AI Detector AI Humanizer ChatGPT Detector AI Content Detector AI Humanizer Detector Accuracy Private Detector How it works Tools

Methodology

How AI Detection Works

Perplexity, burstiness and token-level signals — in plain English.

By the Astra team · Updated 17 September 2026

Related: AI Detector Accuracy · AI Content Detector

Try the Astra AI Detector

Now that you know what the signals mean, see them on your own writing. The main Astra AI Detector returns an AI score with sentence-level highlights — free, unlimited and processed entirely in your browser.

Try the AI Detector

01In brief

The short version

AI detectors measure how predictable and uniform a piece of writing is.

AI detectors measure how predictable and uniform a piece of writing is.

Human writing tends to be uneven and occasionally surprising; machine writing tends to be smooth and statistically regular, because language models pick high-probability words and produce steady, even sentences. A detector turns those patterns into a probability that the text was AI-generated.

It doesn't read a hidden label or watermark, and it can't be certain — which is why the honest output is a likelihood, not a yes/no.

02The signals

The signals AI detectors read

Modern detectors combine several signals rather than relying on any one. Astra shows them per sentence, so a result is explainable rather than a black box:

Perplexity

How predictable each word is to a reference language model. AI text keeps perplexity low and flat because it favors the most likely next word; human writing is more surprising and varied.

Burstiness

The variation in sentence length and rhythm, often measured as the spread of sentence lengths across a document. People mix long, short and fragmentary sentences; models hold a steadier pulse, so low burstiness reads as machine-like.

Token-level probability

Word-by-word likelihood given the words before it — this drives the highlight “heat” you see in the tool, marking which words fit the machine pattern most closely.

Semantic & lexical similarity

Meaning- and wording-level patterns that help catch paraphrased and rewritten AI, where the words changed but the structure didn't.

Formatting & phrasing fingerprints

Repeated sentence openers, heavy use of transitions like “Furthermore” and “In conclusion,” and tidy bulleted “key takeaways” that GPT-family models default to.

Sentence-level analysis

Every sentence carries its own confidence, so any flag can be traced back to the specific evidence behind it.

03Classifiers

Behind the score: classifier models

The strongest detectors train a classifier on millions of labelled samples.

The strongest detectors don't stop at perplexity and burstiness.

They train a machine-learning classifier on millions of labelled human and AI samples, extracting features like word-frequency distributions, syntax and punctuation habits, then validate on held-out data to measure accuracy and control false positives.

That's why detection quality varies so much between tools — it depends on the training data, the models covered and how the thresholds are tuned.

04Reading a score

How to read your score

Treat the percentage as a likelihood, not a verdict.

Treat the percentage as a likelihood, not a verdict.

Look at which sentences are highlighted and how strongly, weigh the context, and remember that longer passages give more reliable results than a sentence or two.

A middling score is a prompt to read carefully — not a confident answer in either direction.

05The limits

What AI detection can't do

It can't prove who wrote something — and here's why.

It can't prove who wrote something.

The signals it relies on aren't unique to AI, so formal, simple or non-native (ESL) English can be flagged as AI — a false positive — while genuinely AI text can slip through, and heavy paraphrasing or humanizing can defeat any detector. New models can also outrun detection until it's updated.

That's why Astra is designed as a transparent guide, and why we publish how accurate AI detectors are rather than a single marketing number.

06Worked example

A worked example: how a detector scores one paragraph

Two paragraphs, same topic — here is what a detector actually computes for each.

Imagine pasting two paragraphs on the same topic, one drafted by a chatbot and one written by a person, and watching a detector score each. Under the hood, it is doing two simple things.

First, perplexity. The detector feeds the text through a reference language model one word at a time and asks, how predictable was that word? In a machine sentence like The results clearly demonstrate that the proposed method is effective, almost every word is the model's top guess, so perplexity stays low and flat. A person might write The numbers were a mess, honestly, until we re-ran the third batch — mess, honestly and third batch are low-probability swerves that push perplexity up.

Second, burstiness. The detector measures how much sentence length and rhythm vary. A human paragraph might run 8, 24, 5, 31, 11 words — a jagged spread. An AI paragraph tends to cluster, say 17, 18, 16, 19, 17 — a narrow spread with a low standard deviation. Low variation reads as machine-like.

Neither number is proof. A carefully edited essay can be flat and predictable, while a chatbot told to write casually can look bursty. That is why a detector blends these signals, weighs them across the whole passage, and reports a probability rather than a verdict — and why the illustrative figures above are patterns, not fixed thresholds.

07Zero-shot

Zero-shot detection: reading a model's own probabilities

Not every detector is trained on labelled data; a whole family reads a model's raw probabilities instead.

Broadly, detectors fall into two families. Trained classifiers, covered above, learn from millions of labelled human and AI samples. The other family is zero-shot: it needs no training set at all, and instead reads a reference model's own probability estimates.

Perplexity thresholding is the simplest version: compute average predictability and flag text that is unusually smooth.

Log-rank and log-probability tests — the approach behind early tools like GLTR — look at where each word sits in the model's ranked list of likely next words. Machine text leans heavily on top-ranked words.

