We wrote the same article three ways
12 August 2026
Some people believe they can recognise machine writing by style alone: em dashes, neat lists of three, words such as "delve". In our collection harness, one Claude model assignment wrote the same article three times under different style rules. We counted five visible patterns. Several disappeared when the prompt banned them; others needed a checking pass. This did not test a detector or ask readers to guess who wrote the drafts.
How the drafts differed
Draft A had no style guidance. Draft B named a person's voice and banned dash characters and stock phrases. Draft C kept those rules, added limits for several sentence structures, then checked the finished text against the list. Drafts A and B had no checking pass.
The topic and argument stayed the same. The drafts differ in length, so these are not like-for-like rates.
| 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 |
| "It is not X, it is Y" 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 the repeated "not X, but Y" shape, not every contrast in the prose.
The words changed before the sentence shapes
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. It 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.
Let's be honest: 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.
What the counts do not show
We did not run an AI-text detector on the drafts. We did not ask readers to guess which draft came from a model. The table therefore cannot show that a detector was evaded or that a person would misidentify the text.
The instructions changed five familiar features. That is all the table establishes. Anyone using those features to judge a passage should remember how easily they can be edited, and that ordinary human writing can contain them too.
Detector results change with the test
A 2023 study tested seven widely used detectors on 91 TOEFL essays written by non-native English speakers. On that English dataset, with those detector versions, the average false-positive rate was 61.22 per cent.
A 2026 Czech-language study tested three detector families and found no systematic bias against non-native speakers in its setting. It 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.
Ask how the article was made
If you need to know how an article was made, ask what was requested, what came back, which claims were checked, what a person changed and who reviewed it. Record those answers while the work happens.
The prepared rewrite on this site is deliberately narrow. For its fixed passage, the rewrite lowers the score from 2.99 to -0.04, below the 1.80 cut-off. It does not show that every rewrite defeats every watermark.
Sources: draft texts, prompts and full pattern counts held by San Digital, measured 12 August 2026; GPT detectors are biased against non-native English writers (Liang and others, 2023); Different Time, Different Language (Al Ali and others, 2026).
Published by San Digital Limited, independently of Anthropic. Do not use this article or any score on this site to make authorship, employment, education, disciplinary, legal or forensic decisions.