How do you actually get AI to write like a human? It's the question more managers and executives are asking themselves as AI tools take over more of their daily writing — and as readers get faster and faster at spotting the machine.

Last month, Leadership IQ launched a free AI writing detector at LeadershipIQ.com, and people immediately started pasting in their own work. Emails to their bosses. Performance reviews. Policy memos. Cold outreach. LinkedIn drafts. The kind of writing that fills a normal workweek.

The submissions added up to 27,645 sentences of real workplace writing, with each one scored as either AI-generated, human-written, or somewhere in between. We then went back and analyzed the patterns. What separates the writing that reads as authentically human from the writing that gets quietly flagged as AI? The answer is not what most leaders assume.

Most people think AI writing gets caught because of vocabulary — the "delve" and "leverage" and "tapestry" words that have become punchlines online. Those words do show up. But they're a small part of the story, and they're the easy part to fix. The real tells are structural, and they're hiding in plain sight in writing that looks perfectly polished.

The most striking pattern in the data was not a word, a phrase, or a punctuation mark. It was the length of sentences and how much that length varied.

In writing flagged as human-generated, 27% of sentences were five words or shorter. In writing flagged as AI-generated, only 4% were that short. Humans use fragments. They write "Thoughts?" and "Happy Friday." and "Not sure yet." AI almost never does.

On the other end, AI-flagged writing concentrated heavily in the 21-to-35-word range. About 40% of AI sentences fell in that band, compared to just 13% of human sentences. AI cruises at a steady, professional sentence length and rarely breaks rhythm. Humans break rhythm constantly.

The technical term for this is burstiness. Human writing bursts. It alternates between very short and longer sentences in unpredictable patterns. AI writing flows at an even pace, and that even pace is the giveaway. A reader cannot articulate why a paragraph feels machine-written, but their brain registers the missing variation immediately.

Across 27,645 sentences of AI-flagged writing, exclamation points appeared zero times. Not rarely. Zero. Human-flagged writing used them at a rate of 3.6 per 1,000 words.

Question marks showed a similar gap. Human writing used 14 times more question marks than AI writing. Parentheses — the ones humans use for asides and qualifications — appeared nearly three times more often in human writing.

The one punctuation mark AI used more often was the em dash, at a 17% higher rate than humans. There's a reason it picked up the nickname "the ChatGPT hyphen."

The pattern is consistent. AI uses punctuation to organize ideas. Humans use punctuation to express how they feel about the ideas. A draft full of commas and em dashes but no exclamation points, no questions, and no parenthetical asides will read as machine-generated even if every word is correct.

Sentences that open with "It," "These," "In," and "This" are dramatically over-represented in AI writing. These are abstract pronoun openers, the kind of starts that announce a thesis without identifying who is doing what.

Sentences that open with "So," "Let," "Hi," "See," "You," "And," and "I" dominate human writing. These are conversational openers, the kind of starts a person uses when they are talking to another person.

The most striking single number: human writers opened sentences with "I" more than twice as often as AI writers did. The first-person voice was nearly absent from AI-flagged writing. AI defaults to a kind of disembodied perspective where things happen and findings emerge and considerations apply, but no specific person is doing anything.

The sentences that scored as most clearly human shared a small set of traits.

They contained fragments. They had typos. They named specific things using specific identifiers like product codes, dollar amounts, dates, and proper nouns. They expressed emotion directly rather than through implication. They used contractions liberally. They sometimes interrupted themselves mid-sentence.

In other words, the writing that read as most human was the writing that was least concerned with sounding professional. The professional polish AI delivers automatically is exactly what's flagging it.

AI is going to keep getting more useful for drafting. The question is not whether to use it. The question is whether what you send under your name still sounds like you.

These four findings are the clearest patterns in the data, but they are not the only ones. The full study identified more than 30 distinct structural and linguistic markers that separate AI-flagged writing from human-flagged writing, along with a system of prompts and editing passes that strips those markers out before they reach the page.

Your readers may not be able to explain why a paragraph feels machine-generated. But they can tell. And once they form the impression that you're sending them AI-written content, every email after that gets read with the same skepticism. The credibility cost is quiet but real.

The fix isn't more effort. It's different effort.