Writing Pattern Analyser

What this page measures, and what it cannot

A genuine AI detector works by running text through a language model and scoring how predictable each word was given the words before it. That requires the model. This page has no model, so it cannot do that, and the numbers below are not an attempt to imitate the result.

What it can do is measure the text itself: how long the sentences are, how much their length varies, how repetitive the vocabulary is, how often the passive voice appears. Those are real properties of a document, and they are worth knowing whether or not anyone is worried about authorship.

Why the widely quoted detectors are not reliable either

It is worth knowing before you trust any percentage score, including one you have seen elsewhere. OpenAI withdrew its own classifier in July 2023, citing low accuracy. A 2023 Stanford study found several widely used detectors labelled more than half of the essays written by non-native English speakers as machine-generated. Detectors tend to flag writing that is formal, plain, or produced under time pressure, because those are the qualities they were trained to notice, not because of who wrote it.

A high or low reading here should not be used to accuse anyone of anything. It is a description of a document, and documents are not people.

The figures, in plain terms

• Sentence length variation is the spread of sentence lengths, reported as a coefficient of variation: the standard deviation divided by the mean. A low figure means the sentences are all about the same length. People tend to mix a short sentence in among longer ones; a flat spread is one of the few properties with any real signal behind it.
• Lexical diversity is reported as a moving-average type-token ratio over 50-word windows. The plain ratio falls as a text gets longer, which makes a long document look less varied than a short one for no real reason. Averaging over fixed windows avoids that.
• Function-word share is the proportion of words that carry grammatical structure rather than content: the, of, and, that. A high share means fewer content words are doing the work.
• Uniform run is the longest stretch of consecutive sentences whose lengths are within 15% of each other. It tells you where the rhythm flattens out.
• Passive voice is an approximation: a form of to be followed within three words by a past participle. Adjectives ending in -ed such as tired or interested are excluded, but the estimate will still miss some passives and catch a few things that are not passives. Treat it as a rough indicator.

Reading the result

There is no threshold at which a text becomes machine-written. Every figure here describes writing that people produce constantly: legal drafting, technical documentation, exam answers, and the work of anyone writing in a second language. The useful move is to look at the figures, then read the sentences they came from, and decide whether the text says what you want it to say. That judgement is yours and no measurement replaces it.

Where to go next

If the figures suggest the wording itself needs work, the text rewriter varies sentence structure and lists every change it makes, so you can compare the two versions here afterwards. For a plain count of words, characters and reading time, use the word counter. If a passage will not paste cleanly, the invisible character cleaner removes the hidden characters that cause it.