Skip to main content
AI Watermark Removal

Community

Do Em Dashes Mean AI Wrote It? What the Data Shows

No. In the one cross-chatbot test this page found, six chatbots given identical prompts produced anywhere from zero to nine em dashes each, which is not what a reliable signature looks like. The rule traces to a single viral social post from February 2025, not to any provider's disclosure, and it resurfaced most recently the same week news of Claude's actual, very different, invisible watermark spread.

By ignacio-sjPublished Revised Community discussion

Correction,

The page said Claude got a real text watermark in August 2026 and that it shipped then, and one answer said the exact technique had not been published. Anthropic announced and documented the watermark in August, but no Claude model was marked until Fable 5.1 and Mythos 5.1 shipped on September 1, 2026, with older models added from September 14. And Anthropic's explainer of August 14, 2026 describes the mechanism as a version of the SynthID-Text approach, keyed word selection with nothing added to the text. The section and the answer now say so.

All corrections

Key takeaways

  • Bailey's June 2025 test used identical prompts across six chatbots and found wildly inconsistent counts: ChatGPT and Copilot 8, DeepSeek 9, Claude 2, Gemini and Meta.ai 0. His conclusion was that no single stylistic tell reliably separates AI from human text, and that factual errors proved more diagnostic than punctuation habits.
  • The idea traces to one viral X post from February 11, 2025, not to any AI provider's statement. It reached 3.7 million views and 11,000 likes according to Daily Dot's reporting, and it stays classified here as rumor and speculation, not a confirmed mechanism.
  • OpenAI's real November 14, 2025 change let users tell ChatGPT to avoid em dashes through custom instructions, which CEO Sam Altman called a "small-but-happy win." That's a style preference toggle, not the removal of an embedded mark, since there was never one to remove.
  • When news of a real invisible text watermark for Claude spread in August 2026, the old joke resurfaced within hours, including a reply from X user @Adrienne asking whether the news was "just em dashes and calling everything load-bearing." Anthropic has never named em dashes, or any punctuation habit, as part of the mechanism.
  • "AI signature" covers at least four unrelated things: a signed C2PA manifest (verifiable, but strippable), an invisible SynthID-style watermark (unreadable without the provider's detector, but it survives metadata stripping), a writing habit like the em dash, and a commercial detector's probabilistic verdict. Only the first two were put there on purpose.

The one real cross-model test

Em dashes per chatbot, identical prompts

DeepSeek
9
ChatGPT
8
Copilot
8
Claude
2
Gemini
0
Meta.ai
0
What this actually shows

Bailey's June 2025 test, identical prompts across six chatbots. Zero to nine em dashes per model is not what a reliable signature looks like. No provider has ever named em dashes, or any punctuation habit, as part of a real watermarking mechanism.

Four unrelated mechanisms, one loose label

What "AI signature" actually means, case by case

MechanismReadable?Trust modelRemovable?
C2PA manifestReadableCryptographically verifiableStrippable
SynthID-style watermarkInvisibleProbabilisticPersistent
Writing tic (em dash)ReadableUnreliableCoincidental
AI detector verdictN/A (a score)ProbabilisticN/A
Why the mix-up produces confident wrong conclusions

These four differ on readable versus invisible, cryptographically verifiable versus probabilistic, and strippable versus persistent. Treating any two as synonyms produces confident wrong conclusions in both directions.

Where did the em dash AI rule come from?

Rumor/speculation

One X post from February 11, 2025, 3.7 million views, and a paper trail so thin every citation of the wording routes through a single article.

On February 11, 2025, X user @PaulSkallas, who posts as LindyMan, wrote: "A shortcut for detecting if something is written with AI is they all use this symbol '—' throughout the writing... It's relatively rare when a human uses it... But AI chats love using it."

It reached 3.7 million views and 11,000 likes, according to Daily Dot's reporting. Writers and editors pushed back immediately in the same thread, pointing out that plenty of humans use the mark deliberately and often.

No AI provider was involved at any stage. This is a pattern one person noticed and posted about, not a technical claim any lab has ever confirmed.

What did the cross-chatbot em dash test find?

Confirmed

Here's the actual data, including the half of it that nobody quotes.

