Two-week experiment: Pangram AI detector flagged zero human-written texts as AI
birchlse · x · 2026-11-17
- Editor-researcher Jesus Fernando reports running Pangram across many scenarios for two weeks for an upcoming paper: not a single reasonably long human-written text—including texts by ESL writers and his own—scored above 10% AI generation probability.
- He calls the widespread refusal to accept this "the oddest intellectual position" of the decade: fooling Pangram into false negatives with humanizing tools and careful editing is easy, but producing false positives on human text is nearly impossible.
- Screenshot shows Pangram 3.3.2 rating a mathematical-economics passage as 100% AI-generated at Medium confidence. Full paper to be circulated the next day.
More from Models
- Developer burns $200 and 12 hours fixing Gemini hallucinations, quits until next model — jasondeanlee · 2026-11-17
- Inkling 2 Small Scores 25.1% on FrontierSWE, Matching Kimi K3 at a Fraction of the Cost — LiTianleli · 2026-11-17
- User speculation: an OpenAI Chat-Work merge could route Plus users to Work and strain compute — Angaisb_ · 2026-11-17
- Dev: Grok Bot is doing to everyone what Claude Code did to coding agents — handotdev · 2026-11-17
- NVIDIA publishes full Cosmos3 collection on Hugging Face — NVIDIAAI · 2026-11-17
- NVIDIA Cosmos world models hit 20 million downloads on Hugging Face — NVIDIAAI · 2026-11-17