Research
The economic impact of AI: Four scenarios
As we transition to an AI-driven economy, careful management is crucial.
Summary
The impact of AI is unclear
ChatGPT and other AI innovations will transform the way we access, utilize, and generate knowledge. However, the economic implications remain uncertain. Is AI really a driver of large-scale investment? Is it really a treasure trove of productivity? And if so, how will the additional productivity gains be distributed? Will these accrue to actual workers, or will they go to companies and capital? And how does this revolution factor into demand? It’s all still very unclear.
The 1990s revisited?
Let’s first segue to the internet revolution of the late 1990s, which also sparked widespread enthusiasm. It marked a time of accelerated technological progress in many areas. Particularly in the United States, the economy’s performance significantly exceeded expectations. Growth was robust and unemployment dropped to the lowest level in a generation. Wages increased without inflating employers’ costs, enabling domestic demand to expand at above-trend rates, seemingly without putting upward pressure on inflation. However, can all of this be attributed to the rise of the internet? Or was it actually the result of a sequence of positive supply shocks, including the integration of much of Asia into the global economy?
In financial markets, the idea of a commercialized internet fuelled the dotcom bubble. Investors were eager to invest, at any valuation, in companies that merely hinted at plans to make money online. This speculative frenzy unfolded against a backdrop of real interest rates ranging from 1.5% to 2%, which were higher than those in the early 1990s but significantly lower than the rates in the 1980s. The bubble eventually burst when the Federal Reserve hiked its policy rate multiple times, from 4.75% to 6.5%, in response to an economy that was heating up more than anticipated and to rising inflation.
Even with the bursting of the dotcom bubble, the IT revolution has profoundly transformed our societies, creating winners (e.g., professionals for whom computers are complementary) and losers (e.g., middle-skilled workers whose jobs were automated). However, quantifying its overall economic impact remains a challenge. For one, there is no counterfactual. Secondly, statistics show that productivity growth has actually slowed since computers entered the workplace. That said, the positive effect of digitalization on productivity and welfare is likely to be underestimated. It has led to a shift from producer output to non-market output, with households using their own time and resources to produce services for their own consumption.
The parallels between the AI and the IT revolution are clear. But there are also numerous uncertainties regarding the potential and the constraints of AI. Many tools and applications are still in their infancy, and many business models are still in a development stage. It’s likely we’re being limited by our own imagination. Conversely, it is a hype right now, and companies that just mention AI in earnings calls know this resonates well with investors. Secondly, companies that lay off workers and attribute it to AI-led productivity gains could simply use AI to cast a positive light on an unfavorable situation. It may take a decade or two before we can say something definitive about AI, if we’re lucky. As such, we won’t pretend to know how this revolution will unfold.
How to proceed?
We can explore how AI could impact real interest rates, given different scenarios for its effects on demand and productivity. In doing so, we consider a 2x2 matrix of high/low demand and of high/low productivity growth. For each quadrant, we explore what this would mean for growth, inflation, central bank policy responses, equilibrium rates and real rates. As such, this piece is a jumping off point for thinking about the potential economic implications.
We’re not focusing on the quantity of jobs. Western economies are grappling with a demographic crunch, and likely running out of workers before running out of jobs (see Figures 5 and 6). These countries are all near or at full employment, particularly for skilled workers. Therefore, it’s not the quantity that’s significant, but the quality. A world where everyone performs tasks that require minimal training differs greatly from one where everyone is engaged in skilled work. Hence, what’s crucial is not the mere existence of work, but the availability of enough skilled workers.
AI: Four scenarios
We analyze our scenarios using Tyler Cowen and Alex Tabarrok's "Dynamic AD-AS model". This is a simple business cycle model with GDP growth and inflation on the X and Y-axes.
The model has three components. The first is the Solow curve, derived from the Solow growth model. We can think of it as Y*, the potential growth rate determined by the real factors of production capital and labor. Y* is independent of inflation or central bank policy, but is a key input for R*, the equilibrium rate of interest. The second is the Aggregate Demand curve, which is total nominal spending. It is defined by linear combinations of inflation and real growth. The third is the Short-Run Aggregate Supply curve, which shows the relationship between real growth and inflation, for a given expected rate of inflation P*. Changes in this SRAS curve allow the economy to temporarily deviate from its potential growth rate Y*.
