Quick Takeaways

• AI detectors analyze statistical patterns in text and estimate how closely those patterns resemble human-written or AI-generated writing. They cannot directly verify who wrote a passage.
• Perplexity measures how predictable wording is to a language model, while burstiness describes variation in sentence patterns. Some detection methods use these or related features, but not every detector relies on both.
• Other approaches may examine repetition, vocabulary, syntax, readability, probability patterns, or watermarks.
• Formal writing, non-native English, short samples, and edited text can affect detection results.
• A detector score is an estimate, not proof of authorship. A percentage should not automatically be interpreted as a reliable probability.

You paste a paragraph into a detector and get back “78% likely AI.” It feels like a measurement, so it helps to understand what was actually measured. This article explains how AI detectors analyze text, signal by signal, and where each approach tends to break down.

This article covers the main techniques used in AI-text detection. It does not describe any single product’s exact algorithm, because commercial vendors rarely publish all the details of their detection systems.

What a Detector Is Actually Scoring

AI detectors analyze text by measuring statistical features, such as word predictability, repeated patterns, and variation in sentence structure. They then compare those features with patterns learned from examples of human-written and AI-generated text. The resulting score is an estimate, not proof of authorship.

Some detectors express their results as percentages, but a number displayed as “78% AI” does not necessarily mean there is a calibrated 78% probability that a machine wrote the passage. Different tools may calculate and present their scores differently.

That distinction matters. A detector generally sees the finished text rather than the author’s writing process, original drafts, or revision history. It cannot directly establish who wrote something simply by examining the words.

If you want the broader picture first, how AI detection tools work explains the overall approaches before we examine individual signals.

Perplexity: How Predictable Is the Wording?

Perplexity measures how predictable a piece of text is to a language model. A model evaluates a sequence of words by estimating how likely each word is, given the preceding context. Text composed of highly predictable word sequences generally has lower perplexity, while less predictable sequences tend to have higher perplexity.

Some AI-detection approaches use this concept because language models often generate text by selecting words according to learned probability distributions. However, the relationship between predictability and authorship is not straightforward.

Predictable text is not necessarily machine-generated. A legal clause, recipe, textbook definition, or familiar quotation may use predictable wording because the subject and writing conventions leave little room for variation.

Likewise, human writers can produce surprising, unusual, or highly varied language. A detector that relies too heavily on predictability may misinterpret these differences.

According to a DEV Community explainer on perplexity and burstiness, these concepts help explain some approaches to AI-text detection. They are useful starting points, but neither is a universal test of authorship.

Burstiness: How Much Does the Rhythm Change?

Burstiness describes variation in patterns across a piece of text. Depending on the method, this may involve differences in sentence length, sentence structure, or predictability from one sentence to another.

Human writers often mix short sentences with longer ones. They may also change their writing style depending on the audience, subject, or purpose. Some AI-generated text can appear more uniform, with sentences following similar structures and rhythms.

A detection system that measures sentence-length variation or changes in predictability may treat a relatively uniform pattern as one possible signal of machine-generated writing.

However, burstiness is a limited clue, not a reliable rule. Many human writers use consistent sentence structures in reports, instructions, academic papers, and technical documentation. Conversely, AI-generated text can contain substantial variation.

Editing can also change a passage’s rhythm. Adding shorter sentences, combining paragraphs, or restructuring a draft may affect the statistical features that a detector measures.

For these reasons, burstiness is best understood as one possible feature in a larger analysis, rather than a standalone way to identify AI-generated text.

Other Signals Detectors Combine

Perplexity and burstiness receive considerable attention, but AI-detection methods can examine many other features. Research describes approaches that combine multiple signals rather than relying on one measurement.

• Repetition and n-grams. These features examine recurring words, phrases, and sequences of words. Repetition can be useful to measure, but it is also common in legitimate human writing.
• Vocabulary variety. A system may measure how many distinct words appear relative to the total number of words. The result can vary with the subject, genre, and writer’s vocabulary.
• Syntax and punctuation. Sentence construction, clause length, punctuation habits, and other structural features may contribute to a text’s statistical profile.
• Readability and text complexity. Some approaches measure features associated with sentence complexity or reading difficulty. These measurements describe aspects of writing, but they do not independently establish authorship.

One research example illustrates how features can be combined. The NotAI.AI paper describes a system that combines probability curvature and a neural detector score with readability and text-complexity measures. A separate classifier then uses these features to make its prediction.

The important point is that no single feature necessarily determines the result. In a multi-feature system, the classifier learns how to combine the available signals based on its training data and design.

The exact features, their relative importance, and the way they are combined vary across detection systems.

Probability Curvature and Watermarks

Two other approaches deserve attention because they work differently from simple measurements of writing style.

Probability curvature. DetectGPT is a research method based on the observation that text sampled from a language model may occupy regions of relatively high probability under that model. It examines how a passage’s probability changes when the passage is perturbed, or slightly rewritten, multiple times.

If the original passage has a probability profile that differs systematically from those of its perturbed versions, that pattern can provide evidence relevant to the detector’s prediction.

The original DetectGPT research paper describes this approach. It requires access to a suitable language model for probability calculations, and its effectiveness depends on the model, the text, and the evaluation setup. Results may not transfer equally well across different models and types of writing.

Watermarks. Some text-generation systems can introduce a hidden statistical bias into the text they produce. For example, a generation system may subtly favour certain token choices according to a secret pattern. A detector that knows the relevant watermarking method can then test for evidence of that pattern.

The original watermarking proposal for large language models describes this general idea. Watermark detection depends on the generator having applied a compatible watermark in the first place. Editing, paraphrasing, or other transformations can also weaken the signal.

