Perplexity and Burstiness Explained: How AI Detectors Score Text

What is perplexity?

Perplexity is a number that describes how surprised a language model is by a piece of text. A language model reads text one token (roughly a word or part of a word) at a time and assigns a probability to each next token. When the actual next word is one the model expected, the probability is high and the surprise is low.

Formally, perplexity is the exponential of the average negative log-probability of the tokens. You do not need the math to use the idea: low perplexity = predictable text, high perplexity = surprising text.

Large language models such as ChatGPT, Gemini and Claude generate text by repeatedly picking likely next words. Their output therefore tends to have low perplexity when it is scored by a similar model. People, on the other hand, make odd word choices, use slang, jump between ideas and reference things only they know, which raises perplexity.

What is burstiness?

Burstiness describes variation across a text, most often in sentence length and structure. Human writing is “bursty”: a long, winding sentence is followed by a short one. Then another. Paragraphs speed up and slow down depending on what the writer cares about.

AI-generated text is often more uniform. Sentences cluster around the same length, paragraphs are similar in size, and the rhythm stays steady from start to finish. A common way to measure this is the coefficient of variation of sentence lengths: the standard deviation divided by the average. A low value means the sentences are all about the same length.

• Low burstiness: five sentences of 18, 20, 19, 21 and 18 words.
• High burstiness: sentences of 31, 6, 22, 3 and 15 words.

Burstiness is also about structure, not only length: repeated sentence openings (“Moreover,” “Furthermore,” “Additionally,”), identical list patterns and evenly balanced paragraphs all lower it.

How AI detectors combine these signals

The terms were popularized in early 2023, when GPTZero described its approach using perplexity and burstiness. Today most detectors use a wider mix of signals, often grouped into families:

• Predictability — perplexity or similar token-probability scores (requires running a language model).
• Rhythm and structure — burstiness, paragraph uniformity, templated openings and conclusions.
• Wording — stock phrases such as “delve into,” “a testament to” or “it's important to note.”
• Specificity — names, numbers, dates and first-hand details that generic AI drafts lack.
• Voice — first person, contractions, questions and informal words.

Our AI detector is open about which of these it uses. It does not compute token probabilities; it scores rhythm, wording, specificity and voice in your browser, highlights the sentences behind the result and shows a confidence level that drops when signals are weak or conflicting. The page How it works explains the full method.

A worked example

Consider this 60-word paragraph: “Artificial intelligence has revolutionized the way businesses operate. Moreover, it plays a pivotal role in enhancing efficiency. Furthermore, organizations that leverage AI can unlock new opportunities. Additionally, AI fosters seamless collaboration between teams.”

A detector would notice four things at once: every sentence is 9–11 words long (low burstiness), three sentences open with formal transitions, the text contains several stock phrases (“revolutionized the way,” “pivotal role,” “leverage,” “fosters seamless”) and there is not a single name, number or example. Each signal is weak on its own. Together, they point strongly toward AI-style writing.

Now compare: “Our support team cut reply times from 9 hours to 2 after we started drafting answers with an AI tool in March. Not every reply works — customers notice canned answers fast — so Priya still rewrites about a third of them.” Varied sentences, specific numbers, a name and an opinion: this reads as human, even if AI was involved behind the scenes.

Why perplexity and burstiness can mislead

Both signals measure style, not authorship. That leads to predictable false positives:

• Non-native English writing often uses simpler, more predictable vocabulary, which lowers perplexity. A 2023 Stanford study found several detectors flagged non-native writers' essays far more often.
• Formal and technical genres — legal clauses, lab reports, manuals — are uniform by design.
• Short texts contain too few sentences to measure variation reliably.
• Famous text (the U.S. Constitution, well-known speeches) appears so often in training data that models find it very predictable.
• Edited AI text can have human-level burstiness after a person rewrites a few sentences.

How to get a meaningful score

1. Paste the complete text, ideally 300 words or more. The minimum on AIDetector365 is 80 words.
2. Choose the closest focus: general text, long-form writing, web content or possibly paraphrased text.
3. Read the verdict together with the confidence level, not the percentage alone.
4. Open the highlighted sentences and check whether the reasons actually apply.
5. Treat the result as one signal. Drafts, version history and a conversation matter more.

Perplexity, burstiness and the future of detection

As language models improve, their output becomes more varied and less predictable, which narrows the gap these signals rely on. Prompting a model to “vary sentence length” or “write casually” also changes them. That is why serious detection guidance — from universities and from detector vendors themselves — stresses that scores are probabilistic and should never be the sole basis for a decision.

For a broader view of reliability, read How accurate are AI detectors? and our complete AI detector guide.

Frequently asked questions