←back to thread

204 points pabs3 | 1 comments | | HN request time: 0.207s | source
Show context
timshell ◴[] No.44093768[source]
I'm a founder in this space (www.roundtable.ai; YC S23)

Behavior is a big missing link. Many CAPTCHA services (including Google reCAPTCHA v3) claim to use behavioral analyses, but you can disprove this using Operator to fill out a form and see reCAPTCHA and other bot detection systems flag it as a human.

At Roundtable, we rely on first-order behavioral markers (keystroke, mouse, scroll, click) etc. When first-order are sufficiently spoofed, analyze higher-level cognitive traits (e.g. incongruent effect in Stroop)

replies(2): >>44094474 #>>44095197 #
_rami_ ◴[] No.44094474[source]
(Author here) Interesting! How do you differentiate between a bot and a screenreader user? Both won't move their mouse, scroll, etc but only send a single click event
replies(1): >>44099238 #
1. timshell ◴[] No.44099238[source]
Yup, we had this with open-end analysis (see: https://www.producthunt.com/stories/how-to-detect-ai-content...)

We had some people use voice software to fill in survey responses. This flagged computer-generated assistance.

Generally speaking, it'll be more than a single click event (e.g. see https://x.com/_magrawal/status/1925985289568211168)