Do AI Detectors Still Rely on Perplexity and Burstiness?

Search for how GPTZero detects AI writing today and most of what comes back still centers on two statistics, perplexity and burstiness, the pair that made the tool famous within weeks of its January 2023 launch. GPTZero's own support documentation now says something different: it stopped using both of them in the autumn of that same year.

So is burstiness still used by AI detectors? Not by the tool that made the term a household word in academic circles. GPTZero's support center states plainly that it no longer uses perplexity or burstiness, having moved to a deep-learning architecture instead. Turnitin, the tool most students actually encounter, has never published either term as part of its method in the first place. The concepts are not gone from the conversation. They are gone from the products doing the actual detecting.

Is Burstiness Still Used by AI Detectors?

No, not by GPTZero, and not by name. GPTZero's own support documentation says so directly: "As of autumn 2023, GPTZero no longer uses perplexity and burstiness for its AI detection because we migrated to a deep-learning based architecture." Turnitin's own guides describe a different method again, built around scoring chunks of a submission with a trained classifier, and neither term appears anywhere in that documentation either. The two statistics that defined the first wave of AI detection are not what the leading tools currently run. That is narrower than saying they were thrown away: GPTZero's own explainer still counts them among seven indicators feeding its model. They stopped being the method without ceasing to be a measurement, and the difference between those two claims is most of what this question turns on.

What Changed in Autumn 2023

The two terms were never obscure. GPTZero's own early explainer defined perplexity as "a measure of how likely an AI model would have chosen the exact same set of words as found in the document," and burstiness as "a measure of how much writing patterns and text perplexities vary over the entire document," a document-level statistic, not a simple sentence-length count. Both definitions are still published on GPTZero's site. What changed is which of them the detector actually runs.

"Migrated to a deep-learning based architecture" is a specific claim, not a marketing phrase, and it describes a real change in kind. A perplexity score is computed directly from a formula: run the text through a reference language model, average the log probabilities, exponentiate the result. A deep-learning classifier is trained instead, shown large numbers of human and AI-written examples until it learns, on its own, whatever combination of features separates the two piles in that training data. Nobody hand-writes the rule in the second approach. The model finds it.

Both retired terms get a full explanation elsewhere on this site: what perplexity actually measures and what burstiness actually measures each cover their own mechanics in depth, including the formulas behind them. This piece stays narrowly on whether either one is still what a detector runs today.

What GPTZero and Turnitin Actually Document Today

The current GPTZero technology page, which describes what is shipped in the product, states that the company employs "an end-to-end deep learning approach, trained on text datasets from the web, education, and AI-generated from a range of LLMs", in which "a sentence-by-sentence classification model determines the probability and confidence that a text was created by AI". Perplexity and burstiness do not appear anywhere on that page, not as a component, not as a legacy feature, not in a footnote.

Turnitin's documentation was never built around either term to begin with. Its guides describe a submission split into chunks of a few hundred words, each chunk scored by a trained model, and the chunk scores averaged into the single percentage a report displays. That is a supervised classifier working on segments of text, the same family of method GPTZero has moved to, not a perplexity calculation and not a burstiness calculation under a different name.

The shift is easiest to see side by side.

	What early explainers described (2023)	What current documentation describes
GPTZero	Perplexity and burstiness, scored against a reference model	A sentence-by-sentence deep-learning classifier outputting a probability and confidence score
Turnitin	Never published; the terms do not appear in Turnitin's own materials at any point	Chunks of a few hundred words scored by a trained classifier, averaged into one percentage

Why the Concepts Still Explain Detectability

None of this makes perplexity and burstiness wrong, only outdated as a description of the current mechanism. Both concepts were always describing something real: Text generated by sampling toward a model's own highest-probability continuations tends to be more predictable, word by word, than text a person writes from scratch; it tends to vary less from one sentence to the next. The underlying regularity hasn't changed because a vendor changed its detector's architecture. What changed is how the detector goes looking for it.

The trained classifier is not explicitly asked to look for perplexity or burstiness. Rather, it's trained on the same basic distinction between human and machine text that those two statistics were designed to capture. And it seems highly unlikely that a classifier trained to distinguish the two piles wouldn't key on precisely that regularity (one of the most reliable differences between the two) as well. This is an inference about why the classifier likely works, not a claim about its published feature list, and the two should stay separate.

Independent research backs the idea that the underlying statistic still has teeth, even outside any single vendor's product. DetectGPT, an academic zero-shot method published in 2023, looks at a close cousin of perplexity, how sharply a model's log-probability function curves around a passage, and reaches 0.95 AUROC on one benchmark against 0.81 for the best prior zero-shot baseline, without training a classifier on labelled examples at all. The core idea, that machine-generated text sits in a statistically distinct region of a model's own probability space, is still active research. It simply is not what GPTZero ships to users.

Why So Much of the Internet Still Gets This Wrong

Most published explanations of how AI detection works were written during or shortly after GPTZero's January 2023 launch, when perplexity and burstiness were the only public framing the company itself offered. Those original explainer pages are still live today. The autumn 2023 retirement notice sits on a separate support page rather than folded into them, so a writer who read the first page and stopped researching would have no reason to know the method changed less than a year later.

A third-party detector review circulating in 2026 claims GPTZero still runs perplexity and burstiness as one layer among several inside a larger system. That claim conflicts directly with GPTZero's own current technology page and its own support documentation, and nothing on GPTZero's own site corroborates it. A company describing its own shipped product still outranks an uncorroborated third-party claim about that same product, even though vendor documentation can be wrong too. This piece follows GPTZero's own pages rather than the review repeating the older story.

What This Means for How You Write

The practical guidance built around burstiness has not become useless just because the word disappeared from GPTZero's own materials. Varying sentence length deliberately is still, mechanically, an argument about avoiding the same statistical uniformity a classifier is very likely trained to notice, whatever it calls that signal internally. The underlying advice survived the vendor's architecture change even though the vocabulary describing it did not.

None of this requires taking anyone's word for it, including this post's. TextPulse's free perplexity checker and the wider free tools collection let you look at where your own draft sits before a detector, of whichever generation, ever does.

If the answer sends you back to your own writing, the practical follow-ups are increasing burstiness deliberately and varying sentence length across an academic paragraph, which is the same job described without the metric.

The mechanism a detector runs will keep changing; GPTZero has already changed it once. What stays constant is the actual question, covered in full in this site's guide to how AI detectors actually work: whether predictable writing is the same thing as machine-written writing, regardless of which formula happens to be fashionable in a given year.

Related research: the findings above are examined at scale in Sentence-Length Burstiness as a Signal of AI Rewriting, a TextPulse Research working paper with open data, code and a citable DOI. Whether prompting alone can make a model write like a person is tested in Do AI Models Speak Human?.