Why the em-dash became the tell
Two things happened at once. Language models produce em-dashes at a much higher rate than most human writers, because their training data is full of edited, professionally typeset prose where the em-dash is standard. And almost nobody types one directly, because it is not on a keyboard. Getting one requires a shortcut, an autocorrect rule or a paste.
The result is a punctuation mark that is common in generated text and rare in typed text. Over 2025 and 2026 it became the single most cited giveaway, to the point where writers who genuinely like em-dashes started removing them defensively.
The same logic applies to curly quotation marks and apostrophes. A plain text box produces straight quotes. Curly ones come from a word processor or a generator. And the single-character ellipsis (…) rather than three typed periods is a third marker from the same family.
What this changes, precisely
An em-dash (—) with surrounding whitespace becomes a spaced hyphen. An en-dash (–) is treated the same way. Curly double quotes (“ ”) become straight ones ("). Curly single quotes and apostrophes (‘ ’) become straight ones ('). The ellipsis character (…) becomes three periods.
That is a visible change to your text, which makes it different from the invisible-character removal that runs alongside it. So it is a separate checkbox, on by default. The report tells you exactly how many of each it changed.
Nothing else about your wording is altered. This is a character substitution, not a rewrite.
The other patterns worth knowing about
Punctuation is only the most visible layer. The style report that runs with every check also flags the phrasing patterns that give text away, and those are harder to fix mechanically because fixing them means rewriting.
Stock phrases: "it's important to note", "in today's fast-paced world", "delve into", "a testament to", "navigating the complexities", "ever-evolving landscape". The report names every one it finds.
Rhythm: models produce sentences of unusually consistent length. Human drafts mix a three-word sentence with a forty-word one. The report measures the variation and flags it when it is too even.
Structure: bulleted lists where every item starts with a bolded term followed by a dash. Three-item lists everywhere. Both are model habits more than human ones.
The tool does not rewrite these for you. That would need a language model, and it would defeat the point of a tool that runs instantly and keeps nothing. It tells you where they are so you can decide.