What "perplexity" actually means in AI detection

If you've ever looked at an AI detection report, you've probably seen the word "perplexity." It's the most technical of the detection signals and the least intuitive. Here's what it actually means and why it matters for your writing.

The simple version

Perplexity measures how predictable your word choices are. Given the words you've already written, how easy is it for a language model to guess what comes next?

If your text is highly predictable (each word is the statistically likely next choice), it has low perplexity (more likely to be flagged). If your text is surprising (unusual word choices, unexpected turns, personal idioms), it has high perplexity and detectors let it pass.

LLMs work by picking the most probable next word. Text generated this way has inherently low perplexity because it is the prediction. Human writing is less predictable because it's not optimized for statistical probability. Humans optimize for meaning, style, humor, emphasis, and a dozen other things that a language model doesn't model well.

How detectors use it

Most AI detectors run your text through a language model (usually a variant of GPT-2 or a similar architecture) and measure how "surprised" the model is by each word. Words the model predicted with high confidence contribute to low perplexity; words it didn't expect contribute to high perplexity.

The detector then compares your text's perplexity to two reference distributions: one for known AI text and one for known human text. If your text's perplexity is closer to the AI distribution, it gets flagged.

Sentence-level vs. document-level

Perplexity can be measured per sentence or across an entire document. Both matter, but they flag different things. A document might have normal overall perplexity but contain individual sentences with suspiciously low perplexity; these are the sentences that get highlighted in a GPTypo scan.

Conversely, a document with high overall perplexity but a few AI-generated sentences mixed in might pass at the document level but fail at the sentence level. This is why sentence-level analysis matters.

Why your writing might have low perplexity

You don't have to be using AI for your text to have low perplexity. Several natural writing patterns produce predictable text:

• Common phrases and idioms: "At the end of the day," "it goes without saying," and similar expressions are highly predictable. A sentence built from common phrases has low perplexity even if a human wrote it.
• Formulaic structure: Phrases like "In this section, we will discuss," "The results indicate that," and "It is worth noting" are predictable by definition - they're conventions.
• Over-editing: When you revise a sentence multiple times, you tend to converge on the "best" (most logical, most clear) version, which is often the most predictable one. First drafts typically have higher perplexity than final drafts.
• Technical writing: Domain-specific text with standard terminology has naturally lower perplexity because the vocabulary is constrained. A sentence about database normalization has fewer possible word choices than a sentence about a childhood memory.

How to increase perplexity naturally

You don't need to write nonsense to have high perplexity. You just need to write like yourself.

The most effective perplexity boost comes not from exotic vocabulary, but from personal specificity. "The meeting was unproductive" has low perplexity. "The meeting felt like watching paint dry on a Tuesday" has high perplexity, and it's better writing.

Specific techniques

Use concrete details instead of abstract summaries. Specific names, numbers, places, and references are harder for a language model to predict. "Revenue increased" is predictable; "Revenue hit $4.2M in Q3, up from $3.1M" is not.

Include personal voice. Asides, opinions, hedges, and informal phrasing all increase perplexity. "This approach is — frankly — overkill for most use cases" reads as more human than "This approach may be excessive for many use cases."

Vary your word choices. If you used "significant" in the previous paragraph, use "notable" or "substantial" in this one. A language model would predict you'd reuse the same word; choosing a different one raises perplexity.

Don't strip out your quirks. If you tend to use em dashes, start sentences with "But," or use parenthetical asides, keep doing it. Those patterns are unique to you and contribute to higher perplexity.

Perplexity in context

Perplexity is one of five signals GPTypo tracks, and it rarely causes a false positive on its own. But when combined with low burstiness and AI vocabulary, it contributes to a higher overall detection score. Understanding what it measures helps you know when to worry about it (rarely) and when to focus your editing elsewhere (usually).