What Actually Moves Perplexity and Burstiness — and What's Just Noise

Two numbers, one guess

Most AI detectors boil a document down to two measurements before they ever print a score. Perplexity asks how surprised a language model is by your word choices — low surprise, low perplexity, more "AI-like." Burstiness asks how much your sentence lengths and rhythms swing from one line to the next — flat and even reads as machine-made, jagged and uneven reads as human.

Neither number is measuring truth. Both are measuring resemblance to the output of models that were trained to pick the statistically likely next word. Once you know that, you can stop guessing at fixes and start looking at what actually pushes those two numbers around.

What moves the needle vs. what just feels like it should

Specificity is the cheapest lever you have

Perplexity climbs when your next word is hard to guess. Generic writing is easy to guess — "the company saw significant growth" is a sentence a model could have written in its sleep. "Revenue went from $2.1M to $3.4M in the back half of the year" is not, because a language model has no way to predict that exact figure from context alone.

This is why detail-heavy writing tends to score as more human than vague writing, independent of who actually wrote it. Names, dates, numbers, brand-specific details, and idiosyncratic examples all raise the surprise factor. It's also why a student's first draft full of hedge words and abstractions can flag even when a human typed every letter — the ideas themselves are predictable, not the typing.

Where specificity tends to hide in a draft

• Swap a general claim for the one data point that supports it
• Replace "experts say" with the actual source or study
• Trade a category ("a popular tool") for the actual name
• Add a concrete example instead of describing a type of example
• Let one sentence carry an opinion or judgment a model wouldn't default to

Building burstiness on purpose

1. 1
   Read your paragraph lengths out loud

   If every sentence takes about the same breath to say, that's the flat rhythm burstiness metrics are built to flag.

2. 2
   Cut one sentence down to three or four words

   A short, blunt sentence next to longer ones creates the variance detectors measure directly.

3. 3
   Let one sentence run long, with a subordinate clause or two

   Not padding — an actual layered thought, the kind that comes from explaining a real exception or caveat.

4. 4
   Re-check the mix

   You're not aiming for a formula. You're aiming for the kind of uneven rhythm that shows up when someone writes the way they'd actually talk through an idea.

The limit that matters most

None of this makes a detector score a verdict. It's a probability estimate built on patterns from training data, and formal human writing — a legal brief, a lab report, a dense academic paragraph — can land in the same low-perplexity, low-burstiness range as machine output simply because careful, formal prose is also predictable. The detector isn't lying in that case. It's just measuring a property that human writing and AI writing can both share.

That's the honest way to use these tools: as a directional signal on one piece of text, not a lie-detector test on a person. A single score tells you how one passage compares to typical model output right now, on one detector's model of "typical." It won't tell you who wrote it.