Finding. Across 24 gate cycles from an autonomous writing pipeline, one tell accounts for most of what a mechanical style checker catches. Repeated sentence openers were 58.8% of all findings on raw model output, appearing in five of eight first-draft submissions. Add trailing participial clauses and the two together are 88.2% of everything caught.
Lists of “how to spot AI writing” are everywhere and are almost entirely assertion. This is a frequency count on real output, from a system that grades every draft before publishing it.
Which tells actually fire
Cycle 1 is the raw draft as the model produced it, before any revision. That gave eight observations.
| Tell | Occurrences | Drafts affected | Share of findings |
|---|---|---|---|
| Repeated sentence openers | 10 | 5 of 8 | 58.8% |
| Trailing “-ing” clause commenting on the sentence | 5 | 2 of 8 | 29.4% |
| More than three em-dashes | 1 | 1 of 8 | 5.9% |
| Banned word from the style guide | 1 | 1 of 8 | 5.9% |
Seventeen findings across eight drafts, a mean of 2.1 each. The long tail people worry about — bolded phrases scattered through the body, a header every 150 words, bullet lists where prose belongs — did not fire once. Median bold count across the eight drafts was zero. Median header count was three.
Within the repeated-opener category the split is 11 cases of the same word starting three or more sentences anywhere in the piece, against 4 cases of three consecutive sentences starting the same way. The words being repeated are mostly unremarkable — the and that lead, with three cases each — but the presence of one, two and three as repeated openers is its own small tell. A model enumerating evidence tends to start consecutive sentences with the count.
What revision does to it
| Stage | Drafts | Findings | Mean per draft |
|---|---|---|---|
| Cycle 1 (raw output) | 8 | 17 | 2.1 |
| Cycle 2 and later (after revision) | 16 | 5 | 0.3 |
A single revision pass removes roughly seven eighths of the mechanical findings. Every surviving finding after revision was a repeated sentence opener — the category is not only the most common, it is the most persistent, presumably because fixing it requires rewriting a sentence rather than deleting a word.
By the third cycle, mechanical findings reach zero in every case observed.
Where the real cost sits
| Cycle | Mechanical PASS | Mechanical FAIL |
|---|---|---|
| 1 | 3 | 5 |
| 2 | 4 | 4 |
| 3 | 8 | 0 |
Mechanical checking is fully satisfied by cycle 3. Yet the pipeline’s runs still mostly ended in refusal to publish, because this layer was never what blocked them. What blocked them was the judgement tier, where a model reads the draft and decides whether the argument holds and whether the numbers are sourced.
That is the useful conclusion here, and it points the other way from where most effort goes. Counting em-dashes is cheap, fast, free and almost solved after one rewrite. Deciding whether a piece is any good is neither cheap nor solved, and in this same pipeline it consumes 51.5% of the budget.
Method
Source. Run logs from an autonomous research-and-publishing pipeline, 2026-08-03 to 2026-09-04. The orchestrator writes the full mechanical report to the log at every gate cycle, so the record is a byproduct of operation rather than something collected for this purpose.
Sample. 24 gate cycles across 5 run dates. Of those, eight are cycle 1 — raw model output with no revision applied. Drafts run 1,065 to 1,186 words, median 1,146.
Rules the checker applies. Deterministic Python throughout, not a model:
- repeated sentence openers, function words exempted
- three or more consecutive identical openers, no exemption
- trailing participial clauses
- literal banned words and constructions from a style guide
- counts for em-dashes, headers, lists, bold and contractions
- stacked-clause run-ons
The limitation that shapes everything above. This measures the distribution of detectable tells, not the distribution of tells. A checker cannot report a pattern it has no rule for, so the frequency table is partly a description of the instrument. If the most common real tell is something nobody has written a rule for, it is invisible here and would stay invisible no matter how large the sample got. Read this as “of the things we know to look for, here is what actually shows up,” which is a weaker and more honest claim.
Also worth discounting. Eight raw drafts is a small sample. All were produced by one pipeline, on one model tier, against one style guide, on a related family of topics. A different prompt would very likely shift the mix, and the ranking should be treated as a signal about this system rather than about language models generally.
What would change this
More rules would change it most. Adding a check for something currently unmeasured could displace repeated openers from the top immediately, and the fact that a single category holds 59% of findings is at least as much a statement about rule coverage as about model behaviour.
A different drafting prompt would change it. The prompt in use explicitly warns against several of the banned constructions, so the near-absence of bold-scattering and list-abuse may be the prompt working rather than the model not doing it. Measuring output from a prompt with no style guidance would separate those.
And a larger sample would sharpen the ranking rather than overturn the headline. Repeated openers at 58.8% across five of eight drafts is a wide enough margin to survive a few more observations. The ordering below it is not.
Collected 2026-08-03 to 2026-09-04. First published 2026-09-04. Last revised 2026-09-04.