Module 1 · Lesson 1.2

Perplexity and Burstiness, Explained Simply

3 min read · Updated

If you read anything about AI detectors, two words come up again and again: perplexity and burstiness. They sound technical. The ideas behind them are simple, and once you get them you'll know what your edits are really changing.

Perplexity: how surprising your words are

In lesson 1.1 you saw that a language model gives every possible next word a probability. Perplexity turns that into one number for a whole passage. It asks: if a model read this text word by word, how surprised would it be?

  • Low perplexity means the model saw most words coming. The text follows the most expected path.
  • High perplexity means the text kept making choices the model didn't predict.

AI-generated text has low perplexity almost by definition, because it was written by picking likely words. Human writing tends to score higher. We use odd phrasings, private jokes, local names and the occasional word that's slightly wrong but feels right.

A quick comparison:

Low perplexity Regular exercise has many benefits for both physical and mental health.
Higher perplexity I started running to get away from my inbox, and somewhere around week six my knees stopped complaining.

The first sentence could finish itself after three words. The second keeps turning. "Inbox" and "knees" are not where a model expects a sentence about exercise to go.

Burstiness: how much your rhythm changes

Burstiness looks at variation. Human writers speed up and slow down. A long, winding sentence is followed by a short one. Then a fragment. Then something medium. AI text tends to hold a steady pace, with sentence after sentence of about the same length and shape.

Here's the same idea as a picture. Each bar is one sentence, and its height is the number of words in it.

Typical AI paragraph Typical human paragraph 30 words 30 words Lengths 19 to 23 words. Spread: 6% Lengths 4 to 31 words. Spread: 60%
The AI paragraph looks like a fence. The human one looks like a city skyline. Detectors notice the difference, and so do readers, even if they couldn't tell you why.

The AI detector on this site draws this chart for any text you paste in. It's one of the fastest ways to see why a paragraph feels monotonous.

How detectors use the two numbers

GPTZero popularized this pair in early 2023, and many tools still use some version of it. The logic goes like this:

Perplexity Burstiness What a detector usually concludes
Low Low Likely AI-written
Low High Mixed. Possibly AI text that was edited for rhythm
High Low Mixed. Possibly formal human writing
High High Likely human

Notice the problem in the middle rows. Plenty of real people write formal, even-paced prose: lawyers, scientists, students taught to follow a five-paragraph structure, and many people writing in their second language. Their writing can land in "likely AI" without any AI involved. Lesson 1.3 looks at where detectors fail in more detail.

What raises both numbers

You don't fix perplexity by stuffing in strange words. That reads badly and fools no one. The changes that work are the ones that make writing better anyway:

  1. Specific details. Names, numbers, places and things that happened are hard to predict. See lesson 3.3.
  2. Your own phrasing. The way you'd say it out loud, not the most common way to write it. The read-aloud test catches this.
  3. Varied sentence length. Put a short sentence after a long one. Lesson 3.2 is all about fixing sentence rhythm.
Try this

Take one paragraph you wrote recently and one an AI tool wrote. Count the words in each sentence and write the numbers down in a row. Which list has the bigger gap between its shortest and longest sentence? Then paste both into the AI detector and compare the rhythm charts.