Co-Scientist Review: Hallucinations Plummet but Limits Remain

A double-blind review of 150 AI-generated papers found Co-Scientist cut severe data hallucinations from 90% to 4% and outright fabrication to 0%, though selective reporting and methodology-code mismatches persist, suggesting AI still has far to go in handling the physical realities of science.

2026-08-29 ~ 2026-08-29 · 3 related posts

• Co-Scientist evaluation: Severe hallucinations drop to 4%, fabrication to 0% — SRSchmidgall · 2026-08-29
• AI Scientist limitations: Selective reporting and code-method divergence remain — SRSchmidgall · 2026-08-29
• AI has a long journey to navigate physical scientific realities, but potential is huge — SRSchmidgall · 2026-08-29