90 to 400 words. A five-line answer among eighty long ones is the only thing on the page nobody has to scroll, and it will often outperform the long ones.
Shared skeletons. If three answers open with the same move, or run claim → three supports → close, the profile reads as templated even when every individual answer is good.
Sentence-level tells
Contrast flips. "It's not X, it's Y." One in a batch is fine. It appears three times per batch by default and has to be cut back.
Triads. Three parallel items, three parallel clauses, three examples. Use two, or four, or one.
Balanced parallel construction generally. Symmetry reads as generated because irritation is asymmetric.
Content tells
Generic operator wisdom. "Do direct outreach." "Focus on quality." "Build relationships." These teach nothing, and they read as machine-written precisely because they carry no knowledge — there is no way to write them from experience that distinguishes them from writing them without any.
Fix: every answer carries at least one mechanic only a practitioner would know. Test it by asking whether a competent competitor sitting in the same thread could have written the line. If yes, it is not method, it is filler.
Nothing unresolved. Human answers leave loose ends: an aside that does not advance the argument, a hedge where the honest answer is genuinely uncertain, a point raised and dropped. A batch with no loose ends anywhere is a batch that was generated.
Behavioural tells
These are not about the writing, and they flag a profile faster than any prose problem.
- Answering incoherent or auto-generated questions. That is what farming profiles do, and it is visible on a profile page at a glance.
- Two answers on the same subject a few days apart.
- A whole seeded cluster answered in one day.
- Fabricated experience with a named product. Permanent damage, because the people who use that product are the ones reading.
The batch read
Last pass before delivery: read only the first two lines of every answer in order, then only the last line of every answer in order. If either sequence reads as one voice repeating a habit, the batch is not ready. It should read as one person on different days.