Deckrd: AI Detection

Detect AI-generated messages in your conversations

Deckrd — AI Detection for Your Conversations Know who you're really talking to. Deckrd silently analyzes chat conversations to detect AI-generated messages, alerting you when your conversation partner may not be human. HOW IT WORKS • Automatic detection on Discord, WhatsApp, Messenger, LinkedIn, and X DMs • 5 behavioral detectors analyze timing, language, coherence, style mimicry, and persona consistency • Floating widget appears when AI probability crosses your threshold • Browser notifications for high-confidence AI detection PRIVACY FIRST • All analysis happens locally in your browser • Your messages never leave your device • No message content is stored — only anonymous scores • Optional server analysis sends only anonymous signal data, never message text (future release) CUSTOMIZABLE • Adjust detection thresholds with simple sliders • Choose when the floating widget and notifications appear • Works on any chat platform with smart heuristic fallback Free to use. Pro tier coming soon with AI consensus deep analysis. ## What's new in v0.9.0003 - **Big accuracy pass.** Fixed a class of false-positive "AI detected" verdicts on conversations where Deckrd had thin signal. The score now blends toward "uncertain" when only a single detector is firing, instead of pegging to 100% confidence. - **Smarter direction detection.** When Deckrd can't tell which side of a chat a message came from, it now skips the message rather than guessing. Stops the user's own writing being graded as if it were the counterparty's. - **Cleaner behaviour on x.com profile feeds.** Deckrd no longer tries to score posts on a single author's profile or timeline, since those aren't conversations. Direct messages still work as before. - **More-stable scoring at the boundary.** A floating-point edge case was silently dropping votes from detectors that returned values near the neutral threshold. Threshold comparison is now exact. ## Fixes - Linguistic detector now requires at least three messages of inbound content before firing, matching the other content-heavy detectors. - Mirror detector no longer falls through to a "perfect mirror" reading when an adapter returns empty messages. - Confidence scaling no longer over-weights detectors that report low confidence, so weak signals don't dominate the verdict. ## Privacy No new permissions. Detection still runs entirely on your device. Message content never leaves your browser.