Binoculars
A zero-shot method that compares how two related models score the same text, with no detector training data. Useful as an independent baseline when you do not trust a vendor's number.
What it is
Binoculars needs no training data. It runs two related language models over the same text and compares how surprised each one is; machine text tends to look predictable to one model in a way that human text does not. The method is published with code, so anyone with a GPU can run it and nobody can quietly retrain it to favour a vendor. It is a common baseline in papers and a useful independent check when a vendor score is disputed.
When to use this entry
Use it as an open, reproducible baseline detector.
Questions
Can I run Binoculars myself?
Yes. The paper is open and the authors published code.
Is Binoculars more accurate than commercial detectors?
The paper reports strong zero-shot results, but like every detector it weakens on paraphrased or short text.
Cite this page
Paraphrase Pass. "Binoculars". Checked 2026-10-31. https://paraphrasepass.com/detectors/binoculars/
Facts on this page carry the date they were last checked. Quote the date with the fact. The primary sources are linked above; cite them for the claim and this page for where you found it.