Today for AI
HOT RADAR
llmHEAT 0.0°

AI CLUSTERED EVENT · 10/10/2026

Tokyo Metropolitan University Study: Pangram Misses 79.8% of Meta Muse-Glimmer Rewritten Abstracts

2 reports archived0 independent sourcesupdated 10/10/2026, 16:01:17
Synthesis & Latest Updates

A new study from Tokyo Metropolitan University reveals that the AI-text detector Pangram missed 79.8% of scientific abstracts rewritten by Meta's Muse-Glimmer model, while maintaining a near-zero false positive rate on human-written texts. The findings demonstrate that detection efficacy is heavily dependent on the specific LLM used for rewriting, with Pangram catching 93.5% of GPT-5 rewrites but failing significantly against newer models, highlighting critical generalization gaps in current AI detection tools.

LATEST/A new study from Tokyo Metropolitan University reveals that the AI-text detector Pangram missed 79.8% of scientific abstracts rewritten by Meta's Muse-Glimmer model, while maintaining a near-zero false positive rate on human-written texts. The findings demonstrate that detection efficacy is heavily dependent on the specific LLM used for rewriting, with Pangram catching 93.5% of GPT-5 rewrites but failing significantly against newer models, highlighting critical generalization gaps in current AI detection tools.

HEAT TRENDHourly heat curve

52 fully observed hours
Now
0.0
Peak
16.717:00
24h change
-100%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 2 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·68 pts
    • AI text detectors exhibit severe 'version overfitting,' failing almost completely against models outside their training distribution.
    • Mainstream tools like Pangram show massive variance in miss rates (from <10% to >79%) when facing different vendors' new models.
    • Relying on single static detectors for academic review is no longer viable; dynamic re-testing mechanisms or multi-modal/behavioral analysis are required.
  2. Rohan Paul (@rohanpaul_ai)T2·68 pts
    • Pangram missed 79.8% of abstracts rewritten by Meta's Muse-Glimmer, with only 1 false positive out of 5,000 human abstracts.
    • Detection accuracy is highly model-dependent: 93.5% catch rate for GPT-5 rewrites vs. massive failure on newer models.
    • General-purpose AI detectors struggle to keep pace with rapid LLM evolution, revealing significant cross-model generalization gaps.
Tokyo Metropolitan University Study: Pangram Misses 79.8% of Meta Muse-Glimmer Rewritten Abstracts | Today for AI