And not out of modesty: such a detector cannot be validated, and one that gets it wrong is worse than none because it gives confidence where there is none. What is measured here is emptiness, which is a property of the text and not of its author — a person can write padding and a model can write something substantial. And it turns out to be more useful, because the real question for someone who receives a document is not «who wrote it?» but «is it worth reading in full?».
Ryan Roslansky told how he received a document made with AI and, rather than read it, passed it to another AI to summarise: the summary had not a single new idea. That is a method, not an impression, and here it is measured: it looks for the smallest set of paragraphs that covers everything the text says, and what is left over is what you can throw away without losing anything. That is the big number.
The tools that already exist look for words: flowery vocabulary, stock phrases, repeated bigrams. That catches padding that repeats LITERALLY. But a language model's padding repeats the same idea in different words, and there a dictionary does not reach: in the automatic-summarisation literature, semantic methods detect up to 90% of redundancy against 19% for lexical ones. That is why meanings are compared here, not letters — and why a model of 150 MB has to be downloaded once.
Raise it and it will only mark paragraphs that are near-copies; lower it and it will start accusing paragraphs that share a topic but say different things. There is no good value, there is a trade-off — the same one as the camera's confidence threshold in Local AI camera. A warning about the numbers: this model compresses similarities, so do not expect 90% between two paragraphs that say the same thing. Measured with control texts, blatant paraphrases land at 62-69% and paragraphs with different facts below 55%; that is why the cut sits at 60. Move it and watch what gets struck out before trusting anything.
«Sales fell 12% in March according to Nielsen» versus «it experienced a significant evolution over the period». You do not need to understand the text to see the difference: it is enough to count figures and proper nouns per 100 words. It is the silliest signal of all and one of the most reliable. Watch one detail: the first word of each sentence does NOT count as a proper noun, because its capital is spelling and not information.
The model downloads once —it is 150 MB, the same one already used by the embeddings, similarity and RAG tools— and from then on everything happens in your browser. It matters more than it seems: what gets pasted here is usually an internal document somebody sent you, and those do not get uploaded to somebody else's server to be scored.
This does not tell you whether an AI wrote it, and not out of modesty: such a detector cannot be validated, and one that gets it wrong is worse than none. It measures emptiness, which belongs to the text and not to its author — a person can write padding and a model can write something substantial. The question it answers is «is it worth reading in full?».