Epoch AI finds stylized AI text evades detectors far more often than plain prompts

dl_weekly · x · 2026-07-26

Epoch AI tested three AI-text detectors and found they miss stylized, human-like AI writing far more often than plain AI text. Scientific writing was especially hard to catch, with false negatives around 26%, versus near-zero misses on basic prompts.

Original post →

More from Research

Research channel →