Detection you can trace back to a formula, not a model
Most AI detectors are themselves machine learning classifiers: a neural network trained on labeled human and AI text, producing a score nobody, including the people who built it, can fully explain. EyesOnText takes the opposite approach. Every signal is a named statistic with a threshold you can inspect, and every correction is a logged, reversible edit.
1. Detect: four independent analyzers
Each analyzer scores a different property of the text and returns evidence, not just a number. The composite score is a weighted combination, and the weights themselves are visible and tunable per preset.
Readability variance
We compute a Flesch reading ease score per paragraph, not just once for the whole document, and measure how much it varies. AI text tends to sit in a narrow band and stay there across paragraphs; human writing swings between dense and simple as the writer's attention shifts.
Vocabulary diversity
Type-token ratio, Shannon entropy over the word-frequency distribution, and Zipf-law compliance together describe how varied and how natural a text's word choices are. AI-generated text systematically underuses rare words and overuses common ones, which shows up as lower entropy and a distribution that deviates from the Zipf curve real writing follows.
Sentence burstiness
We take the standard deviation of sentence length across the document and cross-reference a curated list of AI-telltale phrases and punctuation habits (certain transition words, a spike in em dashes, formulaic list openers). Low burstiness, meaning every sentence is a similar length, is one of the strongest and most consistent AI signals we have.
Stylometric fingerprint
Burrows' Delta, a forensic authorship-attribution technique originally developed to settle authorship disputes, measures deviation in function-word frequencies (the, of, but, which, and so on) from a human-writing baseline. It is adapted here for AI-versus-human attribution rather than author-versus-author, and it catches patterns the other three analyzers miss.
2. Correct: six sequential steps
When correction is enabled, flagged text passes through six steps in order: spelling, style issues, AI-specific phrases, passive voice, vocabulary repetition, and sentence-uniformity smoothing. Each step is applied and logged individually, so you can see exactly which step changed which sentence and why.
3. Verify: re-detection with automatic rollback
After correction, detection runs again on the corrected text. If any analyzer's score got worse instead of better, the responsible step is rolled back automatically rather than shipping a change that made the text look more, not less, AI-generated. What you get back is the actual before-and-after, not a promise.
Why rules instead of a model
A trained classifier degrades as the models it was trained against evolve, and it cannot tell you why it flagged a sentence. A regex, a formula, or a statistical threshold does not drift, and every one of them can be read, audited, and adjusted. The tradeoff is that we will never claim 100% accuracy against every possible phrasing, and we do not: we show you the evidence and let you make the call.