Punctuation is the wrong place to look
12 August 2026
People often treat em dashes and curly quotation marks as evidence of machine writing. We counted both in collected Claude output and in five nineteenth-century novels. In these files, the novels used them far more often.
The count
We compared 457,045 characters from four Claude model assignments with 3,059,329 characters from Austen, Carroll, Melville, Twain and Shelley. The novels came from raw Project Gutenberg files. The Claude files were written straight to disk, without a chat interface changing their punctuation.
| Corpus | Characters | Curly quote and apostrophe characters | Em dashes |
|---|---|---|---|
| Collected Claude output | 457,045 | 0 | 37 |
| Five nineteenth-century novels | 3,059,329 | 20,224 | 3,055 |
Across 120 Claude files, we found no curly quotation marks or apostrophes. The files used straight marks instead. Em dashes appeared at about 0.08 per thousand characters. The novels contained about 6.6 curly marks and one em dash per thousand characters, so their em-dash rate was roughly twelve times higher.
This count describes these files. It does not establish a general punctuation rate for people or models.
The dialogue problem
The novels contain much more dialogue than the full prompt set, which also includes code, JSON, lists and technical prose. So we ran a second, rough comparison: the three dialogue prompts against all five novels, counting curly marks and em dashes together. The gap narrowed, but each model rate was still about 5.5 to 23.4 times lower. The genres are closer, not matched, so treat this as a check on the first count rather than a separate result.
The em dashes are concentrated. Thirty-six of the 37 occur in dialogue prompts and mark interrupted speech. One occurs in an informal post. None appears in the factual prose, lists, code or JSON. The punctuation followed the writing task.
The punctuation may have changed later
The Project Gutenberg files are modern edited transcriptions of older books. They contain typographic quotation marks, but the files alone do not tell us which marks came from the original edition and which came from later editorial work.
Modern text can change after drafting too. A word processor can curl quotation marks. A person, model, editor or style rule can add an em dash. Our Claude files did not pass through a word processor or publishing system before the scan. If punctuation changed later, it tells us about that edit. It does not identify who drafted the words.
Punctuation cannot establish authorship
An em dash or curly quote tells you only that a writer, editor or piece of software used it. It does not reconstruct how the text was made.
Anthropic says that even finding its own supported watermark is not conclusive. Punctuation has no published key, calibrated cut-off or stated error rate. A decision about a person should not rest on the tidiness of their typography.
Use a character scan for character questions
If you need to know which characters are present, inspect their Unicode code points. That will tell you whether an apostrophe is straight or curly, whether text contains a zero-width character and where it occurs.
A character inventory cannot tell you who wrote the words, what they checked or what changed between drafts. You need to record those facts as the work happens.
Read the full byte-level scan, or try the four watermark steps.
Sources: scan corpus, per-model counts and SHA-256 manifest at san-digital/claude-watermark-audit, generated 12 August 2026; human comparison files from Project Gutenberg; How Claude marks AI-generated content (Anthropic Help Centre, page last modified 10 August 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.