Probability curvature — the idea behind DetectGPT — slightly rewrites the passage many times and checks whether those edits make it less likely under the model. AI text tends to sit on a distinctive peak; human text does not.

Zero-shot methods are transparent and cheap, which is why a lightweight statistical read can run entirely in your browser, as Astra does. The trade-off is sensitivity: they lean on one reference model, so text from a very different or newer model can slip past. Trained classifiers cover more models but need constant retraining and rarely show their reasoning. No single family wins outright, which is one reason careful tools blend several signals rather than trust one number.

08Watermarking

Watermarking: the signal some AI text is starting to carry

Some AI text now carries a deliberate hidden marker — here is how it works, and why it is not a cure-all.

Perplexity and burstiness read patterns that AI text happens to leave behind. Watermarking is different: the AI provider deliberately plants a hidden statistical signal while the text is being generated, so it can be recognised later.

The common approach quietly biases word choice. As the model generates, a secret key splits the vocabulary into a 'green' list and a 'red' list at each step and nudges the model toward green words. The shift is far too subtle to notice while reading, but a checker holding the key can measure an unusually high share of green words and flag the text. Google DeepMind's SynthID-Text is a production example of this idea applied to text.

Watermarking sounds like a silver bullet, but it has hard limits for a general-purpose detector:

It only exists if the tool that wrote the text switched it on — most everyday chatbot output is not watermarked.

Paraphrasing, translating or heavy editing dilutes or erases the signal.

Each scheme needs its provider's key, so no independent tool can read every watermark.

That is why Astra, like most browser-based detectors, relies on statistical signals you can inspect rather than on watermarks you cannot. Watermarking is a promising complement to detection, not a replacement for it.

09Under the hood

What AI detectors are not doing under the hood

Half of understanding a score is knowing what the detector never had access to.

A lot of confusion about detectors comes from imagining machinery that is not there. Knowing what a detector does not do makes its score much easier to read.

It does not read a hidden label

Unless the text is watermarked, there is no tag, metadata or invisible marker saying 'AI wrote this.' The detector only sees the words you paste.

The model is not confessing

Detectors do not ask ChatGPT 'did you write this?' They compare your text against statistical patterns, so even a confident score is an inference.

It is not a database lookup

A detector is not matching your passage against a stored library of known AI answers; it scores the writing's own statistical fingerprint in real time.

It cannot see your prompt or history

It has no idea what you asked for, which tool you used, or how many times you edited — only the final text.

These limits cut both ways. They are why a genuinely human paragraph can be flagged and a lightly rewritten AI one can pass. The takeaway is the same throughout this page: a score is a reasoned estimate from surface patterns, best used to start a conversation about a piece of writing — never to end one.

10Questions

Methodology FAQ

Reference models, watermarks, perplexity, burstiness and why tools disagree.

Does an AI detector run the same model that wrote the text?

No. A detector uses its own reference model or trained classifier to judge the text's statistical patterns; it does not need, or usually have, access to the specific model that produced it. That is also why a brand-new model can evade detection until the detector is updated.

Can an AI detector see my prompt or where the text came from?

No. It analyzes only the words you paste, with no visibility into your prompt, the tool you used, or your edit history. Everything in the score is inferred from the text itself, and in a browser-based tool like Astra that text never leaves your device.

What is a watermark in AI-generated text?

A watermark is a hidden statistical signal that an AI provider can embed while generating text, usually by subtly biasing word choice, so a checker with the right key can recognise it later. It only works when the generator enabled it, and paraphrasing or editing can erase it, so most detectors do not rely on it.

Why can two AI detectors give the same text different scores?

They use different reference models, features and score thresholds, and they were trained or tuned on different data, so they weigh the same signals differently. Disagreement is normal, and it is a good reason to treat any single score as one opinion rather than a fact.

How do AI detectors work?

They measure how predictable and uniform text is (perplexity and burstiness), analyze word-level probabilities, and usually run it through a classifier trained on millions of samples, then output a probability that it was AI-generated. It's statistical inference, not a hidden label.

What is perplexity?

A measure of how surprised a language model is by each word. Low, flat perplexity suggests machine-generated text; human writing is more surprising and variable.

What is burstiness?

The variation in sentence length and rhythm across a document. Human writing varies a lot; AI writing tends to be more uniform, which detectors pick up on.

What signals do AI detectors analyze?

Perplexity, burstiness, token-level probabilities, semantic and lexical similarity, formatting and phrasing fingerprints, and sentence-level patterns — ideally shown per sentence.

Can AI detectors detect every model?

They cover major models like ChatGPT, Claude and Gemini but can lag when new models ship, since detection is an arms race. See the model detector.

Why do AI detectors produce false positives?

The signals aren't unique to AI — formal, simple or non-native English can look statistically similar. That's why a score is a signal, not proof — see accuracy.

11Please note

Before you act on a score

A detector result is one signal to weigh — never proof on its own.

A note on accuracy: no detector can guarantee 100% accuracy, whatever it claims. Treat the score as one signal, not proof — never make a decision that affects someone's career or academic standing on a detector result alone. Consider context, and talk to the writer.

AI-generated
H
High Confidence