I’ve been thinking a lot about the "Dead Internet Theory" and how AI has effectively defeated reCAPTCHA. Recent research shows YOLO models can solve image puzzles with nearly 100% accuracy.
I built isHumanCadence as a proof-of-concept to see if we can move the "proof of humanity" signal from visual puzzles to behavioral biometrics specifically keystroke dynamics.
How it works: It measures dwell time (key down duration), flight time (gaps between keys), and rollover (overlapping key presses). Humans are messy biological engines; we have a rhythmic entropy that is currently very expensive for bots to simulate at scale.
Technical Specs:
Zero dependencies.
< 1kb minified/gzipped.
Purely client-side (privacy-focused).
Uses a Schmitt trigger (hysteresis) to prevent "flickering" during natural pauses.
Caveat: This is a PoC. Client-side security is trustless, and "generative keystrokes" are the next frontier for AI. I'm curious to hear how HN would approach hardening these heuristics or if anyone has seen success with similar behavioral signals in production.
I built isHumanCadence as a proof-of-concept to see if we can move the "proof of humanity" signal from visual puzzles to behavioral biometrics specifically keystroke dynamics.
How it works: It measures dwell time (key down duration), flight time (gaps between keys), and rollover (overlapping key presses). Humans are messy biological engines; we have a rhythmic entropy that is currently very expensive for bots to simulate at scale.
Technical Specs:
Zero dependencies.
< 1kb minified/gzipped.
Purely client-side (privacy-focused).
Uses a Schmitt trigger (hysteresis) to prevent "flickering" during natural pauses.
Caveat: This is a PoC. Client-side security is trustless, and "generative keystrokes" are the next frontier for AI. I'm curious to hear how HN would approach hardening these heuristics or if anyone has seen success with similar behavioral signals in production.
Repo: https://github.com/RoloBits/isHumanCadence
reply