Steno / The tells

Twelve tells that give away AI writing

People rarely run your work through a detector. They read two paragraphs and decide. These are the twelve patterns that make them decide wrong, and what to write instead.

Updated 6 August 2026 · About nine minutes

Most advice on this topic is a list of banned words. Ban delve, ban tapestry, ban the em dash, and you are done. That advice fails within a month, because the words move and the habits do not.

What follows is organised by habit. Each one is a thing a language model does when it has been asked to sound authoritative and has nothing specific to say. Fix the habit and the vocabulary sorts itself out.

1. Inflated significance

The model reaches for importance it has not earned. A tool becomes a paradigm shift. A change becomes a watershed moment. Nothing in the sentence supports the size of the claim.

The test: strike the adjective and see whether the sentence loses information or only volume. If only volume goes, it was inflation.

2. Participial padding

A clause gets bolted to the end of a sentence with a present participle. Highlighting, ensuring, fostering, showcasing, underscoring, reflecting. The clause almost never adds a fact. It restates the first half in warmer language.

If you want the reader to draw a conclusion, give them the evidence and let them draw it. Stating the conclusion for them is what makes the prose feel like a press release.

3. Promotional tone in neutral contexts

Travel copy leaks everywhere. A town is nestled in a valley. A restaurant boasts a wine list. A neighbourhood is vibrant. A museum is a must visit. This is the register of a brochure, and it turns up in text that is supposed to be describing rather than selling.

4. Vague attribution

Experts argue. Studies suggest. Reports indicate. It is widely regarded as. Attribution with nobody attached is worse than no attribution, because it borrows authority the text has not got. A reader who wants to check cannot.

Either name the source or drop the claim. If the sentence still works without the authority, it did not need it.

5. The rule of three

Ideas arrive in triplets whether or not there are three of them. Two real items get a third invented for rhythm. Once you notice it you cannot stop noticing it, and it is one of the fastest ways a reader clocks machine prose.

Count your lists. If a group of three has one weak member, cut it and keep two. Two is a perfectly good number.

6. Negative parallelism

Not only X but also Y. It is not merely a tool, it is a philosophy. This is not about speed, it is about trust. The structure creates the shape of an insight without supplying one, and models reach for it constantly at the end of sections.

The fix is to say the second half and delete the first. The contrast was decoration.

7. Uniform rhythm

This is the loudest tell and the least discussed. Sentence after sentence lands in the same length band. Paragraph after paragraph runs to the same three or four lines. Human writing is lumpy. It has a nine word sentence next to a forty word one, a one line paragraph after a long one, because the writer got bored or excited or ran out of things to say.

An analysis published by The Economist in 2026 compared its own reporting with versions produced by ChatGPT, Claude, Gemini and Grok, across roughly 1.2 million words. Verbosity and flat punctuation came out as clearer signals than any single word. Rhythm is the structural version of the same problem.

You cannot fix this by finding and replacing. Read your draft out loud. Where your breath falls into a pattern, break something.

8. Manufactured drama

Runs of short fragments engineered to land. And that changes everything. Here is the thing. But there is a catch. A single one of these is fine. Three in a page reads as a LinkedIn post, and LinkedIn posts are now the most obviously machine assisted writing on the internet.

9. Signposting

Let us dive in. In this article, we will explore. Here is what you need to know. Before we begin, it is worth understanding. None of it carries information. It is throat clearing, and it exists because the model is filling the space where a first sentence should be.

Start with the thing itself. The reader clicked, so they already agreed to the topic.

10. Filler and stacked hedges

In order to is to. Due to the fact that is because. It is worth noting that is nothing at all. Hedges stack until a sentence commits to no position: this could potentially help to somewhat reduce.

One hedge is honest. Three is a refusal to say anything.

11. Copula avoidance

Models avoid is and has as though the words were beneath them. So a thing serves as, represents, constitutes, features, embodies, provides. Almost every one of these can go back to is or has with no loss.

12. Cosmetic tells

Emoji as section markers. Boldface applied by rule rather than by meaning, usually the first two or three words of every bullet. Headings in title case where the rest of the document is in sentence case. Bullets used for prose that is not a list.

The em dash belongs in this category, and it belongs there with an asterisk. It was the famous giveaway for about two years. It is now unreliable, and the reasons are worth reading separately, because a lot of people are still editing for a signal that stopped working.

Using this list

Two ways. As a checklist while you draft, which is slow but produces clean text the first time. Or as an editing pass over something already written, which is faster and is what most people actually do.

The rule that matters in the editing pass: no claim gets dropped and nothing new gets invented. Text should get shorter because the ornament goes, not because information does. If a rewrite loses a fact, the rewrite is wrong, however much better it reads.