Six chatbots given identical prompts produced anywhere from zero to nine em dashes each. Jonathan Bailey, a plagiarism and copyright analyst, kept his June 2025 test simple: identical prompts, six chatbots, count the em dashes in each response. The spread was enormous.

  • ChatGPT: 8
  • Microsoft Copilot: 8
  • DeepSeek: 9
  • Claude: 2
  • Gemini: 0
  • Meta.ai: 0

That range makes the rule unworkable even when comparing AI systems against each other, before you get anywhere near human writers, who have always used the punctuation mark on purpose.

Bailey's own conclusion went further. No single stylistic tell reliably separates AI from human text, and factual errors proved more diagnostic than punctuation habits. This site's Lab runs the same kind of counting test on Claude's output, looking for invisible characters instead of em dashes.

What did OpenAI actually fix about ChatGPT's em dashes?

Confirmed

November 14, 2025 is when ChatGPT started honoring requests to drop the em dash. A style toggle, not a hidden marker being switched off.

OpenAI CEO Sam Altman announced in November 2025 that ChatGPT would finally honor custom instructions asking it to stop using em dashes, calling it a "small-but-happy win."

That's a personalization setting a user turns on, not OpenAI removing an embedded watermark, because there was never a discrete mechanism to remove. Whether ChatGPT text carries a real watermark is a separate question with its own page.

Em dash frequency is a side effect of training data and fine-tuning. A real watermark is inserted deliberately and can be detected deliberately; a stylistic habit is neither. Unlike a real watermark, an em dash habit is also defeated by find-and-replace.

Why did the em dash joke resurface in August 2026?

Community discussion

An eighteen-month-old joke resurfaced within hours of the news of Claude's real watermark. Same news cycle, and not the same kind of claim at all.

In August 2026, Anthropic documented a real invisible text watermark for Claude, a development with nothing to do with any single punctuation mark. Anthropic's support article describes weaving "an imperceptible watermark directly into the text itself", applied at launch to models released on or after August 2, 2026. The first of those, Fable 5.1 and Mythos 5.1, shipped on September 1, and older models are being added: Opus 5 from September 14, and Fable 5, Sonnet 5 and Opus 4.8 from September 30 per Anthropic's 2026-09-25 email.

It names no punctuation mark, no phrasing habit, and no specific character as the mechanism. Independent research into the rollout has turned up no primary source confirming em dashes, phrasing tics, or hidden Unicode as the method.

The news broke on X the evening of August 10, 2026, through independent AI-news accounts rather than any Anthropic announcement of its own. The largest of those posts, from @M1Astra, reached 1.3 million views.

Within hours, a reply on that same thread from X user @Adrienne asked: "Isn't this just em dashes and calling everything load-bearing?"

It's a funny line, and it shows the confusion exactly. A confirmed watermark a provider deliberately embedded and a punctuation habit a stranger noticed are not the same kind of claim, even when they land in the same news cycle.

Does Claude use em dashes more than human writers do?

Rumor/speculation

The Fast Company piece citing an unnamed Economist analysis could not be reached, so the claim about Claude's em dash rate stays unrepeated here.

One claim circulating alongside the meme could not be confirmed: a Fast Company piece reportedly citing an unnamed Economist analysis, saying Claude specifically uses em dashes more than a human baseline.

Every attempt to reach the underlying source failed, since fastcompany.com returned an access error to each method tried. Nobody has produced the Economist analysis itself.

Worth setting beside it: in the cross-chatbot test above, Claude came in second lowest at 2. That doesn't disprove the claim, which compares Claude to humans rather than to other models. This site's invisible-character census also publishes em dash rates for three Claude models from its original run, and finds them varying by more than 2x between tiers of the same provider.

What else do people mean by an AI signature?

The em dash is one of four things called an AI signature. Two were embedded on purpose and can be checked against a published mechanism; the other two cannot.

People use "AI signature" for four unrelated things. The phrase almost never means one specific thing.

  • A cryptographically signed provenance record, such as a C2PA Content Credential. This is a literal digital signature, in the same sense a signed PDF is one.
  • An invisible watermark such as Google DeepMind's SynthID. For text it biases which words a model picks; for images it embeds a signal in the pixels.
  • A stylistic habit, such as repeated words or the em dash. It is a side effect of training data and fine-tuning, not something a company planted.
  • A commercial detector's verdict, from tools like GPTZero or Originality.ai. That is a classifier's guess, not evidence that anything was inserted.

