Science PhD's Double Outperformance of AI But Embrace It For Research

Stanford University released its report on the state of AI yesterday, widely considered the definitive take on trends, performance, education and all related indices. 

There is a wealth of data in the report, much of which is available online, but two notable issues stand out. The first is that China and the US are clearly in the lead in most categories, now almost equal, with the US maintaining a slight lead. The other is that the embrace of AI in all fields continues to grow, but for the time being, PhDs in various sciences continue to outperform AI by a significant degree - approximately double - in multistep workflows. This will probably not be permanent, but it does add to the considerable body of evidence suggesting that AI's global dominance is likely to take longer than Silicon Valley would have the rest of the world believe. JL

Nicola Jones reports in Nature:

Many researchers have started to rely on AI ‘agents’ that autonomously carry out actions including scientific workflows, but the report is skeptical about their performance. AI agents still struggle to reliably perform multistep workflows, it reports, with the best AI agents scoring roughly half as well as human specialists with PhDs. “Agents are wonderful, but we are still far from a place where we use them effectively.” (Despite that) in 2025, more than 80,000 papers, preprints and other types of publication in the natural sciences — which includes life, physical and Earth sciences — mentioned AI, 26% more than in 2024. The subcategory of physical sciences had the largest number of publications that mention AI (33,000). The Earth sciences category had the highest percentage (9%).

In an indication of how quickly scientists are embracing artificial intelligence, the number of publications in the natural sciences that mention AI grew by almost 30-fold from 2010 to 2025, according to an influential annual state-of-the-field report.

The proportion of publications in any given natural-sciences field that mention AI ranges from 6% to 9% (see ‘AI paper boom’), according to the Artificial Intelligence Index Report 2026, released today by the Stanford Institute for Human-Centered AI at Stanford University in California1.

“Scientists have really embraced this AI era,” says computer scientist Yolanda Gil at the University of Southern California in Los Angeles, who led this year’s index report (see ‘Fun AI facts’).

Source: Artificial Intelligence Index Report 2026

Alongside the boom in AI-related science publications, the report also lists a host of newly released science foundation models — AI models that are broadly trained to take on a wide range of tasks, but also specially trained on massive data sets from a specific domain of science.

Many researchers have started to rely on AI ‘agents’ that autonomously carry out actions including scientific workflows, but the report is sceptical about their performance. AI agents still struggle to reliably perform multistep workflows, it reports, with the best AI agents scoring roughly half as well as human specialists with PhDs. “Agents are wonderful, but we are still far from a place where we understand how to use them effectively,” says Gil.

Growing numbers

The report says that in 2025, more than 80,000 papers, preprints and other types of publication in the natural sciences — which includes life, physical and Earth sciences — mentioned AI, 26% more than in 2024. The subcategory of physical sciences had the largest number of publications that mention AI (33,000). The Earth sciences category had the highest percentage (9%).

How AI agents will change research: a scientist’s guide

The boom in AI-related science papers is “not surprising”, says computer scientist Arvind Narayanan at Princeton University in New Jersey, who was not involved with creating the index. But it’s not clear, he says, whether the rise of AI use is productive for science. “Whether or not this explosive growth is meaningful is hotly debated,” Narayanan says. “My view is that it is happening too fast, without giving scientific norms time to adjust, and so the quality of research has taken a nosedive.”

Gil says there isn’t much evidence yet that AI is improving scientists’ productivity. “The studies are limited,” she says. But, she adds, scientists “can’t live without it. If you took AI away from them, there would be a riot. So it must be helping in some way.”

Top models

As uptake has grown, so has the number of AI platforms available to researchers. This past year saw the emergence of many science foundation models, including the first one for astronomy, AION-1, which was trained on more than 200 million celestial objects. This training helps it to classify galaxies or estimate their properties. “When I talked to scientists in 2024 and said ‘There’s foundation models for science’, scientists would not know what that means. They didn’t know they existed. I think we have seen that really advance very quickly,” says Gil.