Recognizing AI Tells in Writing

People can't always name it, but they feel it: a paragraph that sounds fine and says nothing. Smooth, balanced, a little grand. They stop trusting the writer about two sentences in. Once a commissioner or a neighbor suspects a machine wrote a public comment, every fact in it gets a discount, including the ones someone spent three weeks getting right.

The same ear works in the other direction. Form letters to a board, campaign mailers, a sudden flood of near-identical online comments: generated text has a sound, and recognizing it tells you something about who is behind a message and how much work went into it.

Wikipedia's editors keep a long public guide to signs of AI writing, built from thousands of cleanup cases. Two of its observations hold up best. The overused words shift as models change (it notes that "delve" was hugely overused in 2023 and early 2024 and then faded). And the tells are as much about structure as vocabulary. A list of words helps. An ear for the patterns helps more.

The words and phrases

These turn up again and again in machine-drafted text. One alone proves nothing; plenty of humans write "crucial." A cluster is a red flag. Every item below is an example of the pattern, not a recommendation.

Wikipedia's guide also lists overuse of em dashes as a tell. Plenty of fine human writers use them; this course avoids them entirely, which stays out of the argument.

The structures

Swap the words and the skeleton still shows. These are the shapes.

• False contrast. It's not just a library. It's a lifeline. Nobody said it was just a library.
• Habitual triads. The plan is bold, practical and forward-looking. Three adjectives, zero information. Wikipedia's guide calls out the rule of three directly.
• Announced counts. There are three reasons this matters. The reasons themselves would do.
• Rhetorical question openers. What does this mean for taxpayers? The answer is the content.
• Staged objection. Some might argue the cost is too high. But consider this. No name attached to who argues it.
• Hammer endings. Every section closing on a four-word zinger like And that matters.
• Restating close. A final paragraph that repeats the piece in miniature.
• Balanced to a fault. Every claim followed by "however" and an opposing view, even when the evidence is lopsided.

Generic versus specific

The specific version has a date, a vote, a dollar figure, a source, a person and a detail ("Nobody on the board answered her") that only someone in the room could know. It has no adjectives doing any judging. The facts do the work, and a board member can't wave it off as boilerplate.

The specific version takes a position, shows where the math came from and puts the writer in a car on a real road (well, an invented one). It also has a slightly ragged edge. Overpolished copy reads synthetic, and people sense it even when they can't say why.

Rhythm

Generated text drones. Sentence after sentence falls between 15 and 25 words. Human writing swings hard: a 6-word sentence, then a 38-word one that carries a complex idea all the way through with its clauses in order, then a fragment. Humans start sentences with And or But. Read a suspect paragraph aloud and the drone is obvious. The ear catches dead rhythm before the eye does.

What to look for