Module B: Statistical detector
Measures how predictable your text is to two small language models. Machine text tends to be the kind of thing a model would have written next.
How it works and its limits
A zero-shot detector built from two small, real, off-the-shelf language models (gpt2 and distilgpt2): no training and no dataset, just how predictable the text is to them. Inspired by Binoculars (Hans et al., 2024) and GLTR (Gehrmann et al., 2019), independently reimplemented at a much smaller scale and recalibrated for it. Design notes.
It shares the same base-model blind spot as every detector of its kind. Read the limitations.
or try
Results appear here: a verdict with its five measurements, every token shaded by how strongly the model predicted it, and a sentence-by-sentence breakdown.