Home › Case studies › Case 02
Case 02 · The adversarial hole
Every academic AI detector collapses under adversarial paraphrase.
The one attack that breaks GPTZero, Originality.ai, Binoculars, and FastDetectGPT — and where we score 99.6%.
The failure · on the record
The RAID paper (Dugan et al., ACL 2024) showed that adversarial paraphrase reduces every published detector’s AUC by 24–41 points. GPTZero drops to 0.62. Originality.ai to 0.71. Binoculars to 0.76. FastDetectGPT to 0.43. Our detector on German paraphrase: AUC 0.9958, TPR 99.2% at 5% FPR, published and signed in RAID.signed.json.
The numbers
What we do differently
Verifiable in-browser: press Verify on the proof page and every one of 30 configurations (including paraphrase) recomputes bit-for-bit against the signed artifact.
The signature covers the entire result payload — no cherry-picking.
Proof
Run the browser verifier. Open the verifier →
Try it yourself
Score a paraphrased AI text.
curl -X POST https://api.trigeochiral.com/detect \
-H 'Authorization: Bearer $TRIALKEY' \
-d '{"text": "Paste paraphrased AI text"}'Bounty · put money where the claim is
$500
Per paraphrased AI text we score below 0.85 probability. First 100 verified failures paid. /challenge/.
Sources