1. Goldilocks: Productivity and demand rise
In the first scenario, we assume there will be an extensive set of AI tools that empowers workers and enhances their productivity, which requires substantial upfront investment. These tools, developed by a diverse ecosystem, are rapidly and widely adopted within organizations. They provide decision-making guidance and guardrails, integrating coded information and rules with tacit knowledge. This enables a wider workforce to perform tasks and make decisions that were typically reserved for experts, thereby alleviating costly bottlenecks. It significantly alters the value of human expertise. Some areas of expertise may become obsolete, others may rise in importance. But in general, human expertise is extended to a larger pool of workers. This leads to a radical transformation of the labor market, improving equality and productivity.
Aggregate demand: Investment spending increases to fund the AI-transformation. Stock markets rally, with an increase in investor confidence lifting valuations. Wealth effects lead to higher consumer spending, shifting the AD-curve outwards. The middle-skill, middle-class core of the labor market, which has been eroded by automation and globalization, will be revitalized. This dynamic also boosts consumer confidence and spending. The economy shifts from point a to b.
Short-run aggregate supply: The increase in demand and dynamism creates a temporary boom, with revenues initially rising at a faster rate than wages. This encourages producers to expand and to offer higher nominal wages, which in turn does lead to higher prices. As time progresses, an increasing part of the demand shock will be reflected in the inflation rate and less will be reflected in the real growth rate. It leads to higher inflation expectations and an upward shift in the SRAS-curve. Unless monetary policy intervenes, the economy would shift from point b to c1.
Solow: Note that in the long run, the economy must always be on the Solow curve. We assume that the AI revolution improves productivity and reduces inequality. This means that businesses can produce and sell more goods and services. It sets the economy on a higher path of potential growth, leading to a rightward shift in the Solow curve.
Policy responses and interest rates: Monetary policy intervenes when the economy is overheating and inflation expectations rise to well above 2%. A SRAS curve with P* > 2% is inconsistent with the central bank’s target. Policy rates will be raised in order to lower the AD curve to a level at which inflation expectations fall and the SRAS curve intersects with the AD curve on the Solow curve. In layman’s terms: the central bank hikes rates in order to reduce inflation expectations and to slow nominal GDP growth, eventually bringing inflation back on target. The economy shifts from point b to c2. It takes a higher equilibrium interest rate to encourage the volume of saving required for the high investment levels that are needed to sustain a high-growth economy. When inflation expectations settle again at 2%, real rates will have increased.
2. Excess capacity: Productivity rises, demand lags
In the second scenario, there still is that extensive set of AI tools that raises productivity. However, we now assume that the industry has a very steep learning curve, with AI functions becoming widely available at low price points. Although the fixed costs of developing and improving these technologies have been high, they are easily replicated. The marginal costs of many AI tools and user queries fall to zero. That is not good news for AI stocks and it doesn’t feed into an AI investment boom. In a relatively short time span, implementing these tools in the workplace will require no macro-significant upfront investment.
Tasks that now require human input and creativity are replaced by ChatGPT-like tools. Intelligence gets commoditized, but firms, managers, and high-skilled workers succeed in capturing the productivity gains. Service jobs that are traditionally done by middle-skilled workers in developed economies risk being outsourced to AI-assisted workers in developing economies. The profit share of GDP in developed economies increases, to the benefit of asset owners. It is difficult for laid-off workers to find new skilled employment. Their skills have become obsolete. There is more than enough work in sectors that are less vulnerable to AI-disruption, such as in-person service jobs. On balance, workers will shift to lower-paid jobs.
Aggregate demand: Profits rise, but are largely distributed to asset owners instead of being reinvested. Some households are AI winners, but the bulk of the consumers aren’t particularly happy. Inequality increases, offsetting productivity gains. On balance, there is no improvement in aggregate demand. Nominal spending continues at its pre-revolution pace.
