ChatGPT punctuation: the giveaways, the checker and the fixes

Every character below appears more often in model output than in text someone typed by hand. The notes say how much weight each one deserves, because the difference between "interesting" and "proof" is the entire subject.

• —Em dashU+2014The headline tell, and the most misleading one. Models use it as an all-purpose connector, but plenty of humans type it deliberately. Overuse is the signal, not the character.
• –En dashU+2013Rarely typed by hand outside publishing. In an AI draft it usually arrives where a hyphen or the word "to" belongs.
• —Spaced em dashU+2014The British-style " — " with spaces both sides is what models emit most often, because it reads cleanly regardless of the surrounding style.
• --Double hyphenU+002D U+002DThe ASCII fallback. Common in older human writing too, so treat it as a weak signal on its own.
• …Ellipsis characterU+2026The single-glyph ellipsis is more common in model output than the three-dot version, which most keyboards produce.
• “ ”Curly quotesU+201C U+201DSmart quotes are correct typography. On their own they mean nothing; in bulk they suggest text that was generated, not typed.
• ’Curly apostropheU+2019Same reasoning as curly quotes. Strongly depends on where the text was pasted from.
• §Section signU+00A7Legal-writing tell. Models use it far more than the people drafting contracts.
• →ArrowU+2192Very common in model formatting for before/after and cause/effect, uncommon in ordinary prose.
• •Bullet characterU+2022Not a tell by itself. A bullet anywhere a sentence would do is a structural tell.
• ✓Check markU+2713Appears in model-generated summaries and checklists far more than in typing.
• ‑Non-breaking hyphenU+2011Copied out of PDFs and generated documents. Invisible difference, breaks search and sort.

Across repeated sampling of ordinary chat replies, the ranking is stable:

1. Spaced em dash ( — ) — by far the most frequent. It is unambiguous in plain text, so models default to it.
2. Unspaced em dash (—) — less common, and more likely in "polished" rewrites than in first replies.
3. En dash (–) — used incorrectly more often than correctly, usually where a hyphen or the word "to" belongs.
4. Typed double hyphen (--) — appears when the model is told to avoid special characters, or when it is writing into a plain-text field.
5. Ellipsis character (…) — the single-glyph version, rather than three full stops.

The order is the useful part. If you are going to fix one thing, fix the spaced em dash, because it accounts for the majority of what readers notice.

The reason punctuation is over-trusted is that it is visible while the real signals are statistical. Sentence-length variance, lexical diversity and the presence of a specific fact are much harder to fake and much harder to see. Punctuation is just what shows up in a screenshot.

Rather than eyeballing it, count. The checker below reports every punctuation tell alongside word frequency and sentence-length uniformity, so you can see whether one dash is doing all the work or the whole passage is flat.

Once you know what you are dealing with, the cleanup is mechanical. Each of these takes the same input and different output: