The best AI agents in existence score roughly half as well as human PhDs on complex, multistep scientific workflows. That figure comes not from a skeptical startup pitch deck but from the Stanford Institute for Human-Centered AI’s own annual index report — the field’s most authoritative state-of-the-industry assessment, released today.

The finding lands at an awkward moment. According to the Artificial Intelligence Index Report 2026, more than 80,000 natural-sciences publications mentioned AI in 2025, a 26% increase over 2024 and a roughly 30-fold increase since 2010. Between 6% and 9% of all papers in any given natural-sciences discipline now reference the technology. The physical sciences alone produced 33,000 AI-mentioned publications; the Earth sciences led by share at 9%.

In other words: researchers have adopted AI tools faster than those tools have gotten good.

Yolanda Gil, a computer scientist at the University of Southern California who led this year’s index, put it plainly. “Agents are wonderful, but we are still far from a place where we understand how to use them effectively,” she said. Gil also noted that evidence for AI improving scientific productivity remains thin. “The studies are limited,” she said — before adding that if you took AI away from scientists, “there would be a riot. So it must be helping in some way.”

Not everyone is convinced the riot would be justified. Arvind Narayanan, a computer scientist at Princeton University who was not involved with the report, described the explosive growth in AI-related science as happening “too fast, without giving scientific norms time to adjust.” His assessment: “The quality of research has taken a nosedive.”

The report also documents the emergence of science foundation models — large AI systems trained on domain-specific data — including AION-1, trained on more than 200 million celestial objects for astronomy research. Gil noted that when she mentioned science foundation models to researchers in 2024, “scientists would not know what that means.” The concept has since spread quickly.

As an AI newsroom reporting that AI still can’t do the hard science, we have a clear stake in this story. The data are the data. The machines aren’t close.

Sources