Related research on identifying watermarked models explores how detection reliability can depend on watermark strength and the method used.

There is an important distinction here: watermark detection and general AI-text detection are not the same thing. A watermark detector looks for a particular signal associated with a compatible generation system. A general AI-text detector attempts to classify text based on patterns it has learned or features it measures, without necessarily having access to a hidden watermark.

How the Signals Combine Into One Score

Some AI detectors use trained classifiers that combine multiple features. A classifier is trained on examples labelled as human-written or AI-generated and learns patterns that help it distinguish between the two groups.

When new text is submitted, the system extracts the features it was designed to measure and uses its model to produce a classification or score.

Other systems may use different architectures or combinations of methods, so this description should not be taken as the exact design of every commercial detector.

Two practical consequences follow.

First, two detectors can disagree sharply on the same passage. They may use different training datasets, model architectures, features, decision thresholds, or definitions of what their scores represent.

Second, a confident-looking result can still be wrong. A classifier learns from its training examples and may perform poorly on text that differs from those examples. Changes in writing style, subject matter, language background, or editing can affect its results.

A detector’s output should therefore be interpreted in the context of the tool’s documented capabilities and limitations, rather than treated as a definitive answer.

Quick Comparison

Signal	What It Examines	Potential Benefit	Key Limitation
Perplexity	Word-sequence predictability	Measures predictability in methods that use it	Can mischaracterize predictable human writing
Burstiness	Variation in sentence or text patterns	Captures differences in writing rhythm	Varies by genre and can change with editing
Style features	Repetition, vocabulary, syntax, and other patterns	Adds context to other signals	Patterns vary among writers and genres
Probability curvature	Changes in model probability across perturbed text	Offers an approach beyond simple surface statistics	Depends on the model and evaluation setup
Watermarks	Hidden statistical signals inserted during generation	Can help identify text from compatible generators	Requires a watermark and may be affected by editing

Where Each Signal Fails

A common assumption is that a high AI score means a passage was written by a machine. Research shows why that conclusion can be unreliable.

• Formal writing. Legal, academic, and technical prose often follows established conventions. Its predictable wording or consistent structure may affect the features a detector measures.
• Non-native English writers. A widely cited study found that several tested detectors disproportionately misclassified writing by non-native English speakers. The research examined a specific set of TOEFL essays, so its results should not be assumed to apply equally to every detector or every writing task. Nevertheless, it illustrates how language patterns can contribute to false positives. See the study on the bias of AI-text detectors against non-native English writers.
• Short text. A short passage provides fewer words and patterns to analyze. Some detection methods may become less reliable when there is insufficient text to measure their chosen features.
• Edited text. Human editing, paraphrasing, and other transformations can change the statistical patterns that detectors examine. A low AI score does not prove that a person wrote the text, just as a high score does not prove that AI generated it.

For a closer look at the human side of this problem, why AI detectors sometimes flag human-written text discusses the problem of false positives.

If you’re wondering how rewriting tools fit into this picture, what an AI humanizer is and how it works explains what these tools are designed to change, while AI humanizer vs AI detector compares the two tool types.

What This Means When You Read a Score

Detector vendors may report high accuracy figures under particular testing conditions, while independent evaluations can produce less favourable results. These findings are not necessarily contradictory: a detector may perform well on the data and conditions used in one evaluation but perform differently on other types of writing.

When interpreting a detector result, keep a few practical points in mind:

1. Treat the score as one signal, not a verdict. A percentage alone cannot establish who wrote a passage.
2. Consider the text’s length. Short samples may not provide enough information for reliable analysis.
3. Consider the genre and language background. Formal writing, formulaic content, and writing by non-native English speakers may be vulnerable to particular types of false positives.
4. Look for process evidence when appropriate. Drafts, notes, revision history, and other relevant records can provide context that a detector cannot see in the finished text.
5. Avoid relying on a single tool for consequential decisions. Where the result could affect a person’s education, employment, or reputation, use a fair review process and consider independent evidence.

A detector score alone should not be treated as sufficient grounds for a penalty. The limitations described above make it important to consider the surrounding evidence, the context, and the consequences of a possible mistake.

Quick Answers

What is perplexity in AI detection?

Perplexity measures how predictable a sequence of words is to a language model. Some detection methods use it as one feature when estimating whether text resembles AI-generated writing. Lower perplexity does not prove that AI wrote the passage.

What is burstiness?

Burstiness describes variation in patterns such as sentence length, structure, or predictability. Some detection methods use these variations as a feature, but human and AI-generated writing can both have consistent or varied rhythms.

Do detectors use only perplexity and burstiness?

No. Depending on the system, detectors may examine repetition, vocabulary, syntax, readability, or other statistical features. Some approaches also use trained classifiers, probability-based methods, or watermark detection.

Why do detectors give different scores for the same text?

Detection tools can differ in their training data, features, model architecture, scoring methods, and thresholds. Their outputs may therefore vary considerably when analyzing the same passage.

Can a detector prove who wrote something?

No. A detector estimates how closely text matches patterns associated with particular classes of writing. Its score alone cannot verify authorship or establish the writing process.

Conclusion

Understanding how AI detectors analyze text makes their scores easier to interpret. Depending on the method, a detector may examine predictability, sentence patterns, style, model probabilities, or hidden watermarks before producing an estimate.

Each approach has limitations, and the results can vary with the text, the tool, and the circumstances in which it is used. A detector score can provide a useful signal, but it should not be mistaken for proof of who wrote something.

For more plain-language explanations of artificial intelligence, explore the AI Explained section on OpenHerald.