How do AI detectors work?
AI detectors analyze writing patterns, mainly perplexity (how predictable each word is) and burstiness (how much sentence length and rhythm vary). AI text tends to be smooth and predictable, so detectors flag low perplexity and low burstiness as machine-written. They estimate a probability, not certainty of authorship.
Why — the first-principles explanation
An AI detector never sees who typed the text. It only sees the words, so it has to reverse-engineer the style of a machine. The core idea is that large language models generate writing by repeatedly choosing a highly probable next word. That makes their output statistically smooth. Detectors quantify this with two measurements.
Perplexity measures how surprised a language model is by each word. If the next word is exactly what the model would have guessed, perplexity is low. Human writing tends to include odd word choices, tangents, and quirks that raise perplexity; AI writing stays predictable, so low perplexity is read as a machine signal. Burstiness measures variation across sentences: humans mix long and short sentences unevenly, while AI often produces uniform, evenly-paced sentences. Low burstiness points toward AI.
Modern detectors go beyond these two by training their own classifier models on huge sets of labeled human and AI text, learning subtler fingerprints than perplexity alone. But the fundamental limitation never goes away: predictability is a correlate of AI authorship, not a proof of it. A plain, careful human writer produces low-perplexity, low-burstiness prose that looks machine-made, while a paraphrased AI draft can be roughened up to look human. That's why detectors output a percentage or probability and why vendors warn the score is a signal, not a verdict.
An example that makes it click
Imagine guessing whether a song was played by a drum machine or a human drummer, without seeing either. A drum machine keeps perfect, even time. A human speeds up, drags, and adds little imperfections. So you listen for 'too perfect' timing and call that a machine.
AI detectors listen for 'too perfect' writing: word choices that are always the expected ones, sentences that march at an even length. Smooth equals suspected machine. But a very steady human drummer can fool you, and so can a very plain human writer.
Key facts
- Perplexity measures how predictable each word is; low perplexity suggests AI authorship.
- Burstiness measures variation in sentence length and structure; low burstiness suggests AI.
- Modern detectors also train classifier models on labeled human vs. AI text for subtler patterns.
- Predictability correlates with AI writing but does not prove it, so detectors output probabilities.
- Turnitin needs 300+ words because short text has too little statistical signal to score reliably.
▶ The 60-second explainer (script)
How do AI detectors work? They never see who wrote your text, so they analyze the writing itself for patterns that machines tend to leave behind. The main clue is called perplexity: how predictable each word is. AI models pick the most likely next word over and over, so their writing is smooth and predictable, which means low perplexity. Detectors treat low perplexity as a machine signal. The second clue is burstiness: humans mix long and short sentences unevenly, while AI tends to write in even, uniform sentences. Low variation points to AI. Newer detectors also train their own models on labeled human and AI examples to spot subtler fingerprints. But here's the key limit: predictable writing only correlates with AI, it doesn't prove it. A plain human writer can look machine-made. That's why detectors give you a probability, not a certain answer.
What authoritative sources say
People also ask
What is perplexity in AI detection?
Perplexity measures how predictable each word is to a language model. Low perplexity means the words are easy to guess, which detectors associate with AI writing.
What is burstiness?
Burstiness is how much sentence length and structure vary. Humans write unevenly; AI tends to be uniform, so low burstiness suggests machine authorship.
Do detectors read the actual meaning?
Not really. They analyze statistical writing patterns, not whether ideas are true or original, which is why fabricated but smooth text can still look human.
Why do detectors need a minimum word count?
Short passages don't give enough statistical signal. Turnitin, for example, needs at least 300 words before it will produce a score.