How do AI detectors work?
Almost every AI detector leans on two ideas: how predictable each word is, and how much the writing varies. Play with both below.
1. Predictability (perplexity)
A language model writes by picking a likely next word, over and over. So text that a model wrote tends to be made of words the model itself finds likely. A detector runs the text back through a language model and asks, at every position, “how surprised am I by this word?”
Average that surprise over the passage and you get perplexity. Low perplexity – no surprises – points towards a machine. People pick odd words, make jokes, name their cat, misspell things. All of that raises perplexity.
Try it: build the sentence with the safest words, then again with strange ones, and watch the number.
Build a sentence. Pick each next word.
The results of the study were ▍
Running perplexity
–
Percentages are illustrative, not from a specific model. Perplexity = e raised to the average negative log-probability of the chosen words: lower means more predictable.
2. Variation (burstiness)
People write in bursts. A short punchy line. Then a long sentence that wanders through a thought, doubles back, adds a bracket and finally lands. Models tend to keep every sentence at a similar, comfortable length and similar predictability.
Burstiness measures that variation. The trace on the right is VerifyWrite’s rhythm trace: sentence lengths drawn as a line. Human writing looks like a seismograph during a tremor; model writing looks like a quiet night.
VerifyWrite measures burstiness directly, in your browser, alongside stock vocabulary, contractions, transitions, punctuation variety and paragraph uniformity. It does not run a full language model for perplexity; the AI second opinion adds a model’s reading when you want one.
- Variation (CV)
- 0.67
- Shortest
- 2
- Longest
- 23
Short bursts (“Small win.”) sit next to long, rambling sentences. The trace jumps around.
3. Trained classifiers
Commercial detectors add a classifier: a model trained on millions of labelled examples of human and AI text that learns whatever features separate them. This can be more accurate on the kind of text it was trained on, but it’s a black box, and it degrades when new models or new kinds of writing appear.
4. Watermarks
A different approach is to mark text as it is generated, by nudging word choices in a pattern only the provider can check. Google’s SynthID does this for Gemini output. It only works for text from a cooperating provider, and heavy editing weakens it.
What that means for you
All of these signals are statistical. A formal human writer, a non-native speaker, or someone following a strict essay template can look “predictable”. An edited or prompted AI text can look “human”. That’s why a detector score should start a conversation, not end one – see how accurate AI detectors are.
Questions people ask
How do AI detectors work, in one sentence?
They measure how predictable a text is to a language model and how evenly its sentences are built, sometimes combined with a classifier trained on known human and AI examples, and turn that into a probability.
What are perplexity and burstiness?
Perplexity measures how surprised a language model is by each next word; lower means more predictable. Burstiness measures how much that predictability and sentence length vary across a text. AI text tends to be low on both.
Why can AI detectors be fooled?
Because the signals are statistical. Asking a model to vary sentence length, adding personal details, or lightly editing the output raises perplexity and burstiness. Paraphrasing tools were shown to defeat several detectors in a 2023 University of Maryland study (Sadasivan et al.).
Do detectors look for watermarks?
Some research systems embed statistical watermarks in generated text, and Google has deployed SynthID text watermarking for Gemini. Most text in the wild carries no watermark, so general detectors still rely on style statistics.
Try it on real text
Your text follows you between tools, so you can paste once and check everything.