How Do AI Detectors Work? AI Text Detection Explained

AI detectors work by measuring how predictable your writing is. They push the text through a language model of their own, ask how surprising each word is given the words before it, and compare that pattern against what they have learned from large samples of human and machine-written text.

The full answer to how do AI detectors work involves three overlapping techniques: perplexity, burstiness, and a supervised classifier trained on labelled examples. Most tools use some blend of all three.

What comes out the other end is a probability estimate. Not a confession, not a receipt, not proof. A detector can tell you that a paragraph looks statistically like machine output; it cannot tell you who sat at the keyboard.

How do AI detectors work under the hood

Perplexity: how surprised is the model?

Language models predict the next token. Feed them a sentence and they will rank every possible continuation by likelihood.

Perplexity is a measure of how often the actual next word was one the model would have ranked highly. Text generated by a model tends to sit in the high-probability lane, because that is literally how it was produced. Human writing wanders off into odd word choices, dead metaphors, regional phrasings and mistakes.

So low perplexity — smooth, unsurprising word choices — nudges the score toward "AI". That signal alone is weak, which is why nobody ships a detector built on perplexity by itself any more.

Burstiness: the rhythm across sentences

Burstiness looks at variation rather than average. Human paragraphs lurch: a 40-word sentence, then a four-word one, then a tangent. The predictability of each sentence swings around.

Default model output is flatter. Sentence lengths cluster, and each sentence is about as predictable as the last. Detectors measure that variance and treat unusually even rhythm as a signal.

Trained classifiers: learned fingerprints

The heaviest lifting in modern tools comes from supervised classification. The vendor assembles a corpus of human writing and a corpus of model output — often on matched prompts — and trains a model to tell the two apart.

That classifier picks up things no human wrote down as a rule: punctuation habits, transition-word frequency, how often a paragraph opens with a subordinate clause, the particular flatness of a generated conclusion. It is pattern-matching on style at a scale people cannot audit, which is both why it performs well and why its mistakes are hard to explain.

Finally, the tool aggregates. Sentence or passage scores get rolled up into a document figure — which is where a lot of nuance quietly disappears.

What an AI detector score actually means

A result of "85% AI" does not mean 85% of the document was generated. It usually means the classifier's confidence that the text belongs to the machine-written class is 0.85, given everything it was trained on.

That number is only as good as its assumptions. Change the input — a different subject, a heavily edited draft, a writer whose second language is English — and the calibration shifts underneath you.

This is why every result on our tool shows which passages drove the score rather than a single verdict. A document at 60% where two paragraphs are lit up tells you something specific: look at those paragraphs. A flat 60% across an entire essay tells you almost nothing. Our guide to reading detector evidence goes further into how to work with the highlighting instead of the headline figure.

Where detectors fail, and who pays for it

Detection accuracy is not uniform. It degrades predictably in these situations:

• Short text. Under roughly 150–300 words there simply is not enough signal. A single tweet or abstract is close to a coin flip.
• Formulaic prose. Lab reports, legal boilerplate, recipes, meeting minutes. Genres with rigid conventions look low-perplexity because the conventions are predictable.
• Non-native English writing. A widely cited Stanford study from 2023 found several detectors flagged essays by non-native speakers as AI-written far more often than native-speaker essays, because simpler vocabulary and sentence structure read as low perplexity.
• Mixed drafts. Human outline, AI expansion, human edit. The statistics blur and the document score lands somewhere unhelpful in the middle.
• Heavily copy-edited work. Ironically, a professional editor smoothing out a human draft can push it toward the machine end of the distribution.

The cost of a false positive falls on the writer, not the tool. That asymmetry is the reason no score should ever be treated as sufficient grounds for an accusation.

How the well-known AI detectors differ

The tools you have probably heard of use the same family of methods with different training data, thresholds, and reporting styles:

• GPTZero made perplexity and burstiness famous and still surfaces sentence-level highlighting.
• Turnitin's AI checker runs inside an existing academic workflow; the company reports a low document-level false positive rate but has publicly cautioned that sentence-level flags are less reliable and that scores are not evidence of misconduct.
• Copyleaks and Winston AI bundle detection with plagiarism scanning, so a single report mixes two very different kinds of finding.
• Pangram takes the pure-classifier route, training on paired human and model text and publishing low false-positive claims on its own benchmarks.
• Grammarly's AI detector leans on authorship data from its editor — where the text was typed or pasted — which is a provenance signal rather than a statistical one.

Asking what are the best AI detectors is less useful than asking which one shows you its working. Vendor accuracy figures come from vendor benchmarks, on text that resembles their training data. Your document may not.

Do AI detectors work on humanized text?

Humanizer tools exist specifically to break the statistics described above. They swap in lower-frequency synonyms, vary sentence length, insert small irregularities, and sometimes run the text through a second model.

Against perplexity-and-burstiness scoring, this often works. Against a well-trained classifier it works less reliably, because the classifier is watching structural habits that a synonym swap leaves intact. What it does do, consistently, is make the score less trustworthy in both directions — which is another reason to read the passages rather than the percentage.

If you want to see how a specific draft behaves, paste it into the free AI text detector and look at where the highlighting lands. Student work is usually better run through the essay-specific checker, which expects longer, structured prose. For output you suspect came from a particular model, the ChatGPT detector is tuned for that pattern.

Image detection works on completely different evidence

Text detectors read statistics of language. An AI photo detector reads statistics of pixels: frequency artefacts from diffusion upsampling, unnatural noise distribution, inconsistent lighting physics, mangled fine detail in hair, hands, jewellery and reflected text.

Generative image models also leave metadata and, increasingly, signed provenance data. That is stronger evidence than any classifier — but it survives only until someone screenshots the file. The image detector works the same way as the text side: a probability with the regions that drove it.

Frequently asked questions

Do AI detectors actually work?

Yes, in the narrow sense that they perform far better than chance on long, unedited text from mainstream models. No, in the sense that they cannot deliver certainty about a single document. They are screening tools, best used to decide where to look more closely.

Can AI detectors be wrong about human writing?

Routinely. Every detector has a false positive rate, and it is not evenly distributed — plain, formal, or non-native English writing is flagged more often. If your own work gets flagged, keep version history, drafts, and comment threads; process evidence holds up where a score does not.

How do AI content detectors work on text that was edited by a human?

Editing changes the surface statistics, so the score usually drifts toward "human" as more of the text is rewritten. A lightly edited generation often still shows machine-like passages; a thoroughly rewritten one may not be distinguishable at all. Mixed authorship is the hardest case for every tool on the market.

Which AI checker is most accurate?

Accuracy depends on the text type, the length, and which model produced it, so no single tool wins across the board. Run the same document through two or three, and treat agreement as informative and disagreement as a signal that the text sits in a genuinely ambiguous zone.

What to do with a score

Next time a document comes back flagged, do three things before anything else: check the length (under 300 words, discard the result), read the highlighted passages and judge whether they actually sound generated, and ask the writer to talk you through one specific paragraph. A score starts a conversation. Version history and a five-minute conversation end it. If you hit a case you cannot read, send it to us.