Pangram claims third-party validation as only reliable AI text detector, disputed

jackinwarsaw · x · 2026-09-03

Pangram rounds up third-party evaluations: a June 2026 VUB study testing four detectors on 160 academic papers found only Pangram reliably detected fully AI-written (97.5%) and humanized (95%) text, while GPTZero, Turnitin and Copyleaks scored 0% on fully AI content. Critics counter that the NBER paper isn't listed on Pangram's roundup page, and a UMD paper found 2% false positive rate (possibly an older May 2025 model); defenders note the NBER/UChicago paper measured 0.1% FPR on 2k human samples and sliding-window evaluation differences shouldn't be blamed on Pangram.

Original post →

More from Models

Models channel →