UChicago study: Pangram AI-text detector beats GPTZero, Originality and RoBERTa with <5% error rates
_akpiper · x · 2026-11-18
A University of Chicago study benchmarked four AI-writing detectors—Pangram, GPTZero, Originality, and RoBERTa—across multiple genres.
- Pangram wins decisively: near-perfect accuracy on long and medium texts, with false positive/negative rates under 0.05 even on short passages and stubs.
- RoBERTa fails badly: false positive rates hit 77.75% on news, 69.14% on resumes, and 64.88% on blogs.
- GPTZero and Originality land in between, mostly 0-3% false positives per genre but less consistent than Pangram.
For context, the author notes this accuracy vastly exceeds diagnostic tools in mental health, where the best depression and anxiety measures carry 15% false positive/negative rates.
More from Models
- GPT-6.2 Astra hands-on: upgrades million-line legacy app, all tests green in 6 hours — thesaraharminta · 2026-11-18
- GPT-6.2 vs GPT-6 vs day-1 GPT-6: Astra image generation compared side by side — lyraxana · 2026-11-18
- Will Oremus article on negative parallelism fuels AI-writing detection debate — TuhinChakr · 2026-11-18
- Pangram detector keeps confirming obvious AI-writing cases, Oremus pushes back on skeptics — TuhinChakr · 2026-11-18
- Astra 6.2 is much better at computer use than 6.0, says early tester — mallow610 · 2026-11-18
- Epoch: GPT-6.2 Astra scores 174 on ECI, above the pre-Astra reasoning-era trendline — airesearch12 · 2026-11-18