The em dash is not the AI tell people think it is
Four things people treat as proof that AI wrote something. Only one of them holds up.
1. The long dash. This is the famous one, sometimes called the ChatGPT hyphen, and it has cost people marks on real assignments. It is also the weakest signal on the list. Writers have used that punctuation for two centuries, and current models produce it far less than they used to. Removing it changes how people perceive your text, but it says almost nothing about who wrote it.
2. Individual words. Delve, leverage, tapestry, testament. These are worth cutting because plainer alternatives read better, not because any single one is evidence. People used all of them before chatbots existed. What matters is how many appear across a full page, not whether one shows up.
3. Clean grammar and no typos. A careful writer and a language model look identical on this measure. Typos prove tiredness, not humanity.
4. Rhythm. This is the one that holds. The Economist ran a study comparing its own articles against output from several major models, alongside writing from other newspapers and novels published between 1950 and 2022. The comparison covered 55,940 sentences and 1.2 million words. The popular giveaways turned out to be largely red herrings. What separated the machine text was structural: sparse punctuation, and sentences landing at a very consistent length one after another.
Human writing is uneven. A long winding sentence because the thought was long, then four words because the point was blunt. That unevenness is the signal that someone was thinking while writing.
How to check your own draft in a minute: take five sentences in a row and count the words in each. If the counts cluster tightly, the rhythm is the problem, and no amount of swapping punctuation or vocabulary will fix it. The sentences themselves have to be rebuilt.
When trying to identify AI-written content, many people resort to examining punctuation or vocabulary. However, through my experience editing various texts, I've found that these indicators are not reliable on their own. The so-called “ChatGPT hyphen” — the em dash — is a classic example. Writers have used this punctuation for centuries, and modern language models rarely overuse it as some fear. What really caught my attention is the rhythm of sentences. As supported by a detailed study analyzing 1.2 million words, machine-generated text often has a uniform sentence length pattern. This lack of variability contrasts sharply with human writing, where sentences naturally vary: some long and winding to unpack complex ideas, others short and punchy for emphasis. A simple yet effective test I use is reading five consecutive sentences and counting the words in each. If word counts are too similar, the text might have been generated by AI. This rhythm check highlights the structural differences machines struggle to replicate. Also, focusing on individual “fancy” words like ‘delve’ or ‘leverage’ as AI markers can be misleading, as they are common in human writing too. Similarly, perfect grammar or absence of typos is not exclusive to AI — many careful human writers produce clean text, while typos often reflect human fatigue rather than authenticity. In practice, improving writing that feels “robotic” is about varying sentence lengths and punctuation to mimic natural human thought flow. Tools exist that help rewrite such texts to add human tone and rhythm, making content more engaging and believable. Understanding these nuances not only helps in detecting AI content but also benefits writers wanting their work to sound more authentic and human.