What Is Perplexity?

Perplexity is a measurement of how surprised a language model is when it encounters a piece of text — technically, how unpredictable each next word turns out to be given the words before it. Text that follows the statistical grooves of typical writing (common phrasings, expected word orders) has low perplexity; unusual, creative, or irregular writing has high perplexity.

Under the hood, the model assigns every candidate next word a probability drawn from its training data, combines those probabilities across the passage, and reports the result as one score. A sentence like "the committee approved the proposal" scores low because nearly any model would predict each word from the ones before it, while "the proposal outvoted its committee's absence" forces genuinely improbable predictions and pushes the score sharply up. Absolute numbers differ between models, which is why only relative comparisons — this paragraph versus that one — carry meaning.

The connection to AI writing tools comes from detection: because language models generate statistically likely text, their output tends toward lower perplexity than human writing. Detectors exploit this tendency, along with its sibling measurement burstiness. The catch is that humans also write predictably in plenty of contexts — formulaic business emails, standard legal prose — which is exactly why perplexity-based detection produces false positives on stiff human writing and why no detector score should be treated as proof of anything.

That points to the most common misconception: low perplexity does not prove machine authorship, it proves predictability, and a tired human filling in a template produces predictability just as reliably as a model does. For writers using AI tools, the implication cuts both ways. Pasting raw generator output into an essay hands over uniformly predictable text, but even modest manual edits — inserting one specific detail, breaking a stock phrase with an unexpected word — raise individual predictions' improbability and lift the whole passage. Edit for specificity rather than for trickery, and the score tends to follow.