Can AI Save Us from the Debt Endgame?

The productivity revolution may offer a way out. But first, we must survive the journey.

Much of the discussion on AI centres around whether it will destroy jobs. Yet the developed world desperately needs the productivity revolution AI promises because, without it, escaping from enormous government debts will be incredibly difficult.

Furthermore, even if AI delivers in the long run, there are three reasons for caution in the near term. The debt problem may already be too large. AI could make things worse before making them better. And politicians may capture or stymie the productivity dividend.

Global public debt is approaching 100% of GDP. In the US, the IMF expects government debt to exceed 140% of GDP by the early 2030s. Ageing populations, higher defence expenditure and mounting interest costs are adding to the pressure.

Most conventional solutions involve pain of some sort. Taxes can rise, spending can be cut, inflation can erode the real value of debt, or governments can suppress borrowing costs through financial repression. However, there may be a more intriguing prospect: economies could grow their way out of their debt problems.

That is where AI could play an important role. Estimates vary enormously, but the OECD puts the potential boost to annual labour-productivity growth in highly AI-exposed economies such as the US and UK at between 0.4 and 1.3 percentage points over the next decade. I suspect the eventual impact could be considerably larger.

Previous technological revolutions have repeatedly proved difficult to measure in advance. During the great US information-technology boom, labour-productivity growth almost doubled from around 1.6% a year in 1974-95 to just over 3% in 1995-2004, with ICT accounting directly for around 1.5 percentage points of annual productivity growth. Railways and electricity transformed entire economies in ways that narrow contemporary measures struggled to capture. AI could be an even bigger gamechanger.

Railways reduced the cost of moving things. Electricity transformed the production and use of energy. Computers reduced the cost of processing information, and the internet transformed its distribution. AI promises to reduce the cost of cognitive work itself.

More importantly, it could accelerate innovation. AI can already help write software, design chips and discover drugs. If it materially increases the pace of scientific and technological progress, conventional productivity forecasts could prove far too conservative.

At the extreme, futurologists talk about a technological "singularity", where AI becomes capable of accelerating technological progress and perhaps even improving itself. Whether anything resembling this happens is unknowable. But it illustrates why attempts to put a precise number on AI's eventual economic impact should be treated with considerable caution.

AI could also improve government productivity. If it allows healthcare, administration and other public services to be provided more efficiently, the fiscal dividend could come from both a larger private-sector tax base and a more productive public sector.

So, there is genuine reason for optimism. But equally, there are substantial obstacles, the first being the sheer magnitude of the problem. AI doesn't merely have to generate productivity growth. It must generate enough to offset enormous existing debts and continuing political demands for spending.

Nor does greater productivity automatically translate into healthier public finances. If much of the benefit accrues to capital owners and multinational companies, governments may capture less of the dividend than headline GDP growth suggests. Greater inequality could simultaneously increase demands for redistribution. And politicians can simply spend the windfall. In short, a productivity revolution creates an opportunity to escape the debt trap. It doesn't guarantee that governments will take it.

The second problem may be the more immediate one. The investment comes first, but the productivity comes later. Building the AI economy requires extraordinary quantities of capital investment in semiconductors, data centres, electricity generation, transmission grids and associated infrastructure. That investment competes for capital and real resources at precisely the moment governments also need enormous quantities of them.

The initial effect could therefore be inflationary, pushing real interest rates and government borrowing costs higher. Only later do we hopefully receive the productivity dividend. The old fiscal model could break before the new productivity model has delivered. That leaves the next five to ten years particularly vulnerable to a sovereign or financial crisis. Indeed, AI could initially help precipitate the debt crisis it ultimately allows us to escape.

This is why financial repression will probably still be part of the story. Governments might eventually grow their way out of their debts but may first have to suppress borrowing costs to keep the system functioning until that growth arrives. Financial repression could be the bridge between the old world and the new.

The third problem is political. Every major technological advance has destroyed jobs. Yet technological progress hasn't produced ever-increasing unemployment. Rather, it has produced extraordinary increases in prosperity. Part of the reason we repeatedly fear otherwise is a forecasting asymmetry: we can see the jobs threatened by technological progress, but we cannot see the jobs it will create because the industries that will employ them don't exist yet. Imagine forecasting the employment consequences of the internet in 1995. Identifying threatened occupations was relatively easy. Foreseeing app developers, cloud-computing engineers and cybersecurity specialists was almost impossible. We risk making the same mistake with AI.

That doesn't mean unemployment won't rise during the transition. Indeed, employment could follow something resembling a J-curve. Existing jobs can disappear relatively quickly, while creating new industries, retraining workers and reorganising businesses takes time. Employment could initially deteriorate before the productivity gains generate sufficient new opportunities.

That period may also represent the greatest political danger. People will see jobs disappearing while the industries that eventually replace them remain largely invisible. Enter the modern Luddites.

The original Luddites were skilled workers whose livelihoods were threatened by machinery during the Industrial Revolution. Their resistance to change was understandable. This time, those threatened are likely to include politically influential middle-class professionals: lawyers, accountants, programmers, journalists, academics, civil servants, financial analysts and perhaps doctors.

We should therefore expect pressure for restrictions on AI, compulsory human oversight, professional protections and perhaps taxes on automation. Some regulation will plainly be necessary. But there is an enormous difference between regulating genuine risks and protecting incumbents from competition.

The costs of disruption are concentrated and visible. Many of the benefits are dispersed, delayed and currently unimaginable. The person whose livelihood is threatened can lobby politicians today. The entrepreneur who might create an AI-enabled industry employing thousands of people a decade from now cannot. They may not even know what that industry will be.

That creates a powerful political bias towards resisting change, precisely when the economy is near the bottom of the J-curve. Governments cannot simultaneously demand a productivity revolution while protecting every existing job and business model from the disruption required to create it.

There is one caveat to the reassuring lessons of history. If AI eventually becomes better than humans not merely at today's jobs but also at performing the new jobs technological progress creates, the historical mechanism through which technology creates new employment could break down. But if we reach anything approaching that level of capability, today's estimates of productivity growth are also likely to prove well short of the mark.

The bottom line is that nobody knows whether the productivity dividend will be large enough. Nor do we know whether it will arrive quickly enough, or whether governments will allow us to capture it.

At one extreme, AI disappoints, debt continues rising and governments resort to increasingly aggressive financial repression or eventually encounter a fiscal crisis. At the other, AI materially accelerates technological progress, economies become substantially more productive and today's debt stocks progressively diminish in significance. Between those extremes lies an enormous range of outcomes, but at least there is now something that has been conspicuously absent from the debt debate for years: a credible escape route.

For it to work, three things must go right. The productivity dividend must be large enough. We must survive the dangerous period before it arrives. And politicians must resist protecting us from the very disruption required to produce it.

The technology many fear will impoverish us by taking our jobs may instead be our best opportunity to become productive enough to escape the debts we have accumulated. There is some hope. But the race is now between the debts we already know about and a productivity revolution whose scale we can barely comprehend.