Artificial intelligence (AI) is transforming our economies and our way of life. Some observers argue that AI may lead to large productivity gains, including by fostering scientific discoveries (Aghion and Bunel 2024, Davidson et al. 2026). At the same time, there are concerns about the negative side effects of this new technology. In particular, some worry that AI may lead to technological unemployment and rising inequality by triggering a wave of automation (Acemoglu et al. 2026). Consistent with this view, early empirical evidence suggests that while AI is boosting firms’ productivity, it is also inducing firms to replace workers with capital (Yotzov et al. 2026), causing a drop in the share of income going to labour (Minniti et al. 2026).

In a recent paper (Fornaro and Wolf 2026), we argue that coupling advances in AI with an appropriate macroeconomic policy mix is crucial to ensure that AI benefits workers and leads to shared prosperity. Indeed, our research suggests that macroeconomic policies may determine whether the spread of AI will cause an economic boom or a slump.

The making of an AI slump

We develop a macroeconomic framework to study monetary and fiscal policies for AI. Our model has three key features. First, as in Acemoglu and Restrepo (2018), firms can choose whether to perform some production tasks with labour or automate them with capital. We model advances in AI as a rise in the share of tasks that can be automated. This shock increases productivity and potential output, but also opens up the possibility that capital may displace labour in production. Second, as in Mian et al. (2021), the income distribution matters for aggregate demand. In particular, our economy is inhabited by workers and capitalists. As it is the case empirically, workers have a higher propensity to consume compared to capitalists. Third, the presence of nominal wage rigidities implies that output can deviate from its potential level and monetary policy has real effects.

We start by studying a scenario in which macroeconomic policies do not react to the rise in AI, and show that the result may be a demand-driven slump. The slump materialises because higher automation redistributes income from high-spending workers to low-spending capitalists (see also Imas 2026). Absent expansionary macroeconomic policies, the result is a drop in aggregate demand for consumption. Moreover, the prospect of weak demand and low sales reduces firms’ investment. Hence, while the rise in automation boosts productivity, its negative impact on aggregate demand ends up depressing output (Figure 1, right column).

Figure 1 Impact of higher automation on output and consumption

Notes: The left column refers to an economy in which output is equal to its potential value. The right column refers to an economy in which macroeconomic policies do not react to the rise in automation.

What are the other symptoms of this AI-driven slump? First, since productivity rises while output declines, the slump is characterised by high unemployment. Second, the slump affects workers and capitalists very differently. Workers’ income takes a double hit, as it falls both because output drops, and the labour share declines. Capitalists’ income, instead, is supported by the rise in the capital share. Since capitalists are the richest part of the population, the slump is associated with a rise in inequality.

The root cause of an automation-driven slump is that firms’ choice to automate production triggers aggregate demand externalities. That is, firms automate their production to cut costs, so as to increase sales and profits. But each individual firm does not consider that replacing workers with capital depresses aggregate demand, lowering the sales and profits of all other firms. When this general equilibrium effect is sufficiently strong, a paradox of productivity materialises: while automating production to increase productivity is rational from the perspective of an individual firm, the collective adoption of automation technologies decreases firms’ sales and profits.

Two challenges for monetary policy

We next consider a central bank that reacts to the rise in AI by setting monetary policy to maintain the economy at full employment. In this case the adoption of AI triggers an output boom. In the short run, the boom is sustained by an increase in firms’ investment, which is needed to fully exploit the productivity gains offered by the new automation technologies (Figure 1, left panels). Over time, consumption demand also rises. In fact, in the medium run, workers’ consumption increases because higher production more than compensates the drop in the labour share.

There are, however, two challenges that monetary policy faces when trying to sustain full employment during a rise in automation. First, in the short run, higher automation is inflationary. More precisely, the initial phases of the rise in AI look like a cost-push shock, that is, an adverse shift of the Phillips curve increasing the inflation rate consistent with any level of employment. This finding may come as a surprise, because higher automation boosts productivity, and rising productivity is often seen as a deflationary force.