Short-run aggregate supply: Unchanged demand amidst a positive supply shock opens up spare capacity in developed economies. It reduces investment demand and leads to lower inflation expectations. This is a downward shift in the SRAS curve. The economy shifts from point a to b. As time progresses, an increasing part of the supply shock will be reflected in the inflation rate and less will be reflected in the real growth rate. Unless monetary policy intervenes, the economy shifts from point b to c1.
Solow: The AI revolution leads to improved productivity. More value can be derived from the combination of capital and labor that is being supplied. It sets the economy on a higher path of potential growth, leading to a rightward shift in the Solow curve.
Policy responses and real rates: Inflation expectations fall and a SRAS curve with P* < 2% is inconsistent with the central bank’s target. The central bank cuts rates in order to raise the AD curve and to lift inflation expectations, such that the SRAS curve intersects with the AD curve on the Solow curve. Simply put, the central bank cuts rates, looking to raise nominal GDP growth and bring inflation back to target. The economy shifts from point b to c2. The increase in productivity suggests the equilibrium rate rises, but this is offset by the increase in inequality. The upward impact on real and equilibrium rates is not as pronounced as in scenario 1.
3. False dawn: Demand surges, but output sags
Scenario three draws the closest parallels with the computer revolution of the late 20th century. There is major investment and rapid AI development, similar to the first scenario, which boosts demand and lifts animal spirits among investors. In parts of the economy, productivity rises. However, in terms of aggregate productivity and, crucially, overall supply capacity, only incremental gains are delivered.
For one thing, Generative AI, while fun to experiment with, is still unreliable and risky. This may be acceptable for creative tasks where there is a human in the loop, but not for critical tasks which require judgement or display “humanity”. The real-world applicability of AI turns out to be much more challenging than how it has been sold. Extensive roll-out of – and reliance on – AI tools will eventually remain limited to a select number of sectors and jobs.
It is also far from certain that productivity gains will be translated in additional supply capacity. Recall that Keynes once envisioned a utopian future in which technology would be so advanced that workers would only need to work around 15 hours per week to meet their basic needs, providing them with more leisure time. Obviously, we’re still a long way from that , but the adoption of new technologies (and the expansion of the labor force) has shortened workweeks, with productivity gains translating into improved well-being rather than economic welfare. Europe is a prime example of this, and AI may accelerate this globally.
Finally, some of the productivity gains will be realized in non-market production. Services that were originally provided by producers, will now be done by consumers assisted by AI tools. This is a productivity gain that is difficult to capture in statistics, but it does lead to savings. In short, we all tend to perceive the AI revolution as a major macroeconomic change with large aggregate effects, but it could also be a major societal change with significant compositional effects.
Aggregate demand: Investment spending increases to fund the AI revolution. Stock markets rally, with an increase in investor confidence lifting valuations. Animal spirits among investors rise. Financial conditions ease. Consumers get wind of an economic boom. This all leads to an outward shift in the AD-curve. The economy shifts from point a to b.
Short-run aggregate supply: Firms are struggling to find the resources to meet the increase in demand, while AI is of little immediate practical help in large parts of the economy. Pricing power increases, raising margins and profits. This fuels the need for more investment. Inflation expectations rise. It leads to an upward shift in the SRAS curve. As time progresses, an increasing part of the supply shock will be reflected in the inflation rate and less will be reflected in the real growth rate. Unless monetary policy intervenes, the economy shifts from point b to c1.
Solow: The AI revolution leads to improved productivity in parts of the economy, but the quantity of labor that is supplied to the economy drops. There are no changes to the Solow curve.
Policy responses and real rates: Frothy asset markets lead to an overheating economy. Inflation expectations rise to well above 2%. A SRAS curve with P* > 2% is inconsistent with the central bank’s target. Policy rates will be raised in order to lower the AD curve to a level at which inflation expectations drop and the SRAS curve intersects with the AD curve on the Solow curve. The central bank hikes rates in order to reduce inflation expectations and to slow down nominal GDP growth, eventually bringing inflation back to target. The economy shifts from point b to c2. The potential growth rate hasn’t improved, so real rates will remain unchanged, as in the 1990s.
