Anton Korinek, an economics professor at the University of Virginia, tells the students he advises nowadays that they should really begin to master a flourishing technology expected to transform the field of economics. That technology is generative artificial intelligence, or “genAI” for short.
Economists already utilize machine learning, a branch of AI, to analyze data and develop economic projections. But genAI is a separate technology. It underpins ChatGPT and similar tools and has advanced at a break-neck speed in recent months.
Korinek expects it to “revolutionize research,” according to a paper he wrote that was accepted for publication by the Journal of Economic Literature.
“It’s a powerful technology and if you use it, you can solve economic problems that we face as a society, better and more productively. That’s what research is all about,” Korinek told CNN in an interview.
GenAI isn’t just able to help with research. A recent paper by two economists at George Mason University showed that genAI can be useful in teaching economics by solving specific models in the classroom and creating exams. And a separate paper by researchers at the Federal Reserve Bank of St. Louis showed that an AI model they developed “understands the request for inflation forecasts” and “provides an inexpensive and accurate alternative to traditional forecasts.”
From assisting with research projects to forecasting inflation, genAI is more likely to empower economists, rather than take their jobs — at least for now.
Economists carry out several small tasks when doing research and Korinek’s paper argues that large language models, a specific type of genAI, can help in six ways: “ideation and feedback, writing, background research, data analysis, coding, and mathematical derivations.”
The most commonly used genAI tools are OpenAI’s ChatGPT, Microsoft’s New Bing, Google’s Bard, Anthropic’s Claude 2, and Meta’s LlaMA 2, according to Korinek. Any of those chatbots can help economists brainstorm what to research by simply asking the chatbot to provide a list of ideas. It can even evaluate research plans by providing pros and cons.
GenAI is great at copy editing and fine tuning writing, including spotting typos, suggesting titles, and even generating text specifically for social media to promote a paper. The new technology can help a researcher’s writing be more clear, specific, and flow much better, according to Korinek’s paper.
AI chatbots are also excellent at summarizing text. Both versions of ChatGPT (3.5 and 4) can summarize passages of text up to about 3,000 words. Claude 2, however, can summarize up to about 75,000 words, which covers the length of almost all academic research papers, according to Korinek. Economists can ask the chatbot questions on a specific paper such as “What are the author’s main conclusions?” or “What is the specific evidence supporting these points?”
Economic research typically involves technical tasks such as coding and devising mathematical proofs. GenAI tools, such as ChatGPT Advanced Data Analysis, are useful at writing, explaining, translating and even debugging code, especially in languages such as python and R. Chatbots can set up economic models, derive equations, and explain them, though Korinek noted that genAI’s capabilities related to math are limited at this point.
It’s also important to note that the latest version of each chatbot, such as ChatGPT-4, has greater capabilities than prior versions.
And of course, genAI isn’t foolproof. It sometimes spews out inaccurate information, sometimes referred to as “hallucinations.”
A teaching assistant
Economics professors Tyler Cowen and Alex Tabarrok at George Mason University published the paper “How to Learn and Teach Economics with Large Language Models, Including GPT” earlier this year, which explained how genAI can summarize text, improve writing, suggest ideas, and solve simple economic models with explanations, similar to Korinek’s paper.
But the paper also showed that genAI is particularly useful in the classroom.
“ChatGPT and Bing Chat will also create very credible syllabi for a variety of courses including readings, course policies, and grading procedures,” Cowen and Tabarrok wrote in their paper. “In case you are wondering, Chat GPT accepts late assignments with a penalty of 10% per day.”
The economists noted that genAI tools “are not yet ready to solve PhD problems but they are good at solving undergraduate models and for teaching students.”
A recent working paper suggests that genAI is adept at forecasting inflation — even more so than the economists who do it today.
The paper, authored by two policy advisers at the St. Louis Fed, compared inflation forecasts from Google’s PaLM, a large-language model chatbot similar to ChatGPT, to one of the leading sources of macroeconomic predictions: the Survey of Professional Forecasters. That group includes “tenured professionals working in the field of macroeconomic forecasting with advanced degrees in economics or related fields,” according to a spokesperson.
The researchers found that PaLM’s inflation predictions produced fewer errors than those of the SPF, which aggregates forecasts made by a variety of economists and financial analysts.
“These findings suggest that LLM models may provide an inexpensive and accurate alternative approach to generating forecasts of inflation,” the paper said.
Impact on employment
So what could be the impact of genAI’s advancement on employment in economics? It will likely be limited at first, mainly helping economists become more productive and efficient, but it could lead to some eventual job losses.
Jobs site Indeed conducted a recent study gauging the level of exposure of certain jobs to genAI based on the skills needed to perform them. It found that software development jobs, such as software engineers, face the highest exposure.
“I think genAI will create better jobs along the way because we will have gotten rid of the tasks that we don’t like to do,” Svenja Gudell, Indeed’s chief economist, told CNN. “However, getting to that steady state will require a period of change that could be turbulent and painful.”
Gudell said that a company could decide to cut its labor costs by laying people off if genAI helps maintain its usual level of output. Or a company could decide to keep all of its employees and simply enjoy greater output because of genAI. Either way, the technology poses a risk to employment.
Economists use a lot of technology to do their jobs, which are tasks that genAI could also perform, especially as it becomes more and more refined. However, being an economist will always require a “human touch,” according to Gudell. That includes teaching students and doing live presentations before an audience, but also when working with genAI itself.
“In economics, we often talk about the interpretability of a model. It’s one thing to just get an answer, but understanding how you got to that answer can be oftentimes just as important,” Gudell said. “Then the human comes into the loop again when it’s time to actually take that result and put it into context.”