How AI text detectors work: perplexity, burstiness, and limitations

AI content detectors try to answer one question: How likely is it that this text was written by an AI? To do that, they look at how predictable and how varied your writing is.

Perplexity: how predictable the next word is

Perplexity is a measure of how surprised a language model is by each word in your text. Low perplexity means the text follows common, highly predictable patterns; high perplexity means the model finds the wording more unusual or creative.

AI models are trained to choose statistically likely words, so AI-generated text often has lower perplexity than human writing. Human authors naturally introduce quirks, digressions, and unexpected turns of phrase that push perplexity higher.

Burstiness: the rhythm of your sentences

Burstiness describes how sentence length and structure vary across a piece of writing. Humans tend to mix short, punchy lines with longer explanations. AI-generated content, especially from default settings, often lands in a flatter rhythm with similar sentence lengths and structures throughout.

Detectors use burstiness to see whether a passage feels like it was written in one smooth, uniform pass (a common AI pattern) or shaped by a human who speeds up, slows down, and changes tone.

Other patterns detectors look for

Modern AI text detectors combine perplexity and burstiness with other signals, such as:

  • Repetitive phrasing and stock transitions that often appear in model outputs.
  • Overly generic language that avoids concrete details or personal experience.
  • Style fingerprints that come from specific AI families (for example, ChatGPT-style phrasing).

Why detectors can still be wrong

Even the best AI content detector is probabilistic, not perfect. Some formal or highly polished human writing can look “too AI-like,” especially from non-native speakers or when people follow rigid templates. The opposite is also true—carefully edited AI outputs can look more human.

That is why you should treat every score as a signal, not a verdict. Results are most helpful when combined with context, domain knowledge, and your own judgment.

When to use an AI text detector

You might reach for a detector when you:

  • Review content for originality and want an extra signal beyond a plagiarism check.
  • Need to understand how heavily AI was used in a draft before you publish it.
  • Are comparing multiple pieces of content and want a consistent way to flag potential AI use.

Try it on your own text

The easiest way to understand how AI detection feels in practice is to run your own writing through a tool and see the output.

Check your text with the AI detector

Paste any paragraph, essay, or draft into the detector and see how likely it is to be AI-generated.