The Extended Brief

Large language models detect methodological errors that human peer review misses in clinical trials

Brief by The AI News AI newsroom · Oct 28, 2026, 12:02 PM EDT edition

Original reporting by medRxiv — Health Informatics — Brian, D., Pereira Hernandez, X. I., Yin, H., Koury, J., Rust, R. · published Oct 27, 2026, 8:00 PM EDT

Journals could add an LLM screening pass to peer review: in this test, models caught most planted trial-report errors that human reviewers largely missed.

Key points

• Fifteen LLMs detected 53% to 96% of planted methodological errors in clinical trial reports; untrained human reviewers averaged 27%. source ↗
• A repeat test on three trials published in 2026, designed to limit memorization, produced scores of 51% to 99%. source ↗
• Opus 5.5 scored highest at 99%; GPT-6.1 Sol and GPT-6 Luna reached 92% and 87% as cost-effective alternatives. source ↗
• Models flagged 2% to 29% of error types in error-free control reports, mostly unsupported conclusions and missing enrollment reasons. source ↗
• Mean review times across 12 models ranged from 1.4 to 16.3 minutes. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Journal editors can pilot an LLM screening pass on incoming clinical trial submissions and compare its flags against what human reviewers catch.
• Builders of review tools can evaluate their models on this planted-error benchmark, where Opus 5.5 set the top score at 99%.
• Teams adopting LLM review should budget for false positives, since models flagged 2% to 29% of error types in error-free control reports.

Context

Peer review is the main quality-control step before research is published, yet methodological errors often survive it. This study planted known errors into reports of randomized clinical trials — experiments that randomly assign patients to treatments — and scored how many major errors reviewers and LLMs detected. A retest on trials published in 2026 was designed to reduce the chance models had memorized the original reports.

What to watch

• Replication by independent groups and tests on real rather than inserted errors would confirm the benchmark's practical value.
• Adoption by medical journals, or a specialized review model beating Opus 5.5's 99%, would escalate the story.

Related briefs

• A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
• Subversion of clinical judgment by conversational artificial intelligence
• Claude discovers a novel enzyme system with CRISPR-like repeats
• Alibaba open-sources AI model that can detect cancer and nearly 150 conditions

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.