Methodology

We don't detect AI.
We verify humanness.

LLMs can mimic any written output. They cannot mimic the behavioral fingerprints of a real human being sitting at a keyboard.

97.3%
Human detection accuracy
< 1%
False positive rate
~5 min
Average test duration

The problem

Why LLMs beat every traditional filter

Traditional bot detection looks for artifacts: known bot IPs, malformed headers, request patterns, missing user agents. All of this is learnable.

Large Language Models don't just pass keyword scans — they write polished cover letters, tailor resumes to job descriptions, and answer behavioral interview questions in context. They are not bots in the traditional sense. They are statistical mirrors of human output.

CogQuiz doesn't try to identify AI. It verifies that a real human completed the application — by measuring things that are trivial for people and impossible to fabricate at scale.

Test design

Four measurement layers

01 — Typing biometrics

Humans type with irregular micro-timing. A person hits 40–150ms of random variance per keystroke. Bots: zero variance, or perfectly uniform intervals.

Signal: Inter-key variance, error-correction bursts, pause curvature.

02 — Temporal & social reasoning

"March. What month is 7 months from now?" AI computes it. Humans feel it — contextually, imprecisely, naturally. We probe reasoning paths, not just answers.

Signal: Calendar logic, sequence reasoning, absurdist humor comprehension.

03 — Emotional nuance

"Thanks for getting back to me I guess." Every AI model reads this as neutral or positive. Humans hear the edge. We test tone classification under genuine ambiguity.

Signal: Sarcasm detection, implied meaning, subtext interpretation.

04 — Behavioral confidence

AI is always "certain." Humans show variance — hesitation, self-correction, backtracking. Overconfidence under ambiguity is a reliable AI signal.

Signal: Confidence variance, answer revision patterns, response time spread.

Privacy

What we collect. What we don't.

We store

  • +Keystroke timing patterns (not typed content)
  • +Mouse movement curvature
  • +Answer text and response timing
  • +Confidence ratings per question
  • +Aggregate behavioral scores

We never store

  • Names, emails, or any PII beyond session
  • Typed keystroke content (only timing)
  • IP addresses beyond session lifetime
  • Device fingerprints
  • Resume or cover letter content

See it in action.

The demo takes 5 minutes. No account needed.