A new interactive quiz posted to Hacker News is putting AI detection tools on blast—and the results are brutal. The site, Truly Typed (trulytyped.com/quiz), challenges visitors to distinguish human-written passages from AI-generated ones across three rounds. Each round presents two texts and asks users to guess which came from a machine. But here's where it gets interesting: the same six passages were also run through popular detection tools including ZeroGPT, QuillBot, and GPTZero. Every single one was classified as human-written, even though half of them weren't.

How the Quiz Works

The setup is elegantly simple. Participants read two passages side-by-side and click on whichever they suspect is AI-generated. The quiz provides a hint button if you're stuck, and reveals the answer after each guess with an explanation. In one round covering spaceship topics, for instance, Passage A was generated by asking Claude to write about the same subject as Passage B—which turned out to be the original human-written text. Users who picked correctly often did so through intuition rather than any detectable pattern, which is exactly the point.

Why AI Detectors Keep Whiffing

The implications are significant. If established tools like ZeroGPT and GPTZero—services specifically designed to solve this problem—can't reliably flag AI content at better-than-chance rates, what does that mean for educators, editors, and platform moderators? The quiz suggests the answer is: not much. Detection appears to be more art than science right now, relying heavily on 'vibe checks' rather than verifiable signals. This creates a genuine problem for anyone trying to maintain authenticity in written communication.

The Real Issue: Provenance Over Detection

Truly Typed isn't just trying to build a better detector—they're arguing the entire approach is flawed. Their pitch cuts deeper: instead of guessing after the fact, readers should be able to see exactly how any piece was written. 'This is what reading on the web should look like,' they state. 'Now that any text can be generated in seconds, readers deserve to see how a piece was actually written. Not a vibe-check. The actual record.' It's a provenance-first philosophy rather than a detection-after-the-fact one.

Key Takeaways

  • Current AI detectors (ZeroGPT, QuillBot, GPTZero) misclassified all six test passages as human-written
  • Human intuition performed no better than random chance in many rounds
  • The detection approach may be fundamentally flawed compared to provenance tracking
  • Truly Typed proposes transparent authorship records rather than post-hoc analysis