A possible explanation, at least a partial one, for the secular decline in US productivity growth over the past half century is that our innovation system itself has become less productive. More research effort, in terms of people and resources, is required to generate a given amount of technological progress.
A classic example concerning Moore’s law comes from the 2017 paper “Are Ideas Getting Harder to Find?”: “The number of researchers required today to achieve the famous doubling every two years of the density of computer chips is more than 18 times larger than the number required in the early 1970s.”
While throwing more resources at the problem is one option—and increasing federal science investment would be a smart idea—another path forward is making scientists themselves more productive. Good news on that front from the new paper, “AI in Science: Early Insights,” from Google and Google DeepMind, in collaboration with MIT FutureTech.
From the Google paper:
We provide early insights on this from three data sources: a sample of 15 million Gemini interactions, an inventory of over 2,600 specialized AI models across disciplines, and a survey of over 600 scientists. We map these data to a new taxonomy of scientific tasks to study how scientists are using AI.
For starters, the researchers find that scientists already use AI more than most other occupations, and nearly half of those surveyed say they use some form of AI every day. This goes well beyond using chatbots to summarize papers. Researchers also use specialized AI models for tasks such as predicting protein structures, designing materials, analyzing medical data, and running scientific simulations.
Moreover, AI usage seems to have boosted productivity. Roughly three-quarters of surveyed scientists say AI saves them time, averaging almost seven hours a week, or nearly a full workday. Importantly, they mostly spend that saved time doing more research, rather than simply working fewer hours.
AI is also a silo breaker: “About 68 percent of scientists report more access to insights from other disciplines,” the Google paper states, “consistent with hopes that AI may lower the burden of knowledge that pushes researchers into ever-narrower specialties.”
There are some downsides or reasons for concern. Yes, AI-enabled scientists can produce more hypotheses and analyses, but experiments, data collection, and checking AI outputs remain slow and labor-intensive. A key stat on that front from the paper: “89 percent of those who save time spend more than a tenth of it checking AI outputs, and 46 percent spend more than a quarter.” (More here on the AI verification issue.)
But more concerning to the paper’s authors is the finding that nearly half of scientists “say AI pushes them toward safer, more incremental projects where benchmarks are established and results are reliable, against 28 percent who say it lets them take on riskier questions.” Not great if your ultimate goal is pushing forward the science frontier and finding the next big idea. Still, early days for this powerful general-purpose technology.