The core insight: machines write predictably
A language model generates text by repeatedly choosing a likely next word. That is its entire job, and it leaves a fingerprint: machine text is, on average, more predictable than human text. Detection inverts the process. Run a text through a language model and measure how surprised the model is by each word. Low surprise across a whole passage suggests a machine chose those words; spiky, irregular surprise suggests a human did. That measure of surprise is called perplexity, and it is the first pillar of every detector.
Burstiness: the rhythm tell
The second pillar is variation. Humans write in bursts: a seven-word punch, then a forty-word wander, then a fragment. For emphasis. Models smooth that out, producing sentences that cluster around a comfortable mean length with comfortable structure. Burstiness measures that variation, and low burstiness is one of the most reliable machine tells in unedited AI prose. You can feel it yourself when reading: machine text has a metronome under it.
Classifiers: learning the rest
Modern detectors do not stop at two statistics. They train a classifier on large corpora of known-human and known-machine text, letting it learn thousands of weaker signals: vocabulary distribution, punctuation habits, discourse-marker frequency, paragraph shape. The classifier outputs a probability, which is what the 0 to 100 dial on this site displays. The strength of this approach is breadth; the weakness is that it learns the average machine and the average human, and individual writers are not averages.
Watermarks: the future that has not arrived
A watermark would change the game: the generating model subtly biases its word choices according to a secret key, and a verifier with the key can prove generation cryptographically. The research exists, and image generators ship related schemes like SynthID. But there is no public, verifiable watermark in mainstream chatbot prose today, no cross-vendor standard, and trivial paraphrasing weakens most proposed schemes. Until that changes, text detection remains statistical, which means probabilistic, which means fallible.
Where it all breaks
Four failure modes account for most detector mistakes. Edited AI text: human revision injects irregularity back, and the text genuinely becomes mixed authorship; the dial honestly reads it as such. Formulaic human writing: legal prose, lab reports, cover letters and academic register are uniform by design and read machine-like. Non-native English: writers working in a second language use safer structures and higher-frequency words, and peer-reviewed research has documented sharply elevated false positive rates on their writing. Short samples: statistics need sample size, and below a couple hundred words every tool is guessing. This is why our scan has a 120-character floor and why scores between 40 and 69 are labeled inconclusive rather than rounded to a verdict.
A worked example
Take a single paragraph and watch the instruments read it. The sample: a five-sentence introduction about renewable energy, written by a current chatbot and left unedited. Perplexity first. The scoring model reads the passage word by word and finds almost every choice unsurprising: after the words solar and wind, the model expected power, and power arrived. Across the paragraph the surprise stays low and, crucially, flat: no word spikes. Burstiness next. The five sentences run 18, 21, 19, 20 and 19 words, a metronome a human writer almost never sustains. The classifier then folds in its thousands of weaker signals: the paragraph opens with a definition, closes with a summary clause, uses two stock transitions, and never once interrupts itself. Each signal is weak alone. Stacked, they push the probability high, and the dial reads 88.
Now give the same paragraph ten minutes of honest human revision: one sentence cut to six words, a concrete example with a number in it, a clause that doubles back to qualify a claim. Perplexity spikes where the example appears, the sentence lengths now run 6 to 31, and the classifier's stock-pattern signals drop out one by one. The dial reads 41: the bottom of the inconclusive band, which is the statistically correct description of a text that is now genuinely mixed authorship. That movement, repeated across millions of texts, is the entire detection story, including the part where determined revision walks any text out of reach.
A small glossary
Perplexity: how predictable a text is to a language model; low values suggest machine generation. Burstiness: variation in sentence length and structure; humans are uneven, models are smooth. Classifier: a model trained on labeled human and machine text that outputs a probability. False positive: human writing wrongly flagged as AI, the costliest failure mode. Watermark: a deliberate statistical signature embedded at generation time, verifiable with a key; not present in mainstream chatbot prose today. Inconclusive band: the score range where an honest tool admits it cannot tell; on this site, 40 to 69. Keep these six terms straight and you can read any vendor's documentation, including ours, critically.
Each model generation writes more naturally, eroding the statistical gap detectors measure. Expect inconclusive bands to widen industry-wide over time. A detector that never says I do not know is not more advanced. It is less honest.