1,900 AI-for-science papers in a week: why a small test error proves almost nothing

bravo_abad · x · 2026-11-20

In Discovery at Scale, Jorge Bravo Abad counts 1,900 AI-for-science papers in one week (one every 5 minutes) and extracts a cross-domain pattern: a model can pass its held-out test and still be wrong about what it was built to predict.

Test error only measures reproduction of the training distribution — not whether labels are right, whether what matters is even in the data, or whether the celebrated gain came from the idea in the title. Six papers across solar-cell materials, electrochemistry, drug-like molecules, and volcanic ash show this:

None of these findings needed a new algorithm — just the oldest experimental move: change one thing, hold the rest fixed. The piece distills five questions a test error cannot answer: are labels right, what can't the model see, are test cases new, does the gain survive a placebo, and better than what.

Original post →

More from Research

Research channel →