How do AI detectors work?

The short answer, the signals detectors look for, and why no detector can be 100% accurate.

Short answer: an AI detector is a machine-learning model trained on large amounts of human-written and AI-generated text. It learns statistical differences between the two, such as how predictable the word choices are and how evenly the sentences flow, and outputs an estimate of how much of a new text is AI-written.

The signals AI detectors look for

Predictability

Language models generate text by choosing likely next words. The result is text that is, on average, more predictable than human writing. Researchers call the measure of this perplexity: lower perplexity means more predictable text.

Uniformity

Human writers vary sentence length and structure more; AI writing tends to be evenly paced. This variation is often called burstiness.

Learned patterns

Modern detectors don't rely on a single hand-made rule. They are themselves language models, fine-tuned to tell human and AI text apart, and they learn many subtle patterns at once, including style, word choice and structure.

How TextFlag works

TextFlag uses two transformer models. A context model reads each sentence with its neighbours and sets the document's AI percentage. A sentence model scores each sentence on its own and decides the highlights, and a sentence is highlighted only as part of a run of consecutive AI-like sentences. The percentage is then calibrated against 658 documents of known authorship, so it describes the share of AI-written text rather than a raw model probability. Details are on how TextFlag works.

Why AI detectors make mistakes

  • Overlap: some human writing (formal, formulaic, heavily edited) is as predictable as AI text.
  • Editing: paraphrasing and rewriting remove many of the patterns detectors rely on.
  • Length: short texts carry too little signal for a confident estimate.
  • Language: detectors are usually calibrated on English and are less reliable elsewhere.

How to read a detector's result

Treat the score as an estimate with a margin of error, look at which passages were flagged, and never use it as the only evidence. See what an AI detection score means.

FAQ

Questions people also ask

What do AI detectors look for?

Statistical patterns that separate AI text from human text: predictable word choice (low perplexity), even sentence rhythm (low burstiness) and many subtler patterns that a trained classifier learns from examples.

Do AI detectors use perplexity and burstiness?

Early detectors relied on them directly. Modern detectors, including TextFlag, are fine-tuned language models that learn these and many other signals from labelled examples.

Can AI detectors be 100% accurate?

No. Human and AI writing overlap, so every detector has some false positives and misses. Good detectors publish their error rates.

Check your own text for AI

Free for up to 10,000 words. Highlighted report and PDF included.

Start a free check