What the bond market has to say about generative AI

What you’ll learn: 

• Long-term U.S. Treasury yields decline more than 10 basis points through 15 trading days following AI model releases — an “economically large” figure.
• Falling bond prices could be an indication that investors don’t believe AI will have a widespread positive effect on domestic living standards, particularly consumption.
• Market watchers remain positively surprised by some aspects of AI’s progress.

Ever since ChatGPT was released nearly three years ago, generative AI has attracted intense interest from policymakers and businesses alike, raising important questions about AI’s effect on the economy and labor market.

One of those questions: What do market participants think about AI’s potential to transform domestic living standards, particularly regarding consumption growth or existential risk?

“There’s a huge amount of speculation about AI at the moment, and a huge amount of investment going into AI data centers and the like,” said MIT professor of economics Isaiah Andrews. “But are people acting like they believe this most transformative range of outcomes from AI is something that could plausibly happen?” 

To find out, Andrews collaborated with MIT Sloan associate professor of finance Maryam Farboodi on new research that used financial data to examine the effect of AI model releases on long-term asset prices, such as U.S. Treasury bonds.

“Financial market data was a very natural place to look because financial markets are supposed to be forward-looking,” Andrews said. 

“If market participants think that AI may have large growth effects, then new information about the trajectory of AI should impact long-term asset prices,” the researchers write. That includes assets that are not directly related to AI, such as U.S. Treasury bonds.

But, in examining the market reaction when new AI models were released, Andrews and Farboodi found that bond prices fell, which could be an indication that investors didn’t believe that AI would have a positive effect on future consumption growth.

Bonds fall in the wake of major model releases

In “Do Markets Believe in Transformative AI?” Andrews and Farboodi studied how U.S. Treasury bond yields were affected by 15 major model release dates from five major AI labs (OpenAI, Anthropic, Google DeepMind, xAI, and DeepSeek) from January 2023 to December 2024 — after the release of ChatGPT 3.5 but before the advent of the current presidential administration.

They looked in aggregate across different releases to minimize the effect of idiosyncratic events happening around the same time. (For example, one of the releases happened close to the failure of Silicon Valley Bank.)

They found that long-term Treasury, TIPS, and corporate bond yields fell in the wake of those AI releases, on average, and remained lower for weeks. The decline in Treasury yields following AI model releases exceeded 10 basis points, or 0.1 percentage points, an “economically large” figure that persisted through 15 trading days following the model release, they write. (The researchers found similar results using an extended sample that ran through the end of May 2025.) 

“Given all the speculation on what AI can do, I thought that if we found something, maybe we’d find growth expectations rising,” Andrews said. “But to find yields decline rather than increase, I found [that] very, very surprising.”

Farboodi was also taken aback to see yield declines consistent across long-term TIPS, Treasury bonds, and corporate bonds.

“We don’t have direct evidence on investor beliefs, but lack of large changes in short-term yields suggests to me is that they don’t think these changes [in consumption] are going to happen today,” Farboodi said. “They think these changes are going to emerge in a few years.”

AI and labor market disruption 

Farboodi said that it’s unclear why investors lowered their consumption growth expectations around the model releases, “but it does suggest that the market identifies some risk.”

Although the authors did not find any evidence that investors are experiencing greater uncertainty about future consumption growth, one possible explanation for the decline in long-term yields could have something to do with AI’s effect on the labor market. 

“People expect AI to have labor market disruptions,” Farboodi said. While AI can create new jobs and productivity gains, it can displace workers, require new skills, and widen inequality if not managed properly. 

CEOs have publicly discussed how AI is reducing hiring needs. And new research shows a rising unemployment rate among college graduates, in addition to tepid job growth across several white-collar sectors, hinting at AI’s growing effect on role displacement in certain industries. 

If AI displaces workers, that implies that “aggregate consumption will not grow, and inequality will go up. People are concerned with that,” Farboodi said. “So that is one story that could be consistent with our results, though we don’t have any specific evidence for that.” 

She added, “An alternative story consistent with our results is that AI disproportionately benefits entrepreneurs who are highly invested in it, at the cost of average population. But we do not have any specific evidence for this story, either.”

Leading the AI-Driven Organization

In person at MIT Sloan

Register Now

Investors are still keen on AI

To gauge overall sentiment on AI, the authors separately analyzed AI-progress forecasts from Metaculus, an online prediction platform. They focused on a question that asked users of the platform to predict the date on which the first weak artificial general intelligence system that satisfies specific criteria will be publicly announced. 

The authors examined how Metaculus forecast distribution changed around AI event release dates. They found that Metaculus participants positively updated their beliefs about AI progress around the AI model releases in their sample. 

“On the dates around the AI model releases, we see larger [positive] movements in Metaculus participant beliefs,” Andrews said. 

But that’s also when the researchers saw larger drops in bond yields.

“The model releases give people hope that artificial general intelligence is coming closer and will arrive sooner than we thought, but it does not appear that positive news about AI progress corresponds to positive views about consumption,” Andrews said.

Looking at stocks next 

Although the current research did not involve stocks, the authors did look at the S&P 500 and noticed modest increases in the index following the releases of the AI models in their sample. In the future, they plan to follow up on this research by looking at data on equities, but they acknowledged that doing so will be more complicated.

Maryam Farboodi is an associate professor of finance at MIT Sloan. She is an applied theorist whose research focuses on the economics of big data with applications to macroeconomics, finance, and financial intermediation. Farboodi received the 2024 Elaine Bennett Research Prize for her work in understanding how economic growth interacts with the evolving role of big data in the financial sector.

Isaiah Andrews is a professor of economics at MIT. He is an econometrician whose research focuses on developing methods to derive reliable conclusions from economic data, including in settings where the data may not be very informative or the researcher’s model may be wrong. He received a MacArthur fellowship in 2020 and the John Bates Clark Medal in 2021.