This story first appeared in TIME's new artificial intelligence newsletter, In the Loop. You can subscribe herefree to receive regular news like this straight to your inbox.
How much will AI contribute to the development of the US economy?
New research by Anthropic, reviewed exclusively by TIME ahead of its release today, offers at least a partial answer to that question.
By looking at aggregate data about how people use Claude as part of their work, Anthropic researchers have come to estimate how much AI could contribute to annual productivity growth—an important contribution to the overall growth rate of the overall economy—as the technology becomes more widely used.
Their answer: Current generation AI models could increase the annual rate of labor productivity growth in the United States by 1.8%, double the average growth rate since 2019. Assuming that labor makes up 60% of total productivity in the economy and that AI reaches full adoption within a decade, “this implies overall total factor productivity growth of 1.1% per year,” the researchers write. That number, the study's authors told TIME, is a close approximation of how much AI could contribute to overall economic growth. “In these models, labor productivity will generally equal GDP growth,” assuming labor supply remains constant, Peter McCrory, head of economics at Anthropic and co-author of the study, told TIME.
How is the research going? But take these numbers with a grain of salt, because the method by which they were arrived at is unorthodox. First, anthropology researchers created a tool (called Clio), which allows them to extract detailed information about Claude's actual usage, which they say helps maintain privacy. Armed with a sample of 100,000 conversations, the researchers analyzed them to check what tasks Claude performed in each one. To calculate how much time Claude saved in each conversation, they asked a separate version of Claude to estimate how long each task would take with and without AI assistance. They then consulted existing economic data to calculate the average wage per time saved by occupation. Finally, they extrapolated these time savings, weighted by the importance of each task to the entire economy, thus arriving at a number showing the efficiency gains achieved by AI across different task profiles.
What are the restrictions? — For starters, the study's main limitation is that it assumes workers spend whatever time they save by using AI on more productive work rather than, say, spending more time with their children or doing laundry. It also does not account for the time people spend on tasks other than talking to Claude, including time checking the factual accuracy of his answers. Another limitation is that the researchers rely on Claude to estimate how long tasks will take to complete, although they tested Claude's estimates on real data and found them acceptable. Finally, the study does not take into account the rapid improvement in the capabilities of AI tools, but instead assumes that AI will remain at its current level of capabilities over the next decade. Other things being equal, this suggests that the study may be underestimation AI's contribution to productivity growth in the next decade.
Should I be afraid? — The document makes no mention of unemployment, which is especially notable since Anthropic CEO Dario Amodei said in May that AI could eliminate half of all entry-level white-collar jobs in the next one to five years and skyrocket unemployment to as much as 20%. I asked McCrory whether these concerns were supported by new data. “In our work, we have not yet specifically examined this issue of attribution [the causes of] “Productivity growth is positive for the economy, and we are also clear about how technology can impact the labor market,” he told me. “That's what we're trying to do here: just add more facts to the conversation.”
