Announcement
Introducing Botfilter
Botfilter flags AI generated text on websites you visit. Filter AI slop in your feeds, AI spam in your inbox, AI generated articles, and more.
Blog
Notes on Botfilter, the free browser extension that flags likely AI-written text on your device: how it works, how it was trained, and how to use it.
More posts
Technique
How we trained Botfilter's on-device AI text detector
At Balanced, Botfilter's 23.6 MB detector flagged 94 to 100% of AI-written test text by model group, and 0.2 to 1.8% of human text by source.
Published
Technique
A private AI detector that runs on your device, with no account
Botfilter scores text with a 23.6 MB model inside the extension and never uploads the text you read. On your device it saves only a hash of the text and a few numbers about it.
Published
Comparison
A free, private Pangram alternative for everyday reading
Pangram is built for checking documents you send it. Botfilter is built for reading: it marks likely AI-written text as pages load, without uploading them.
Published
Comparison
Pangram vs GPTZero vs Originality.ai on price and privacy
Pangram, GPTZero, and Originality.ai are hosted services that receive the text you check; Botfilter scores the pages you read inside your browser, with no account.
Published
Guide
How to block AI-generated posts on X, LinkedIn, and Reddit
Mutes on X, LinkedIn, and Reddit match words and accounts; Botfilter reads each post on your device and hides, folds, or outlines the likely AI-written ones.
Published
Guide
How to spot and filter AI-written email in Gmail
Gmail’s spam filter looks at senders and user reports, not at who wrote a message. Botfilter reads the writing and marks likely AI-written email on your device.
Published
Explainer
What on-device AI means, and why a small model can be enough
Botfilter's AI-text detector runs in your browser with 22.7 million parameters, about 1/167th the size of Phi-4-mini, a small LLM, and sends no page text anywhere.
Published
Explainer
Are AI detectors accurate?
AI detectors can catch most AI text they were tested on, but every published test also finds human writing flagged, and where AI text is rare, a 2% error rate makes most flags wrong.
Published