I’ve always been suspicious of AI detectors like GPTZero and Originality. They just feel unreliable, sometimes flagging human work as AI. Well, it turns out they’re even easier to fool than I thought!
It’s not about complex sentence structures or vocabulary. The secret is that these detectors are heavily biased against specific formatting “markers” that AIs tend to spit out. It’s a surprisingly simple flaw.
🔬 The AI Detector Traps
The author found that detectors are basically just looking for these special characters. By removing them, AI-generated text suddenly looks human.
🚫 Smart quotes (“ ”) instead of regular ones (‘ “).
🚫 En & em dashes (–) instead of a simple hyphen (-).
🚫 The special ellipsis character (…) instead of three periods (…).
🚫 Invisible characters like non-breaking spaces and zero-width spaces.
🚫 Fancy bullets (•) and full-width characters (ＡＢＣ).
The author even built a tool that scrubs these markers, and pure AI text started scoring 90-99% human. It’s an awesome way to prove how flimsy these detection methods really are.
This isn’t about promoting cheating; it’s about showing that the detectors we’re told to trust can be tricked with what’s basically a find-and-replace. It’s a total game-changer.
Check out the full Reddit post for all the details!
I found a jailbreak to bypass AI Detectors
byu/Officiallabrador in