Start by cleaning the punctuation and stripping hidden characters, then cut the stock words and filler transitions. Vary your sentence lengths, flatten over-built lists, and add at least one specific detail per section.
To do it fast, paste your draft into a free AI humanizer tool that handles the punctuation, hidden characters, and giveaway words in one pass, then finish the voice edits by hand.
Common giveaway words include delve, robust, tapestry, leverage, seamless, comprehensive, pivotal, meticulous, and elevate. Stock phrases count too: "in conclusion," "furthermore," "moreover," and "it's important to note."
No single word is proof. A cluster of them in one short piece is the real signal, since a person writing plainly rarely uses that many high-register words at once.
Yes. Detectors weigh many of the same patterns covered here, especially low burstiness, meaning uniform sentence length, and predictable word choices. Some also scan for invisible characters and formatting quirks that come straight from model output.
That's why fixing the surface tells and adding real variety helps on both fronts: it reads better to people and scores lower as machine text.
You can get very close. Strip the hidden characters, fix the punctuation, swap the stock words, vary your sentence length, and add real details, and most casual readers and many detectors will read it as human.
The one thing no tool fixes for you is substance. Genuine, specific, first-hand information is what makes writing read as truly human.
Not by themselves. Plenty of skilled writers use em dashes on purpose. What signals AI is the frequency: two or three per paragraph, used where a comma or period would work fine.
Treat heavy em dash use as one clue, not proof. Weigh it alongside the vocabulary, rhythm, and detail tells before you decide.
Look for the common chatgpt tells stacked together. Em dashes in most paragraphs, curly quotes, phrases like "it's not just X, it's Y," and rule-of-three lists like "clear, concise, and correct."
Add the habit of restating your question before answering, plus a tidy summary at the end, and you're almost certainly looking at model output rather than a human draft.
The biggest tell is text that says a lot without any specifics. AI writes in generalities because it wasn't there, so you get confident prose with no names, no real numbers, and no small concrete moment a person would actually remember.
Punctuation and word choice are easier to spot, but missing detail is the one that's hardest for a model to fake.
The clearest signs are punctuation habits like heavy em dash use and curly quotes, stock vocabulary like "delve" and "robust," filler transitions like "in conclusion" and "furthermore," and a flat sentence rhythm where every line runs about the same length.
Structure gives it away too: even paragraphs, over-built lists, and empty intros. The strongest tell is a total lack of specific, first-hand detail.