Two questions separate them: was anything deliberately embedded, and can it be checked against a known, published mechanism? A C2PA manifest and a SynthID watermark answer yes to the first. Style habits and detector verdicts answer no.

A C2PA Content Credential is hashed and signed with the producer's private key on X.509 certificate infrastructure, the system behind HTTPS. Any change to the file or its record after signing breaks the hash chain, which is how tampering gets caught.

Leica's M11-P, announced on October 26, 2023, was the first camera to sign photos with a Content Credential at the moment of capture. Anyone can check such a manifest at verify.contentauthenticity.org. A verified manifest shows what the signer claimed, not whether AI touched the file at an earlier, unsigned stage.

A SynthID watermark cannot be seen by opening a file, because nothing is stored as metadata. SynthID-Text biases the sampling a model uses to choose each next word, so a matching detector can later recover a statistical signal from the pattern of word choices.

Google made the SynthID-Text implementation available on Hugging Face in October 2024. Outside developers can run that detection code, but only with the configuration used to generate the text.

What goes wrong if you confuse the different AI signatures?

Mixing them up produces two opposite errors: dismissing real evidence and overstating weak evidence.

The first error is assuming a stripped C2PA manifest means there is no sign of AI. A SynthID-style watermark, if one was applied, stays in the pixels or words whatever happened to the metadata.

The second error is treating an em dash, or a positive score from a commercial detector, as though it carried the weight of a verified signature or a matched watermark. Neither is backed by a known, checkable mechanism. The difference between a watermark and a detector has its own page.

So before treating an "AI signature" as evidence, ask which of the four is meant. How much the claim is worth depends on the answer.

FAQ

Is heavy em dash use proof that ChatGPT or Claude wrote something?

No. Jonathan Bailey's June 2025 test found counts ranging from 0 to 9 across six models on identical prompts, and no provider has ever described em dash frequency as an intentional signal. It's a byproduct of training data, not a watermark.

Did an AI company ever confirm em dashes as a marker?

No. The idea traces to one person's viral social media post in February 2025, not to any provider disclosure, which is why it's classified here as rumor and speculation rather than a confirmed mechanism.

Why did OpenAI let users turn off em dashes if there's no mechanism?

Because plenty of users found the habit annoying regardless of whether it means anything. The November 2025 change was a style preference exposed through custom instructions, a UX response to complaints, not evidence of a hidden marker being switched off.

Does Claude's real text watermark use em dashes?

Nothing Anthropic has said supports that. Its explainer of August 14, 2026 describes the mechanism as a version of the SynthID-Text approach: a secret key plus the preceding words choose among candidate next words, and nothing is added to the text. It never names em dashes or any single punctuation mark, and its exact parameters are unpublished, so treat any claim tying the watermark to em dash frequency as unsupported.

Does a C2PA Content Credential prove whether something is AI-generated?

Not by itself. A C2PA manifest, the kind Leica's M11-P camera has signed at the point of capture since October 2023, records what the signing tool or device claims happened. Its signature proves that record has not been altered since signing. It does not show that no AI touched the file at an earlier, unsigned stage, and the manifest is lost entirely if a platform strips metadata on upload.

Can I see a SynthID watermark by opening the file, the way I can inspect C2PA metadata?

No. SynthID Text biases which words a model chooses during generation, and SynthID for images embeds a signal in the pixels, so there is no metadata field to open. Recovering the signal requires the provider's matching statistical detector.

If I strip the metadata from an AI-generated file, have I removed its AI signature?

It depends which sense you mean. Stripping metadata removes a C2PA Content Credential, since that signature lives entirely in the metadata layer. It does nothing to a SynthID-style watermark, which is embedded in the pixel or word values themselves. The two mechanisms fail in opposite ways.

Is a positive score from an AI detector the same as finding a watermark or signature?

No. A detector like GPTZero or Originality.ai outputs a probabilistic classification based on statistical or stylistic modeling of the text. It does not confirm that an intentionally embedded mark such as a C2PA manifest or a SynthID watermark was found, and those vendors' own sites do not mention SynthID, C2PA or watermarking at all.

Next steps

Sources and citation status