Perplexity: what it means and why detectors care
Perplexity is a measure of how predictable a piece of text is. If a language model can guess each next word easily, the text has low perplexity; if the words keep surprising it, perplexity is high. AI detectors lean on this number because text written by a model tends to be made of the words a model would have chosen, which is another way of saying it is easy to predict. Humanize 360 reports word predictability as one of the eleven signals in its free AI detector.
Take two sentences. "The weather today is sunny and warm, perfect for a walk in the park." Every word follows from the one before it. Now: "The weather today is arguing with itself, so I brought a hat and an umbrella." The second sentence is not better writing, but it is harder to predict, and a detector would treat it as more likely to be human. That gap, repeated across a whole essay, is what perplexity captures.
Where the number comes from
Perplexity started as a way to score language models, not writers. A model assigns a probability to each word given the words before it. Average those probabilities across a passage, flip and exponentiate, and you get a number you can read as "how many equally likely choices the model was juggling at each step". Two choices is very predictable. Two hundred is not.
A detector does not need the model that wrote your text. It uses its own reference model and asks how surprised that model is. Different reference models give different numbers for the same passage, which is one of the reasons two detectors can disagree about one essay.
Why it matters for writing that reads human
People write with specific nouns, private references, odd verbs and the occasional joke. A model, left on its default settings, picks the safest word at every turn. Nobody notices one safe word. Six hundred safe words in a row read flat, and both readers and detectors feel it. Raising perplexity does not mean using long words. It means saying the particular thing rather than the general one.
How Humanize 360 uses it
- The free detector reports word predictability as its own line in the eleven-signal breakdown, so you can see whether this signal or another one is dragging the score.
- The humanizer modes replace generic phrasing with more specific wording, and Heavy strength pushes harder on word choice than Light.
- A custom style profile measures your own vocabulary spread from 300 words of your writing, so the output aims at your level of surprise rather than a generic one.
- The Changes view shows every substitution side by side, so you can reject any that go too far.
Common misunderstandings
- High perplexity is not the same as good writing. Random nonsense scores very high. The goal is text that is specific, not text that is strange.
- Perplexity on its own is a weak detector, which is why it is almost always paired with burstiness and other signals.
- A perplexity number is relative to the model that measured it. There is no universal scale, so comparing numbers across tools is meaningless.
Common questions
Is low perplexity proof that text was AI-written?
No. Simple instructions, legal boilerplate and textbook definitions are all low perplexity and written by people. It is one signal among several, and a good detector shows you the others too.
Does using rare words raise perplexity?
A little, but rare words used in predictable slots do not help much. Specific nouns, concrete details and unexpected but sensible verbs move the number more than a thesaurus does.
Which Humanize 360 signal is perplexity?
Word predictability. It sits next to sentence-length variance in the detector breakdown, and the two together explain most low scores we see.
Can I see my text's perplexity for free?
Yes. Paste up to 300 words into the detector above, three times a day, with no account. The breakdown shows word predictability alongside the other ten signals.