Study Finds AI Humanizers Evade Pangram Detection, and Its 15% FPR Claim Is Shaky
paulnovosad · x · 2026-09-24
A recent study shows "AI humanization" tools make it hard for Pangram to detect AI-written text, and the authors admit the 15% false positive rate attributed to Pangram for 2023-2025 likely isn't a true FPR.
paulnovosad frames it as an arms race: Pangram keeps iterating, so cheaters can't rely on today's working humanizers staying effective tomorrow. He also mocks the inconsistency of people who are deeply skeptical of AI detection but have no skepticism of humanizers.
ScottBarkowski adds that FPR/FNR confusion persists because most social media users won't put in the minimal analytical effort required.
More from Safety
- Democrats propose requiring federal permission to train AI models, with up to 20 years in prison — Polymarket · 2026-11-11
- Bill Gates says AI harms may need to be endured first, fueling fears of government control — nikola_mr64990 · 2026-11-11
- Investigator open-sources 100k+ tweets as 120 AI influence accounts get deleted — ZeroStateReflex · 2026-11-11
- AI Defense Shifts to Scale Economics: Attacks Buried in Legitimate Traffic — DavidLinthicum · 2026-11-11
- AI safety researcher hires paid, credited RA/co-writer to speed up memos and blog posts — davidmanheim · 2026-11-11
- Google DeepMind formally verifies Signal's core cryptographic guarantees in 4 weeks — Ars Technica AI · 2026-11-11