What is perplexity in AI?
The number that measures how surprised a model is by the text it reads, and why that says nothing about whether an answer is grounded.
Perplexity measures how well a language model predicts text: a model that expects each next word with high confidence gets a low score, and a model that keeps getting surprised gets a high one. Lower is better. The score is computed over a fixed text, which is why it compares models under specific conditions, and why it never tells you whether a finished answer is right, only how well the model anticipates language.
A score for next-word prediction.
Training a language model is the process of improving exactly one skill: given the words so far, predict what comes next; what training data is describes what that practice material is. Perplexity is the measurement of the same skill on text the model has not seen. A score of 10 roughly means the model hedges between ten options on average at each step, while a score of 500 means it is very often taken by surprise. That is the whole idea. It is cheap to compute, it does not require human judgments, and it has been the standard sanity check since long before models could hold a conversation.
What the number does not measure.
Perplexity says nothing about truth. A model can predict language beautifully and still produce confident fabrication, because stating a plausible-sounding fact is itself a form of fluent text prediction. The score also says nothing about whether the model will stick to the documents you gave it, refuse when the evidence is missing, or show its sources; those are properties of the surrounding structure, and the fixes for making things up are structural, not statistical. A low score tells you the model speaks well. Whether it can be trusted on a specific question is a separate question entirely.
Where you will actually meet it.
Perplexity shows up in model cards, where publishers report it as evidence the base model learned language well. Bigger models generally score lower, which is one reason the count matters; what 7B means in a model name covers that side. But a card's number describes the base model on the benchmark's text, with the benchmark's tokenizer, not your documents and not your questions, so treat it as a quality hint rather than a promise about the answers you will get.