What this logic misses is that higher automation reduces firms’ demand for labour. To support full employment, therefore, real wages have to fall. But since nominal wages are rigid, the drop in real wages has to happen through a rise in prices. For this reason, during the first phases of a rise in automation, the central bank faces a trade-off between containing inflation and supporting the labour market, and the economy may face an unusual combination of rapid productivity growth, weak labour market, and high inflation.

The second challenge for monetary policy comes from the path of the natural interest rate, that is the interest rate consistent with full employment. In the short run, high investment boosts the natural interest rate. In the medium run, however, low consumption demand by workers depresses the natural rate. If this effect is sufficiently strong, the natural rate goes into negative territory and monetary policy becomes constrained by the zero lower bound. Once this happens, the economy enters an automation-driven liquidity trap, in which weak demand depresses output, employment and investment (Figure 2).

Figure 2 An automation-driven liquidity trap

Notes: This figure compares an economy constrained by the lower bound on the policy rate (solid lines) to a counterfactual scenario in which the lower bound is absent (dashed lines).

Employment subsidies as macroeconomic policy

Finally, we ask how fiscal policy can help monetary policy. While the array of potential fiscal interventions is large, we focus our attention on employment subsidies, or equivalently on cuts in labour taxes. In our context, employment subsidies are a particularly promising tool because they contain inflation by reducing firms’ labour cost, and support demand by boosting workers’ income. In fact, we show that, during periods of rapid automation, employment subsidies and cuts in labour taxes can sustain employment and output, while at the same time keeping inflation closer to its target.

Figure 3 shows the impact of subsidising employment in response to an increase in automation. The figure compares a baseline economy without fiscal interventions (solid lines), to one in which the government subsidises employment at a constant rate (dashed lines). The subsidy serves a dual purpose. In the first phases of the transition, the subsidy helps to contain inflation. Over the medium run, the subsidy mitigates the output losses due to weak aggregate demand. As a result, subsidising employment sustains employment and output, while at the same time keeping inflation closer to its target.

Figure 3 Impact of employment subsidies during rise in automation

Before concluding, let us clarify that we are not predicting that AI will necessarily lead to a slump, or even that this is the most likely outcome. Rather, we argue that macroeconomic policies may determine whether we will experience an AI slump or an AI boom. Moreover, our analysis shows that monetary policy, on its own, may have a hard time converting an AI slump into a boom. Accomplishing this task is likely to require a careful mix of monetary and fiscal policies. In particular, employment subsidies or cuts in labour taxes could be a useful tool to contain inflation and support demand over the coming years.

References

Acemoglu, D, D Autor, and S Johnson (2026), “Building pro-worker artificial intelligence,” NBER Working Paper.

Acemoglu, D, and P Restrepo (2018), “The race between man and machine: Implications of technology for growth, factor shares, and employment”, American Economic Review 108(6): 1488–542.

Aghion, P, and S Bunel (2024), “AI and growth: Where do we stand?”, policy note.

Davidson, T, B Halperin, T Houlden, and A Korinek (2026), “When does automating AI research produce explosive growth? ”.

Fornaro, L, and M Wolf (2026), “Macroeconomic policies for AI”, CEPR Discussion Paper 21412.

Imas, A (2026), “Can advanced AI lead to negative economic growth?”, Ghosts of Electricity, Substack, 8 January.

Mian, A R, L Straub, and A Sufi (2021), “Indebted demand”, Quarterly Journal of Economics.

Minniti, A, K Prettner, F Venturini, D Bloom (2026), “AI and the distribution of income between capital and labour”, VoxEU.org, 3 March.

Yotzov, I, J M Barrero, N Bloom, P Bunn, S J Davis, K M Foster, A Jalca, B H Meyer, P Mizen, M A Navarrete, P Smietanka, G Thwaites, and B Z Wang (2026), “Firms predict an AI productivity boom is coming”, VoxEU.org, 12 March.