Copyright Project Syndicate
CAMBRIDGE – Speculation about AI’s implications for society has reached a fever pitch. While OpenAI co-founder Sam Altman speaks of a future of radical abundance, others, including another OpenAI co-founder, Ilya Sutskever, warn of existential risks. Meanwhile, economists continue to debate whether AI will usher in unprecedented prosperity or widespread unemployment. What are we to make of such wildly divergent predictions? While we cannot know which future will materialize, we can examine what those with “skin in the game” believe will happen. Financial markets aggregate the views of a multitude of sophisticated investors who are risking money on their convictions. If they genuinely thought AI might transform the economy, this should be reflected in asset prices. In some respects, the market has certainly delivered a clear signal. Nvidia’s market capitalization has soared past $4 trillion as Microsoft, Google, Amazon, and other tech giants invest hundreds of billions annually in AI infrastructure. AI firms’ sky-high valuations suggest that investors expect enormous profits. But who will ultimately benefit, and through what mechanism? Do markets believe that AI will broadly boost productivity, raising living standards across society, or will it generate large, concentrated profits without widespread growth? Economic theory suggests that bond markets may hold the answers to these questions. If investors genuinely expect much higher future incomes from AI-driven productivity growth, they should save less today, driving up interest rates on long-term bonds. Think of it this way: If you are confident that you will be wealthy in 30 years, you will be less inclined to save aggressively now, and you will demand higher compensation from long-term borrowers for giving up consumption today. Similarly, if AI raises the perceived probability of extreme scenarios – whether the end of humanity or a post-scarcity economy where all material needs are met – that should also push up long-term rates, because the value of future consumption relative to present consumption declines. Either way, the implication is that bonds should offer a higher return to convince investors to forgo current consumption. In recent research, we tested these predictions by examining US Treasury yields around major AI model releases in 2023 and 2024. Each time OpenAI, Google, Anthropic, xAI, or DeepSeek released a new flagship model, markets received concrete information about AI’s capabilities and the pace of progress. The results were striking: Rather than rising around the time of AI announcements, long-term bond yields (10-, 20-, and 30-year maturities) fell, typically by more than ten basis points. If investors were becoming more optimistic about AI-driven growth, this is not what theory would predict. The declines persisted for weeks after the announcements, and similar patterns appeared in corporate bonds and inflation-protected government-issued securities (implying that inflation is not a key reason for the decline in long-maturity Treasury yields). Moreover, these yield changes were quite large compared to changes around other news releases, such as those following Federal Reserve Open Market Committee meetings. These data are a puzzle. Falling long-term yields suggest that investors are revising expectations downward – either by lowering anticipated consumption growth rates, reducing the perceived probability of extreme events, or both. Yet that seems to contradict both the enthusiasm visible in tech valuations and the impressive capabilities recent models have demonstrated. Although we can interpret these yield changes through standard economic models, any quantitative translation requires assumptions about investor risk preferences. Under conventional assumptions, our results suggest that the average model release led markets to reduce expected annual consumption growth by somewhere between 0.04 and 0.2 percentage points (depending on how risk-averse we assume investors are), or to lower the perceived annual probability of extreme transformation by 0.2 percentage points. Importantly, we find little evidence that rising uncertainty about future consumption growth drives these patterns. Instead, the changes appear driven by investors’ growth expectations, rather than their uncertainty. Could disappointing AI progress explain these results? This seems unlikely. Using data from Metaculus, a platform where participants predict AI milestones, we found that forecasters generally updated toward faster progress around these releases, suggesting they were positively, rather than negatively, surprised by the degree of technical progress revealed by most updates. This leaves us searching for explanations. Perhaps investors worry that AI will produce disruption and displacement without generating substantial growth. Alternatively, they might think AI’s benefits will flow primarily to tech companies and their shareholders while costs (job losses, downward wage pressure, and increased inequality) spread widely across the population, leading to lower aggregate consumption even as certain sectors prosper enormously. Another possibility is that we are observing reactions to something else correlated with AI announcements. While we cannot rule that out entirely, any alternative explanation must produce large, persistent downward movements concentrated in long-maturity bond yields, coinciding specifically with these release dates – itself a pattern requiring explanation. We think our findings cut through the speculation in a useful way. Financial markets do appear to take AI’s economic potential seriously – given the economically large and statistically significant reactions beyond tech-sector enthusiasm. But the signal from bond markets suggests investors are unconvinced that AI will generate sustained or widespread growth, or more extreme outcomes, even as equity markets highlight the profit potential for specific firms. In other words, markets appear to think the path from impressive AI capabilities to widespread prosperity is less straightforward than optimistic narratives suggest. Understanding whether investors’ caution reflects concerns about disruption, distribution, or other factors is crucial for designing appropriate policy responses. Speculation about what AI will mean for the economy will continue and perhaps intensify. As market participants place their bets, the resulting picture is more complex, and more concerning, than current headlines suggest.
