AI leaders have predicted that it will bring about disruptive scientific progress: curing cancer, doubling human lifespans, colonizing space, and achieving a century’s worth of advancement in the next decade. With U.S. federal science funding being cut, the timing seems fitting, since AI may even replace the need for large-scale scientific labor.
At least within tech circles, there is a common-sense view that as AI is applied to every step of the scientific process—surveying existing literature, generating new ideas, analyzing data and running experiments to test those ideas, writing up results, and conducting “peer” review—it will dramatically accelerate scientific progress.
But when it comes to the impact of a new technology on existing institutions, many early predictions that seemed like common sense later turned out to be badly wrong. The Catholic Church once welcomed the printing press, believing it could strengthen its authority by printing the Bible. Early expectations for social media were similarly naive and optimistic: people thought that after the Arab Spring, democracy would spread widely across the globe.
Likewise, AI’s impact on science may be surprising. Even if individual scientists benefit from adopting AI, that does not necessarily mean science as a whole will benefit. When thinking at the macro level, we are dealing with a complex system with emergent behavior. This system operates in unexpected ways because it is not a market. In some respects, it is better than a market—for example, it rewards truth—but in others it is worse, such as in its ability to respond to technological shocks. So far, on balance, AI has been abadshock for science, pushing many of science’s processes to the brink of collapse.
Any serious attempt to predict AI’s impact on science must confront the production-progress paradox. The number of published scientific papers has been growing exponentially: from 1900 to 2015, it increased 500-fold. But by any available measure, real progress has remained flat—or even slowed. So we have to ask: how is AI affecting, and how will it affect, the factors that drive this disconnect?
Our analysis in this article suggests that AI will likely widen that gap. That does not apply equally to every field of science, and it is far from inevitable. If action is taken in time and with enough seriousness, as we propose below, it may still be possible to reverse course. Unfortunately, AI companies, science funders, and policymakers do not seem to understand where the real bottlenecks to scientific progress lie. They are trying to accelerate “production” only, which is like adding lanes to a highway when the real blockage is a toll booth—it only makes the problem worse.
Table of Contents
- Science Has Been Slowing Down—the Production-Progress Paradox
- Why Is Progress Slowing? Can AI Help?
- Science Is Not Ready for Software, Let Alone AI
- AI Might Prolong the Reliance on Flawed Theories
- Human Understanding Remains Essential
- Implications for the Future of Science
- Final Thoughts
Science Has Been Slowing Down—the Production-Progress Paradox
The total number of published papers is growing exponentially, doubling every 12 years. The number of researchers who have written papers is growing even faster. And between 2000 and 2021, investment in R&D by the world’s seven largest funders—the U.S., China, Japan, Germany, South Korea, the U.K., and France—quadrupled.1
But does that mean progress is accelerating? Not necessarily. Some papers lead to fundamental breakthroughs that change the direction of science, while others only make small improvements to already known results.
Real progress comes from breakthroughs in understanding. For example, we did not truly understand plate tectonics—that is, the idea that continents move—until the middle of the last century. Before that, geologists were not even able to ask the right questions. They were trying to explain the effects of the Earth cooling, believing that features like mountains were formed that way. Without plate tectonics, no number of discoveries or papers within the old geological paradigm could have produced the same kind of progress.
So it is entirely possible for the number of papers to grow exponentially while progress does not increase at the same rate, or even slows down. How can we tell whether that is happening?
One major challenge in answering this question is that, unlike research output, progress has no clear, objective metric. Fortunately, an entire field—“the science of science,” or metascience—is trying to answer this question. Metascience uses the scientific method to study science itself. It examines questions such as: What proportion of findings can be reproduced? Which factors affect the quality of researchers’ work? How do academic incentives shape scientific outcomes? How do different models of research funding affect progress? And how fast is progress actually happening?