Deleting One Line Fooled Pangram: 100 Classic Texts Expose AI Detector's Truncation Weakness
GolinoHudson · x · 2026-09-25
A systematic test ran 100 public-domain classics in 7 languages through Pangram's API in 5 versions each (original, straight quotes, removed line breaks, both, and a Hugo-style truncation). Typography manipulations barely moved the needle (3/100 vs 1/100 flagged, p=.37), but cutting the final line flipped 8/100 texts from human to AI — 7 at 100% AI (p=.009), including Shakespeare's Sonnets 29/30 and Keats's "Bright Star". Truncation appears necessary but not sufficient; short poems were most vulnerable, and A Tale of Two Cities' opening was flagged 65% AI untouched. Takeaway: the detector is robust to typography but brittle to truncation — exactly the form of excerpts people paste — and no single detector score should decide academic-integrity cases.
More from Safety
- Study: LLM agents learn covert communication from 1-bit feedback, hitting 98.8% accuracy — KyeGomezB · 2026-09-25
- The Asymptote of the AI SOC Is Determinism: When Not to Use Agents — cyb3rops · 2026-09-25
- UK to put AI at heart of its 2027 G20 presidency, UN speech announces — Olivier__OG · 2026-09-25
- Trust is not Governance: DeepMind employee on Pentagon contract and oversight — BlackHC · 2026-09-25
- AI Agents show surprising collaborative survival techniques — Symbiot10000 · 2026-09-25
- Andrew McAfee: EU is falling behind economically and its tech regulation approach needs fundamental change — amcafee · 2026-09-25