Signs of AI Writing: 12 Patterns With Reproducible Thresholds (2026)

TL;DR: One sign is a coincidence. Two is a hypothesis. This guide gives you twelve measurable patterns — each with a threshold you can reproduce on your own texts — so your verdict rests on counts, not vibes.

Why measure at all?

“Something feels off” is where most AI-spotting advice starts and ends. That is not good enough when you are editing your own draft or reviewing someone else’s work. Every pattern below can be counted with a word processor and two minutes of patience. The thresholds come from our own engine’s calibration work, not from folklore.

Vocabulary signs

1. Style-word pileups. Words like delve, tapestry, and underscore are not banned — but more than 3 Tier-1 style words per 1,000 words is a strong signal. Human frequency in our reference corpus: under 1 per 1,000.

2. Corporate verb inflation. Utilize, leverage, facilitate — when “utilize” outnumbers “use”, you are likely reading machine-polished prose. Humans default to the short verb.

Cliché signs

3. Empty openers. Phrases like “in today’s rapidly evolving world” are templates with the nouns swapped. More than one per page and the pattern flag goes up.

4. Fake authority. “Studies show” with no study, “experts agree” with no expert. Count claims with no verifiable source: 2+ per 500 words is the slop zone.

5. Pseudo-wisdom filler. “At the end of the day, it’s about the journey.” Aphorism density above 1 per 300 words correlates strongly with generated filler.

Structure signs

6. Em dash density in the wrong register. The em dash is not an AI tell by itself — journalists love it too. But above 1.2 em dashes per 100 words in plain business or blog prose, suspicion is fair. (We measured this properly in a separate piece.)

7. Rule-of-three on autopilot. Triads are rhetoric’s comfort food. More than 2 “X, Y, and Z” constructions per 500 words suggests template thinking.

8. Sentence-length uniformity. Compute the standard deviation of sentence lengths. Human prose usually lands above 5 words; generated prose often sits between 3 and 5 — every sentence the same size, like pegs on a rack.

Substance signs

9. Hedging without commitment. “Can help”, “many”, “various factors”. More than 4 hedges per 500 words with zero concrete numbers is the classic profile.

10. Missing specifics. No dates, no names, no figures. Count concrete nouns that could only appear in this article: 0–2 per page is slop territory.

11. Zero lived experience. Nothing in the text could only have happened to its author. No “I tried this in March and it failed”. Presence of a first-person failure story is a strong human signal.

12. The conclusion that concludes nothing. Final paragraph restates the intro with more adjectives. If you can swap the conclusion between two articles without anyone noticing, flag it.

Putting it together

Do not convict on one sign — humans overuse em dashes too. The engine in our detector fires when two or more independent factors cross their thresholds, and shows you every match so you can disagree with reasons. That is the whole point: measurable signs, not vibes.

Related reading

• The AI Words List: 120+ Phrases ChatGPT Overuses — the full dictionary behind the vocabulary thresholds above.
• Is the Em Dash an AI Tell? We Measured It — the data behind the punctuation pattern.
• The Slop Taxonomy: 5 Types of AI Slop — how these patterns cluster into recognizable content types.