Researchers find Claude Opus 5.5 detects 96% of planted clinical-paper errors
TL;DR:
- Researchers tested 15 language models on clinical manuscripts containing deliberately inserted methodological errors.
- Claude Opus 5.5 had the highest detection rate, finding 96% of the inserted errors.
- Repeated model reviews and differences in access to newer guidelines complicate comparisons with human reviewers.
Summary
A team including USC neuroscientist Ruslan Rust tested whether language models could find methodological mistakes inserted into clinical manuscripts. All 15 models detected substantial shares of the errors, with Opus 5.5 leading at 96%. The findings suggest a possible screening role for journals, whose reviews can take months. However, repeated model runs and access to newer standards make the human comparison uneven.
The Bigger Picture
The headline tells you what happened. Discover the facts behind it, the forces at work, the connections that matter, and what comes next.
Why This Matters
The useful result is broader than one model's winning score: even the weakest tested system found roughly half the planted mistakes. That makes manuscript screening worth testing as a workflow, rather than treating AI only as a writing tool. But detecting inserted flaws isn't the same as reliably reviewing real submissions. Journals would need evidence on false alarms, missed natural errors and reviewer time saved before treating these results as a case for changing editorial practice.
Where the Story Goes Next
- ScenarioWatch for prospective journal trials reporting detection rates, false positives and review time on unaltered submissions under comparable human and AI conditions.