Binoculars
A research Python package that scores whether text was LLM-written, using two language models and no training data.
Research implementation of a detector that compares two LLMs, with a Gradio demo and fixed threshold.
Binoculars is the code behind a research paper on spotting machine-generated text. It is for researchers and technically able educators who want to experiment with a detector, not for anyone who needs a verdict to act on.
What it is
It scores text by comparing outputs from two related decoder-only language models, so it needs no training data and works across domains. You install the Python package, call compute_score or predict on a string or list of strings, and get a score plus a label such as "Most likely AI-Generated". A Gradio demo is included.
Who it's for
It suits researchers, journalists and academic integrity teams who can run Python and want to study detection methods. The README says it is for academic purposes only and not a consumer product, and warns against using any detector without human supervision.
Key features
Zero-shot detection
Needs no training data and is described as domain-agnostic.
Two-model scoring
Scores text using Falcon-7B and Falcon-7B-Instruct by default; other scoring models can be passed in.
Fixed global threshold
Ships with one threshold chosen using the Falcon-7B pair for classifying text.
Batch input
Pass a list of strings to compute_score or predict to score many samples at once.
Gradio demo
Run app.py to try detection in a simple browser interface.
Up and running in 3 steps
Our honest verdict
What's great
- It is a published research method with a linked paper, so you can read exactly how scores are produced.
- No training data or fine-tuning is required to run it.
- You can swap in different scoring models through an argument to the Binoculars class.
- Batch scoring is built in for lists of passages.
Where it falls short
- The README states it is academic code, not a consumer product, and it should not be used without human supervision.
- It is more proficient on English than other languages, according to the README.
- Running two 7B-class models needs a capable GPU and Python skills; the README gives no hardware table.
- The threshold was tuned for Falcon-7B scoring models; changing models means rethinking the threshold.
Pricing & license
Binoculars is free under the BSD-3-Clause licence. The only cost is running it yourself on hardware that can hold the scoring models; the README shows no paid plan.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I use Binoculars to prove a student used AI?
No. The README says no detector is perfect and strongly cautions against using Binoculars or any detector without human supervision.
Does it work for languages other than English?
It can, but the README says it is more proficient on English text.
Is there a hosted version?
The README links a Hugging Face demo space. The repository itself is code you run yourself.
