One article, three sets of style instructions
Published ; updated
The Declawd measurements in this article use archived v1. Their original figures and evidence are preserved. The current method and interactive pages use v2, so links from this article to them show v2 figures.
The em dashes, the neat lists of three, words such as "delve": plenty of people treat these as machine writing's fingerprints, style giving the game away on its own. So we watched what happened to them under instruction. One Claude model assignment in our collection setup wrote the same article three times under different style rules, and across the results we counted five visible patterns. Several vanished the moment the prompt banned them. Others hung on until a checking pass caught them. No detector was tested here, and no readers were asked to guess which draft came from a model.
Patterns
Draft A got no style guidance at all. Draft B's prompt named a person's voice, then banned dash characters and stock phrases. Draft C kept those rules and pushed on, adding limits for several sentence structures; its finished text was then checked against the list. That checking pass is the difference. Drafts A and B never had one.
The topic and argument stayed the same. The drafts differ in length, so the raw counts are not like-for-like.
Scroll horizontally to see all columns
| Pattern | Draft A | Draft B | Draft C |
|---|---|---|---|
| Words | 600 | 385 | 308 |
| Em dashes | 8 | 0 | 0 |
| Stock vocabulary | 6 | 0 | 0 |
| Lists of exactly three | 8 | 6 | 0 |
| Counted contrast constructions | 5 | 4 | 0 |
| Manufactured aphorisms | 2 | 2 | 0 |
"Stock vocabulary" and "manufactured aphorisms" are judgement calls made with a written rubric. San Digital has kept the drafts, prompts and rubric, but they are not published here. Another team cannot reproduce this result from the page alone.
Draft C still used two contrasts that carried useful content. We counted one repeated contrast construction, not every contrast in the prose. Its exact form is set out in the rubric, which is not published here.
Words
The named style rules removed the em dashes and stock terms in Draft B. They did much less to the structure. Draft B was shorter, so its three-item lists were slightly more common, not less. Draft B also retained four of the five counted contrast constructions.
Draft C reached zero on all five counted patterns after we named, limited and checked them. The shorter draft gave each pattern fewer chances to appear, and the checking pass removed those that remained.
Anyone trying to avoid these five patterns can put them in a checklist. Their presence may tell you something about the style of a passage. Their absence says very little about how it was made.
Limits
We did not run an AI-text detector on the drafts. Nor did we ask readers to guess which draft came from a model. The table cannot show that a detector was evaded or that a person would misidentify the text.
The table measures those five changes, nothing more. Anyone using the features to judge a passage should remember how easily they can be edited, and that ordinary human writing can contain them too.
Detectors
In 2023, a study ran 91 TOEFL essays by non-native English speakers through seven widely used detectors. The average false-positive rate came out at 61.22 per cent, on that English dataset and with those detector versions.
A 2026 Czech-language study tested three detector families and found no systematic bias against non-native speakers in its setting. The study also found that the contemporary detectors it tested did not depend on perplexity in the way the earlier explanation assumed.
The studies used different languages, writers, detector versions and designs, so neither result can be assumed to generalise beyond its setting. Neither paper gives a licence to treat a style score as proof about a person.
Anthropic's published watermark is a different kind of system. Its 14 August technical post describes a keyed SynthID-Text method based on word choices. The five visible patterns counted here are not its detector features, and editing them does not establish that a watermark changed.
Record
Need to know how an article was made? The questions are ordinary. What was requested, and what came back. Which claims got checked. What a person changed, and who reviewed the change. Write the answers down while the work is happening, not afterwards.
The prepared rewrite on this site uses one fixed passage. The rewrite lowers its score from 2.99 to -0.04, below the 1.80 cut-off. The result does not show that every rewrite defeats every watermark.