How our score works, and its limits

Plain language. Read this before you rely on a score.

What the score is

The TrueScribe score is our own estimate of how likely it is that a text was written by an AI, from the low end (reads like a person) to the high end (reads like an AI). It is a probability, not a verdict. A high score does not mean that most of the text is AI. It means the writing has patterns that are common in AI text, and that texts like it are AI-written more often than not.

How it is made

• The text is split into passages and sentences.
• A language-model classifier we trained, and a second statistical check on how predictable the wording is, each give a view. A small calibrated model combines them into the score.
• Sentences that push the score up are highlighted, with the reason, so you can judge for yourself instead of trusting a number.
• It all runs on our own server. Your text is not sent to another company for detection.

Where it gets things wrong

• Short texts. Under about 100 words there is too little to go on. Treat the score as a rough hint.
• Formal and technical writing. Legal, scientific and templated writing is predictable, which looks like AI. People get flagged for this.
• Non-native English writers. Simpler vocabulary and set phrases can look machine-like. This is a known problem for every detector, ours included, which is why we measure it separately and publish it.
• Edited AI text and AI-assisted drafts. Light editing can lower a score. Heavy editing makes the question meaningless.
• New AI models. A detector is tested against the models that existed when it was tested. Newer ones may behave differently.

What to do with a score

Use it as one signal, together with drafts, notes, a conversation and the person's usual writing. Never use a score alone to accuse someone. If you are the one being scored, you can report a wrong result with the feedback button.

How we know how accurate it is

We do not ask you to take our word for it. We test on a fixed, versioned set of texts that we do not tune on, and we publish the results, including the failures, on the Benchmark page. If a number is not there, we have not measured it, and we will not claim it.

The humanizer

The humanizer rewrites text to sound more natural while keeping the meaning, names, numbers, quotes and citations. It does not add typos or hidden characters. We do not guarantee that any detector will or will not flag the result. We measure how much a rewrite changes our own score and how well it keeps the meaning, and publish that too.