Good morning,
The debate continues on whether generative A.I. large language models (LLMs) like ChatGPT, designed to understand and generate human-like text, will impact the future of work, including finance.
Researchers at OpenAI, the company that launched ChatGPT, and the University of Pennsylvania released a working paper last week titled, “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.”
It predicts that about 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of LLMs. And “approximately 19% of jobs have at least 50% of their tasks exposed when considering both current model capabilities and anticipated tools built upon them,” according to the report.
They measured the "exposure" of detailed work activities, which is like a job description, and tasks to LLMs' capabilities. No exposure means using an LLM via ChapGPT didn’t reduce the amount of time to complete an activity. Direct exposure means it reduces the time at least by half. And a third category indicated that access to an LLM alone wouldn’t reduce the time by half—but additional software developed on top of the LLM could achieve that goal.
The areas that have exposure to LLMs include jobs working in securities commodities contracts and other financial investments, monetary authorities, insurance carriers, data processing hosting and related services, and publishing, to name a few. And some occupations with no exposure include truck drivers, agricultural equipment operators, and glass installers. However, the researchers didn't assess if or whose jobs would be lost with the increased exposure to LLMs.
The findings are based on a Department of Labor database of 1,016 occupations, in addition to 2020 and 2021 Bureau of Labor Statistics employment and wage data. The researchers used both A.I. models and people to assign exposure levels.
Also noted in the report—predicted effects of LLMs span all wage levels, with higher-income jobs potentially facing greater exposure to LLM capabilities and LLM-powered software.
Daron Acemoglu is a professor and labor economist at MIT. Acemoglu is not an author of the ChatGPT report but is the coauthor of research on automation and wage inequality. I asked his opinion of the ChatGPT research.
"The predictions of the new working paper are reasonable, though, of course, nobody can see the future," Acremoglu says. He noted the optimism of the researchers, but 50% is a "huge" prediction. “Automation of certain tasks in the past has not always boosted productivity as much as its proponents had hoped. It remains to be seen whether LLMs will do better here.” He also noted a risk for demographic groups who previously specialized in automated tasks to have depressed wages.
When it comes to generative A.I. and the finance function, Michael Schrage, a research fellow at the MIT Sloan School Initiative on the Digital Economy, used the LLMs ChatGPT and DallE2 to demonstrate financial planning and analysis scenario generation, he says.
“Frankly, what we did was demonstrably superior to the ‘humans-only' scenarios,” Schrage says. “In my mind, there’s no question that—just as macros automated huge swathes of mission-critical Excel spreadsheets—we’ll soon see ‘prompt engineering’ and engineered prompts automate even greater portions of financial process workflows. To me, the wild card is to what extent ‘fine tuning’ models (for example, using your own models to complement or supplement the LLM) get developed, trained, and deployed.”
Sheryl Estrada
sheryl.estrada@fortune.com
Sign up here to receive CFO Daily weekday mornings in your inbox.
