We measured 14,268 words of human writing for AI tells. Moby Dick scored highest.

What we ran, and how you can re-run it

Fifteen sources in total, chosen because their authorship is not in question, and sampled the same way each time: strip markup, discard the Project Gutenberg licence header, then take a 20,000-character slice and post it to our own public endpoint with rewriting disabled.

The endpoint needs no key and no account, so this is reproducible by anyone who wants to check us: `POST https://claudewatermark.xyz/api/process` with `{"text": "…", "rewrite": false}`.

The offset matters and we should have said so first time. The first run sliced from 2,000 characters in and recorded 87 em dashes in 3,357 words of Moby Dick. A second run from 3,000 characters in recorded 89 in 3,366. Both are correct — they are different windows of the same book — but if you re-run this and get 89, that is why, and it is not us adjusting a number. Every figure below is from the 3,000-character offset.

The results

Ten novels, em dashes per 1,000 words: Moby Dick 26.4 · Ulysses 25.0 · Alice in Wonderland 7.3 · The Picture of Dorian Gray 6.8 · Great Expectations 6.1 · Sherlock Holmes 3.9 · Frankenstein 2.2 · War and Peace 2.0 · Pride and Prejudice 0 · Dracula 0.

Every one of those returned zero deterministic artifacts — no HTML class names, no editor data-attributes, no zero-width characters.

Five non-fiction sources on top of the novels: RFC 2616 (the HTTP/1.1 specification, 1999) returned zero of everything, including zero stylistic hits, which is what a plain-ASCII technical document should look like. The Wikipedia articles on digital watermarking and typography returned zero deterministic artifacts and 1 and 4 stylistic hits respectively.

Running total: roughly 50,000 words of human writing spanning 1813 to 2026, and not one deterministic artifact.

The distribution is the part that matters more than any single number. The most-cited AI tell ranges from zero to twenty-six per thousand words across ten canonical authors. Austen and Stoker never use it. Melville and Joyce use it more heavily than most chatbots do. A signal with that spread has no baseline to accuse anyone against.

Why this matters if you have been accused

The em dash has become the most-cited giveaway of AI writing. Melville's 87 in 3,400 words is not a defence of any particular passage, and it does not prove your text was human. What it does show is that the signal people are treating as damning is one that a canonical human author produces at a rate no modern writer approaches.

The same applies to curly quotes and typographic punctuation. Austen scores 57. Any word processor with smart quotes enabled produces them by default, which is most word processors, by default.

These are the classes our own tool labels stylistic, and we label them that way because they are not evidence. A tool that scores you on them and returns a confident verdict is reporting your typography, not your authorship.

What the zero column actually means, and what it does not

The classes that returned zero are the ones that indicate the text passed through a chat interface: provider-specific HTML class names and editor data-attributes. Those are the useful signal, and they did not appear once in 14,268 words of human prose.

Two honest limits. Fifteen sources with one sample each is enough to show these classes do not fire on clean human writing; it is not enough to publish a false-positive rate, and we are not claiming one. And Project Gutenberg text is a favourable case, since decades of normalisation would have removed stray Unicode anyway — the harder tests were Wikipedia and the RFC, and those also returned zero.

We also went looking for the opposite result and found it. Non-breaking spaces are byte-exact, but they occur in ordinary web copy all the time because HTML ` ` is standard typography: measured on the same day, bbc.com/news carried 2, gov.uk 7 and smashingmagazine.com 8. So a non-breaking space is not evidence of a chat interface either, and our own command-line tool used to imply otherwise until this measurement made us fix it.

None of this is about Anthropic's watermark

Worth stating plainly, because the two get merged constantly. Everything above concerns copy-paste artifacts and stylistic tells, both of which anyone can check by reading the bytes.

Anthropic's statistical watermark is a different mechanism entirely, carried in which words the model chose. Nobody outside Anthropic can verify it, because detection requires a key that has not been released. We cannot check it, and neither can any tool currently selling you a check.

Questions

Does using em dashes mean my writing looks AI-generated?

It is the most-cited tell, and it is weak. Herman Melville used 87 em dashes in 3,357 words of Moby Dick in 1851, which is a higher rate than most AI output. Em dashes are a style, not evidence.

So what does actually indicate text came from an AI chat interface?

Provider-specific HTML class names and editor data-attributes, which survive copy-paste out of a chat UI and cannot be typed by accident. Those returned zero hits across 14,268 words of human writing in this test.

Can I reproduce these numbers?

Yes, and we would rather you did. The endpoint is public and needs no account: POST to https://claudewatermark.xyz/api/process with {"text": "…", "rewrite": false}. The sources are Project Gutenberg, the IETF RFC archive and Wikipedia.

Is this a false-positive rate?

No, and we are not presenting it as one. Five sources with one sample each demonstrates that these classes do not fire on clean human prose. A rate would need a far larger and more varied corpus.