How Do AI Detectors Work? Perplexity, Classifiers, and Limits

AI detectors work by analyzing statistical patterns in text and comparing them with patterns typical of human and machine writing. Most combine measures of predictability (how expected each word is) with a classifier trained on large sets of labeled examples. The output is a probability that the text, or individual sentences, resemble AI-generated writing, not proof of who wrote it.

That summary hides a lot of interesting detail. Understanding it will help you read detector results with the right amount of trust.

The core idea: AI text is predictable

Language models write by choosing words one at a time. At each step, the model estimates which word is most likely to come next and usually picks one of the top candidates. This makes the text smooth and fluent, but also statistically "safe."

Human writers are less consistent. We use odd word choices, inside jokes, sudden asides, and sentences that wander. We also make mistakes. These quirks leave fingerprints in the statistics of our writing.

AI detectors try to measure that difference. They ask, in effect: "How much does this text look like what a language model would naturally produce?"

Signal 1: Perplexity

Perplexity is a measure of surprise. Take a sentence and feed it, word by word, to a language model. At each step, the model predicts the next word. If the actual word is one the model expected, surprise is low. If it is unusual, surprise is high.

• Low perplexity: the text is full of predictable choices. Example: "The weather today is sunny and warm."
• High perplexity: the text contains unexpected choices. Example: "The weather today is sulking like a teenager."

AI-generated text often has low perplexity because the model that produced it favored likely words. But plenty of human writing is also low-perplexity: instructions, definitions, legal language, and simple sentences by careful writers. That overlap is one of the main sources of false positives.

Signal 2: Burstiness

Burstiness looks at variation across a passage rather than within a single sentence. Humans tend to write in bursts: a long, winding sentence followed by a short one. Then maybe a fragment. AI models often produce sentences of similar length and structure, one after another.

Detectors measure how much sentence length, complexity, and perplexity fluctuate across a text. Very even text nudges the score toward AI. Highly varied text nudges it toward human.

Again, this is a tendency, not a rule. A student writing a structured lab report will produce even, consistent sentences on purpose.

Signal 3: Trained classifiers

Modern detectors rarely rely on simple formulas alone. Most use a classifier, which is a machine learning model trained to tell two categories apart.

The training process looks roughly like this:

1. Collect examples. Gather large amounts of text known to be human-written (essays, articles, forum posts) and text generated by various AI models.
2. Label them. Mark each example as human or AI.
3. Train the model. The classifier learns which features, often thousands of subtle ones, distinguish the two groups.
4. Test and calibrate. The developers check performance on examples the model has not seen and set thresholds for what counts as "likely AI."

The classifier may use perplexity and burstiness as inputs, along with vocabulary patterns, punctuation habits, and features no human would think to name.

Pengesan AISkor AI di peringkat ayat

Signal 4: Vocabulary and phrasing patterns

Language models have favorite words and phrases. Certain transitions, abstract nouns, and grand statements appear more often in AI text than in typical human writing. Detectors can learn these preferences.

This is also why AI-sounding phrases weaken writing in general. A paragraph that leans on vague, important-sounding words says less than one built on specifics, whoever wrote it.

From sentences to a document score

Many detectors, including ZeroTell's, score text at the sentence level and then combine those results into a document-level number. This approach has two advantages:

• Transparency. You can see which sentences raised the score.
• Mixed documents. If only one section was AI-generated, sentence scoring can show that instead of blending everything into one number.

The way sentence scores are combined varies by tool. Some average them, some weight longer sentences more, and some apply extra logic for clusters of flagged sentences. That is one reason different detectors can produce different numbers for the same text.

Why detectors make mistakes

Every step above involves estimates, so errors are built in.

There is also a built-in trade-off. Set the threshold low and the tool catches more AI text but flags more humans. Set it high and it protects humans but misses more AI. No setting removes both kinds of error. For a fuller discussion, see how accurate AI detectors are.

What detectors do not do

It's easy to imagine detectors as more powerful than they are. They do not:

• Search the internet for the "original" AI conversation
• Access the chatbot's logs or your account history
• Know anything about the writer, the assignment, or the drafting process
• Prove intent or misconduct

They look only at the text in front of them. That is why human judgment, drafts, and conversation remain essential.

A worked example

Consider two short passages about the same topic. The first is typical chatbot output: "Regular exercise offers numerous benefits for both physical and mental health. It can improve cardiovascular fitness and reduce stress levels." The second is from a person: "I started running to fix my knees. What it actually fixed was my sleep, which surprised my doctor more than me."

A detector would likely score the first as more AI-like. Each word is highly predictable, the sentences are similar in length, and the vocabulary is generic. The second has a personal angle, an unexpected turn, and uneven rhythm. But notice that a person could easily have written the first passage too, perhaps in a health leaflet. The detector sees patterns, not people.

How plagiarism checkers differ

People often mix up AI detectors and plagiarism checkers. A plagiarism checker compares your text against a database of existing sources to find matching passages. An AI detector looks for statistical patterns without comparing to any specific source. AI-written text is often "original" in the plagiarism sense, so a plagiarism checker may not flag it at all. Our guide on how plagiarism checkers work explains the other side.

Using this knowledge well

If you are a writer, the lesson is not to game the statistics. It is that vague, uniform, phrase-heavy writing is weaker writing. Specific examples, honest opinions, and natural variation make your work more readable and more clearly yours.

If you are a reviewer, the lesson is to read the highlights, consider the genre and writer, and treat the score as one piece of evidence.

Key takeaways

• Detectors estimate how closely text matches AI writing patterns.
• Common signals include perplexity (predictability), burstiness (variation), vocabulary habits, and trained classifiers.
• Sentence-level scoring shows which parts drove the result.
• Errors come from short text, formulaic genres, second-language writing, editing, and new models.
• A detector result is a probability, never proof of authorship.