RESEARCH·Nov 18, 2026, 10:26 PM ET·Emerging signal

Claude Opus 5.5 detects 96% of planted clinical-paper flaws in study

TL;DR:

  • Researchers tested 15 language models on clinical manuscripts containing deliberately introduced methodological flaws.
  • Anthropic’s Claude Opus 5.5 detected 96% of the inserted errors, while the lowest-scoring model detected 53%.
  • Researchers had the models review each paper repeatedly.
  • The models also identified errors in versions of the manuscripts that hadn't been altered.

Summary

A preprint posted on medRxiv on Oct. 28 tested whether language models could catch deliberately introduced errors in clinical manuscripts. Anthropic’s Claude Opus 5.5 led the 15-model group with 96% detection. The models outscored human reviewers, but repeated reviews and greater familiarity with current trial guidelines complicate that comparison. The results support investigating AI-assisted manuscript screening rather than treating this as proof that models can replace reviewers.

Open science.org in a new tab ↗