The Science chapter of the 2026 AI Index records sustained growth in AI-related scientific publications and the spread of foundation models across biology, chemistry, physics and astronomy.

The picture changes on end-to-end work: models can exceed human averages on some questionnaires while scoring well below experts on paper replication, real bioinformatics and other tasks involving long decision chains.

For R&D teams, the practical implication is to evaluate the whole workflow rather than the fluency of an answer. Data provenance, code execution, intermediate checks and specialist review must remain part of the architecture.

Why it matters

For R&D, the findings support workflows in which literature search, calculation, verification and decision-making remain separate, observable and reviewable.

Evidence limit

The report aggregates heterogeneous benchmarks and does not directly measure performance in company-specific nutraceutical workflows; local evaluation remains necessary.

Editorial note. This content is intended for industry professionals. It does not constitute medical advice, therapeutic guidance or regulatory advice. Read our editorial method.

Read the original source · Stanford HAI ↗