How AI Detectors Use Perplexity and Burstiness

AI detectors do not discover authorship hidden inside a document. They measure patterns and estimate how compatible those patterns are with examples of human or machine-generated writing. Perplexity and burstiness are two popular ways to explain that process, but they are not a universal recipe used by every detector.

That distinction matters. A detector score can support a review, yet it cannot prove misconduct, identify a writer, or reconstruct how a document was produced. If you want to inspect a passage, use an AI detector as one signal and keep the surrounding evidence in view.

Perplexity in plain English

A language model assigns a probability to each possible next token. Perplexity summarizes how surprised the model is by the sequence it actually sees. In simplified terms, the calculation takes the average negative log probability of the tokens and then converts it back to an easier-to-compare scale.

Lower perplexity means the reference model considered the sequence more predictable. Higher perplexity means it encountered more unexpected choices. The important phrase is reference model: the same paragraph can receive different perplexity values from different models or tokenizers.

Suppose a sentence begins, “The committee reached a...” A model may rank “decision” as a very likely continuation and “crossroads” as less likely. A passage that repeatedly selects high-ranked continuations may look statistically smooth. That can be a machine-generation clue, but it is not an authorship certificate. Templates, legal clauses, weather reports, and beginner writing can all be highly predictable.

The 2019 GLTR paper visualized token ranks to help people notice generation artifacts. Its authors described statistical assistance for human review, not a rule that every predictable word must be machine-written.

What people mean by burstiness

Burstiness usually refers to variation across a passage. Depending on the implementation, that may include sentence length, syntactic shape, vocabulary, punctuation, or changes in token surprise from one region to another.

There is no single industry-standard “burstiness formula.” One product may compare the variance of sentence lengths. Another may combine many learned features and use the word only in its explanation. Asking how a particular score was calculated is therefore more useful than treating burstiness as a fixed scientific unit.

Burstiness is best treated as a descriptive clue. Human prose is not always varied, and generated prose is not always smooth.

What modern detectors may combine

Current systems can use signals that go far beyond a single perplexity number. Product details are often proprietary, but research methods show several broad families.

DetectGPT, for example, tests the curvature of a model’s log-probability function around a passage. That is different from declaring “low perplexity equals AI.” A commercial detector may also combine stylometric features, embeddings, classifiers, and calibration layers before it produces one score.

A before-and-after pattern example

Consider this deliberately generic draft:

Remote work offers many benefits for modern teams. It improves flexibility, increases productivity, and supports collaboration. In addition, it helps businesses attract talent. Overall, remote work is an important solution for the future.

Now compare an edited version:

Remote work solves one problem and creates three more. Hiring is no longer tied to a commute, but decisions need written context, quiet work needs protection, and new employees need deliberate introductions. The benefit is not “more collaboration.” It is access to a wider team—if the operating habits can support it.

The revision changes the rhythm, removes stock transitions, and replaces broad benefits with conditions. It is also more useful to a reader. These are editorial observations, not measured detector results: we did not submit the passages to commercial tools and are not claiming that one is “human” while the other is “AI.”

This distinction prevents a common mistake. Editing only to chase a detector score can introduce unsupported claims or distort the author’s meaning. Edit for clarity, factual accuracy, and reader value first. Our guide to humanizing content without losing facts provides a claim-preservation workflow.

Why short, edited, and unfamiliar text is difficult

All classification needs enough relevant information. A very short passage contains fewer patterns, so one sentence can swing the score. Quoted material, bullet points, code, citations, or mixed authorship make the interpretation harder still.

Editing creates another distribution shift. A human may revise generated text, or a model may lightly edit human text. The final document is not cleanly represented by a binary origin label. The RAID benchmark, which includes more than six million generations across models, domains, decoding strategies, and adversarial attacks, found that detectors can be vulnerable under unseen conditions and attacks.

Language and time also matter. A 2026 EACL study in a contemporary Czech setting did not find the systematic bias reported in every earlier setting and suggested that newer detectors may not rely primarily on perplexity. That result is a useful warning against universal claims in either direction: detector behavior must be measured for the actual language, date, domain, and product version being used.

How to interpret a detector score

A percentage usually describes a model output or calibrated risk estimate. It is not automatically the probability that a particular person cheated. Before acting on it, ask:

• What text length and language does the detector support?
• Was the tool evaluated on this genre and on current generation models?
• What false-positive rate applies at the threshold being used?
• Does the result stay stable when quotations, references, and formatting are removed?
• Is there independent evidence such as drafts, document history, citations, or an author explanation?

For a low-stakes content review, a score may simply prompt a closer read. For grading, employment, publishing, or discipline, the cost of a false accusation is much higher. No single detector result should carry that decision by itself.

Improve the writing, not the illusion

Perplexity and burstiness can help explain why a passage feels statistically ordinary or unusually uniform. They cannot tell whether its claims are correct, whether its sources are credible, or whether it helps the intended reader.

Use a readability scorer to locate dense sentences, then review the substance yourself. If AI assistance is allowed, preserve drafts and disclose it according to the relevant policy. Strong writing comes from specific evidence and accountable revision—not from trying to manufacture a “human” pattern.

Methodology and disclosure

This article is a technical synthesis of the primary research linked above. It does not report a new benchmark of commercial detectors, and the header is an original conceptual illustration rather than a detector screenshot. Product implementations can change, so readers should consult each detector’s current documentation before relying on a score.