The signals in AI-written copy that readers actually notice

Which tells matter for comprehension and trust, which are noise, and the five-minute editorial pass that fixes the ones that count.

ES6 QA AI editorial · 14 October 2026 · 6 min read

Readers notice structure before vocabulary

Editors spend a lot of energy on the wrong layer. They hunt for the word 'delve' and miss that every paragraph is the same length, every sentence opens the same way, and the piece makes four claims in a row without ever landing on one.

Uniformity is what readers feel. A human draft has a jagged rhythm: a long explanatory sentence, then a short one that lands. Model output, left alone, produces a metronome. That is the signal worth fixing first, because it affects comprehension rather than merely revealing the tool.

The tells that change meaning

Some patterns are not stylistic at all — they are risk. These are the ones to stop the press for.

  • Unsourced statistics. A number without a source is a liability, not a detail. Either link it or remove it; do not soften it to 'many'.
  • Absolute claims: guaranteed, 100%, never fails, proven. They invite complaints and, in regulated industries, fines.
  • Hedging that contradicts the headline. If the body is full of may and arguably, the promise at the top is not supported.
  • Invented specifics: a plausible-sounding study, a named expert who does not exist, a date that never happened. Check every proper noun.
  • Left-in scaffolding: [insert client name], TBD, {{variable}}, and the model's own disclaimers about being a model.

The tells that are just fashion

The phrase list is useful as a heuristic and dangerous as a rule. 'Robust', 'seamless' and 'leverage' are bad writing, not evidence of a machine, and plenty of human copy is full of them. Flagging them is fine; accusing on the basis of them is not.

Em dashes get the same treatment. They are a punctuation choice, widely used by careful writers, and treating them as a signature produces false accusations. Count them as one weak input, never as the verdict.

A five-minute editorial pass

Read the first sentence of every paragraph in isolation. If four in a row start the same way, rewrite three. Then read the piece aloud: the place where you run out of breath is a sentence to split.

  • Delete the first paragraph. Generated intros almost always restate the title.
  • Replace every abstract noun with a concrete one where you can.
  • Check that each number has a source and each superlative has a reason.
  • Cut every hedge, then re-add only the ones that are honest.
  • Read the last line. If it summarises what you just said, delete it — the reader already knows.

These rules are the same ones the console runs. Check a file against them now.

Start with one file

Find out what your AI output is hiding.

The console is free, needs no account and uploads nothing. See the findings first, then decide whether you want a specialist to sign it off.