utm_source=chatgpt.com links). Runs in your browser.The result is an estimate, never proof. Short messages and edited AI text are hard for any detector.
On linkedin.com the extension checks every post LinkedIn has loaded, in view or not, and labels each one. The free checks (fingerprints, writing style) run for every post at once; then every post of 30 words or more is checked by Claude (Sonnet), one at a time from the top of the feed, at most 40 posts an hour. Shorter posts say “Too short to judge”, posts without words of their own “No text to check”; ads are checked too and tagged “Ad”. Switch it off with the On/Off button in the popup.
Privacy: with the switch on, every loaded LinkedIn post of 30 words or more is sent to the X2 Labs server (ai-text-detector.x2labs.ai) and from there to Anthropic's Claude (Sonnet), whether or not you look at it, up to 40 an hour. The server keeps nothing. Nothing else is sent. Nothing else is stored except a cache of results, kept in this browser for 7 days (at most 1 MB) and keyed by a hash of the post's text. The whole text is not stored, but a cached result can quote the post: fingerprint snippets (up to 80 characters each) and up to three Claude reasons (up to 120 characters each).
Claude's check runs on the X2 Labs server (ai-text-detector.x2labs.ai), which asks Claude Sonnet with the lab's key. Nothing to install and no key of your own. The server stores and logs nothing (no text, no result); the hosting uses your address only to limit requests, and Anthropic handles the text under its API data policy.
Right-click a message you know someone really wrote and pick Save as written by…. A comparison needs at least 5 samples and 300 words. Samples stay in this browser only.
Experimental: tested on real Enron emails of people it was not tuned on, it called 2% of their own new emails “unlike” them and caught 3% of AI rewrites of their emails; it never decides the result on its own. On web pages the card says “this person”, never the name.