Iteratively co-written LLM papers score 33%-92% on Pangram, exposing detector gray zones
mjdramstead · x · 2026-09-24
Debates about Pangram's reliability focus on false positives, but phl43 highlights false negatives and why they arise. His workflow: iteratively co-write reading notes with GPT and Codex — dictating what to write, flagging issues, requesting targeted revisions. Papers produced this way score 33%-92% AI-generated on Pangram, even though sentence-level production is 100% AI. The gap exposes how ambiguous "AI-generated" is across granularity levels and reinforces his earlier doubts about the detector.
Related event: Pangram AI detector criticized: catches crude AI text, misses polished work(4 posts)→
More from Models
- IFM unveils PAN 2, an open foundation world model that predicts physical outcomes — nikola_mr64990 · 2026-11-11
- Indie dev flips from $20 Pro back to $200 Gemini Ultra after two days of deep use — yihui_indie · 2026-11-11
- Leak: Google's upcoming model lineup includes Hydrogen, Krypton, Helix and a video model — koltregaskes · 2026-11-11
- Rumor: OpenAI's Astra 6.2 near, spotted in Codex logs and Google Cloud quotas — haider1 · 2026-11-11
- Same prompt head-to-head: Fable 5.2 spends 3x time, 2x cost for auteur flair — iamfakhrealam · 2026-11-11
- Krafton open-sources Raon-1 nano, a 1.5B on-device LLM distilled from MoE — DimitrisPapail · 2026-11-11