4. Struggling stagnation
In this rather dystopian economic scenario, the advent of AI does not lead to a productivity boom, but rather exacerbates societal inequality. The technology is useful, but unevenly distributed and inadequately adopted across the economy. This gives a competitive edge to only a select number of companies and a small share of the workforce who fully understand how to utilize AI and are able to extract economic rents. As a result, many workers are displaced from their previous employment and forced into lower productivity and lower income jobs elsewhere in the economy. The growing divide between winners and losers leads to increased wealth and income inequality, fuelling social discontent. This sparks social unrest, creates a political backlash, and sees the introduction of high regulatory hurdles that limit further diffusion of AI technologies.
Aggregate demand: Investment spending remains concentrated among a number of AI winners. Stocks may rally, but there is a clear distinction between large caps and small caps and between high-tech and low-tech companies. This ‘asset-rich income-poor’-variety of wealth creation is not supportive of demand, as economic anxiety increases. This leads to an inward shift of the AD curve. The economy shifts from point a to b.
Short-run aggregate supply: Spending is weak and inflation expectations fall. The outlook for investment is subdued. It leads to a downward shift in the SRAS curve. As time progresses, an increasing part of the supply shock will be reflected in the inflation rate and less will be reflected in the real growth rate. Unless monetary policy intervenes, the economy shifts from point b to c1.
Solow: There are no changes to the Solow-curve.
Policy responses and real rates: This scenario resembles the asset-rich income-poor economy of most of the 2010s. Falling inflation expectations and an SRAS curve with P* < 2% would prompt a monetary policy response to stimulate the economy. We’ll see interest rate cuts in order to lift inflation expectations and to get nominal GDP growth going again. Unless monetary policy is constrained by the zero-lower bound, the economy shifts from point b to c2. The potential growth rate hasn’t improved, so real rates will remain unchanged.
Conclusion
In the 1990s, the IT revolution sparked an asset bubble. Initially, the Federal Reserve refrained from intervening, given seemingly benign inflation trends. However, as price pressures rose and inflation expectations risked losing their anchor, the central bank stepped in, leading to the bubble’s eventual burst.
Computerization left an indelible mark on economies and societies, but quantifying its magnitude remains a challenge. While ICT-driven productivity has evidently increased, the benefits have been distributed unevenly, resulting in a polarized impact. Professionals have benefited from this shift, as computers complement their skills, but a significant portion of middle-skilled, middle-class jobs in factories and offices have been displaced. Consequently, these workers have often had to find employment in sectors that are difficult to automate, such as hands-on non-expert jobs, which tend to be low paid.
As we transition to an AI-driven economy, careful management is crucial. While productivity gains are likely, the risk lies in the unequal distribution of these gains. Lessons learned from the late 20th and early 21st century, when political leaders embraced the gospel of markets and scaled back social security provisions, will be essential for shaping policy responses. If mishandled, AI could widen the gap between economic winners and losers: between those for whom AI is a complementary tool and those for whom it is a substitutive one. This would exacerbate social unrest, driving further political polarization. It will then be hard to see how AI would set us on a path of higher growth, despite its promises.
It is in the policy response to the AI revolution where our scenarios drift apart. Key factors that could shape this response are antitrust and regulation, to promote widespread technological diffusion, and redistribution, to protect workers who have invested their lives in careers that are now displaced by AI. The political and economic winds have changed in recent years. The cold logic of the market has lost popular appeal. This a sharp contrast with the 1990s, when the solution to the problems created by automation were largely found in laissez-faire nostrums.
We would therefore think Goldilocks and False dawn are more plausible scenarios than Excess capacity or Struggling stagnation, which were more relevant between the 1980s and the 2010s. That means that, if we had to hazard a guess, the two most likely scenarios are those in which inflationary pressures rise and policy rates need to be raised.