Researcher finds frontier models startlingly bad at end-to-end science

AI researcher Andrew White tested frontier models by having them expand a detailed paper outline into full text, and found the results embarrassing—diffuse structure, poor prose, dull conclusions. He concludes AI-generated papers are a mirage the world doesn't need.

2026-09-06 ~ 2026-09-06 · 4 related posts

• AI researcher: frontier models are appallingly bad at end-to-end scientific research — andrewgwils · 2026-09-06
• AI researcher fed paper outlines to frontier models to test end-to-end research — andrewgwils · 2026-09-06
• Frontier models given paper outlines produced embarrassing results, researcher reports — andrewgwils · 2026-09-06
• 'The world doesn't need AI-generated papers': researcher calls AI papers a mirage — andrewgwils · 2026-09-06