The em-dash panic was a distraction. Here’s what actually signals it.
TextTE
If you learned to spot AI writing by hunting for em-dashes, you learned a superstition. The em-dash has been in use since the late sixteenth century. Dickens used it. Woolf used it. Newspaper editors have used it for generations. The punctuation, on its own, says nothing about whether a human or a machine wrote the sentence around it. In late 2025, OpenAI finally made ChatGPT actually obey custom instructions telling it to stop using em-dashes, a tacit admission that the mark had become a witch-hunt target rather than a real signal. Yet sometime in early 2025 the internet decided the em-dash was a confession, and writers started deleting a perfectly good piece of punctuation from their own work to avoid suspicion.
That panic is a useful lesson in how detection goes wrong. You fixate on a surface feature, the surface feature is cheap to fake or innocent to begin with, and you spend your attention in the wrong place. So let's be honest about which tells still mean something in 2026, which ones never did, and where the durable signal actually lives.
Large language models do have habits. Researchers who combed through 15 million PubMed abstracts found that after ChatGPT arrived, certain words spiked hard: "delve" showed up nearly thirty times more often than the pre-ChatGPT trend would predict by 2024, with "underscore" and "intricate" climbing alongside it. The same analysis also put a floor on machine involvement: at least 13 percent of 2024 biomedical abstracts looked like they'd been processed with a model in some way. A research group at Florida State University traced the habit back to the human-feedback training that polishes these models: the system learns to prefer the words that pleased its raters, and the preference-trained model was measurably less "surprised" by buzzword-laden text than the base model.
So yes, there is a recognizable register: the "moreover" and "in conclusion" scaffolding, the "rich tapestry," the dutiful "it's important to note," the reflexive "testament to." If a short paragraph contains three of these at once, your suspicion is reasonable.
But treat these as a smell, not a verdict, for two reasons.
First, the tells decay the moment they go public. After "delve" became a punchline in early 2024, its frequency in new academic writing dropped: authors noticed, and started editing it out or steering the model away from it. A tell that everyone knows is a tell the careful prompter erases in one extra sentence of instruction.
Second, surface vocabulary is the easiest thing in the world to change. "Write this without buzzwords, in plain language, the way a tired person actually talks" is a one-line prompt. The output that follows will pass every word-list test you own. By mid-2026, anyone trying to deceive you has already done this. The tics catch the lazy, not the deliberate.
Below the vocabulary, the prose has a shape, and the shape is harder to disguise.
Over-hedging and false balance. Default model output is congenitally diplomatic. It presents "on one hand / on the other hand," refuses to commit, and lands every argument in the safe middle. Real opinion writing takes a side and accepts the cost. A piece that carefully balances every claim against its opposite, and risks nothing, often has no human behind the wheel.
Listy scaffolding. Models love to organize. Three neat points, each with a bolded lead-in, each the same length, each ending on a tidy summarizing clause. Human argument is lumpier: one point sprawls, another gets a single sentence, a third is just a good aside.
Uniform rhythm. Read a suspect passage aloud. Machine prose tends toward sentences of similar length and similar grammatical build, a kind of metronomic evenness. Human writing varies its pace: a long winding sentence, then a short one. Then a fragment.
Generic transitions. "Furthermore." "Additionally." "Ultimately." Connective tissue that could bolt onto any topic is a sign the connections aren't load-bearing.
These are sturdier than vocabulary because they survive a casual cleanup. But they too can be prompted away, and, importantly, plenty of competent humans write in lists and hedge their bets. None of these is proof. Each one only shortens the odds.
Here is the tell that survives good prompting, because it is an epistemic limit, not a stylistic habit.
A model can generalize fluently about anything and know almost nothing in particular. It will tell you a neighborhood is "vibrant" and "full of character." It will not tell you that the corner store stopped carrying the good bread in March, or name the specific bus that's always late, or get the texture of a thing it supposedly witnessed. Real writing is studded with details that are checkable and non-obvious: the kind a person accumulates by being somewhere, doing something, getting something wrong and remembering it.
When you read something and your gut says "machine," it's usually this: confident, well-formed, and strangely weightless. Lots of claims, no evidence of contact with the world. The fix isn't to scan for "delve." It's to ask: does this writer appear to know anything that couldn't be guessed from a summary? Can I verify one concrete thing they assert? Machine text tends to evaporate under that question. Lived writing gets more solid.
This is the same move that settles AI images: provenance and specifics beat surface forensics. (We make that case in detail for pictures in how to spot AI images.)
The obvious shortcut (paste it into a detector) is the worst option on the table, and it's worth being blunt about why.
OpenAI built a classifier to flag AI-written text and quietly shut it down in July 2023, citing a low rate of accuracy; by its own measure it correctly caught AI text only about a quarter of the time. That's the company that makes the models telling you it couldn't reliably detect them.
Worse, the failures aren't random. A Stanford study found that detectors flagged the majority of TOEFL essays (written by non-native English speakers) as AI-generated, with an average false-positive rate above 60 percent; all seven detectors tested unanimously misclassified about a fifth of those essays. The mechanism is ugly: detectors lean on "perplexity," and writing with a smaller, simpler vocabulary reads as more predictable, so it gets punished. The tools systematically accuse people whose English is plainer. When the same essays were rewritten with richer vocabulary, that false-positive rate collapsed by nearly fifty points, proof that the detectors were grading word choice, not authorship. Run one on a real student's work and you may ruin their week over a number that means nothing.
We go deeper on the math and the failure modes in why AI detectors fail. The short version: a confident percentage from a detector is a guess wearing a lab coat, not evidence.
Drop the punctuation forensics. Notice the register (the buzzword clusters, the relentless balance, the even rhythm) and let it raise your eyebrow, not your gavel. Then do the thing that scales: look for the specificity gap, and try to verify one concrete claim. Provenance and detail are the durable signal. Everything else is a tell with a short shelf life. For the broader habit of staying oriented as this stuff improves, see living with AI media.
If you want to practice the instinct on something lower-stakes than a colleague's report, that's roughly what the daily puzzle is for.
Does using an em-dash mean something was written by AI?
No. The em-dash predates ChatGPT by roughly four and a half centuries: Dickens used it, Woolf used it, newspaper editors have used it for generations. Plenty of careful human writers lean on it, and OpenAI's GPT-5.1 update in late 2025 finally made ChatGPT's own custom instructions against em-dashes work reliably after months of the model ignoring them, a tacit admission that the punctuation says nothing about who, or what, did the writing. Judging text by its punctuation is how you end up accusing a novelist of being a chatbot.
Should I just run suspicious text through an AI detector?
Don't trust the result. OpenAI shut down its own classifier in 2023 for a low rate of accuracy, and a Stanford study found detectors flagged the majority of essays by non-native English writers as AI-generated. A detector can point you somewhere, but it cannot settle anything.