Investments to bring the power of machine learning to drug discovery have soared in the past year, according to a benchmark report that tracks trends in the development of artificial intelligence.

The money committed to companies and projects in this area increased to $13.8 billion, more than 4.5 times that invested in 2019, according to the Artificial Intelligence Index, an annual report produced under the auspices of Stanford University’s Institute for Human-Centered Artificial Intelligence (HAI).

“The pandemic is part of what drove that,” notes Erik Brynjolfsson, an economics professor, senior fellow at HAI, and director of the Stanford Digital Economy Lab. “We have all benefited from machine-learning techniques that have helped identify new drug options and helped with vaccine development.”

The A.I. Index showed that while A.I. startups received a record amount of funding in 2020, with more than $40 billion invested globally, that money went to an increasingly small number of companies. Fewer than 1,000 A.I. startups received funding in 2020 compared with more than 4,000 in 2017, which was the high-water mark for the number of A.I. startups. Brynjolfsson said this was an indication that A.I. was beginning to mature as a technology and was moving from high-tech startups into more established businesses.

The A.I. Index also showed the continued demand for A.I. expertise in business globally. In 2019, the latest year for which figures were available, 65% of North American Ph.D.s in A.I. went to work in industry, up from 44.4% in 2010. An analysis of 2020 LinkedIn data from 14 countries shows that the hiring of those with A.I. skills is significantly higher than in 2016 across almost every country, with Brazil, India, Canada, and Singapore showing the largest increase over that period. Despite the pandemic, LinkedIn indicated continued hiring across all 14 nations in the sample.

Nor does the pandemic seem to have dented business enthusiasm for A.I.: The A.I. Index cited a McKinsey survey in which half of business leaders said the pandemic would have no effect on their A.I. spending, while 27% said it was actually prompting them to increase spending, as companies accelerated digital transformation efforts to deal with remote workforces, supply chain disruptions, a jump in e-commerce, and the need to run manufacturing operations with fewer staff physically on factory floors.

Despite this surge, Brynjolfsson emphasized that adoption of A.I. was still at an early stage in American industry. In a survey of 850,000 U.S. companies that he worked on, Brynjolfsson said that adoption of most advanced technologies was in the low single-digit percentages. He said that only 1.3% of the firms in that survey reported using any kind of robotics, for instance.

He said that the fact that adoption of A.I. and other forms of automation has not yet had an impact on U.S. economic data, such as productivity, is likely a function of two things: First, he said, conventional economic statistics are not very good at capturing some of the value from A.I. But he also said that he thought productivity gains from new technologies followed a J-curve shape and that with A.I., we were still at the bottom of that curve. “A technology breakthrough often needs a lot of complementary investments in other technology, in human skills, and in reorganization of business processes before you can start to see big productivity gains,” he said.

The A.I. Index showed that the technology is continuing to become increasingly powerful in many ways. This was particularly true of so-called generative systems, which can automatically create new images or write passages of text that are often indistinguishable from similar examples made by humans.

For certain tasks that involve both visual and language skills, A.I. systems have also made a big leap forward in capabilities. On a benchmark test in which the software is given an image and a question about that image that it must answer correctly, top A.I. software now answers with 76% accuracy, up from 40% in 2015. Humans score about 81% on the test. In another test in which the software is given an image and then asked a difficult question and required to justify its answer with reasoning, the best machines now score 70.5%, up from just 44% in 2018. Humans average about 85% on this task.

The report also highlighted the continued technological arms race between China and the U.S. in A.I.: China surpassed the U.S. in 2020 in terms of the number of A.I. research papers its scientists published in academic journals, but the U.S. scientists’ papers were more frequently accepted for prestigious conferences and were more highly cited by other researchers globally. U.S. universities remain a key factor in the country’s prowess in the technology, but they are heavily dependent on foreign students: In 2019, 64.3% of A.I. Ph.D.s in North America were foreign students, 4.3% more than the year before. But of those graduating, 82% remained and took jobs in the U.S.

Diversity remains a big challenge among those working on A.I. Almost half of all new A.I. Ph.D. students in the U.S. were white, while just 2.4% were Black, and 3.2% were Hispanic, the report found.

And A.I. ethics remains a fraught area, the report indicated. It said that while an increasing amount of attention was being paid to bias, fairness, and ethics in A.I., the field lacked a consensus around benchmarks that could be used to measure progress. It also noted that there was a far stronger interest in A.I. ethics among researchers and civil society groups than there was among those working in businesses using the technology.