Can we afford to support workers whose jobs are displaced by AI?

Not necessarily. But there is hope with the right policy choices.

A lot is currently being made about the impact of AI on work. “The bots are coming for your job!”. And early trends are worrying: some economists think there is already a link between falling employment and sectors exposed to AI, while leading firms are explicitly citing AI as a reason for cutting back jobs, and surveys suggest others expect to follow suit. But many AI evangelists, including Sam Altman, think there is an easy solution: if AI is so good it can take your job, it will probably also be so good that it will also cause a massive economic boom. We can just take some of the surplus from that - they argue - give it to the newly unemployed, and make sure everyone is better off. Presumably we would then all start book clubs and community gardens with our new free time while chanting “we love you Sam”.

I have more than a few questions about this story. Is this even a future we want? Maybe work is about more than just income, but also about connection, empowerment and purpose. Or at least work could be about those higher things. But let’s set these concerns aside for now and focus on a simpler question: is this a future we can afford? After all, today is budget day in the UK, and we all know you can’t spend without taxing first.

To balance the budget while making everyone better off, we need to believe that AI will drive more growth than the value of the wages of people who lose their jobs, and that the government will capture that wealth and redistribute it to those in need. That means the key question is not just whether AI makes the economy bigger, but whether it also makes the government richer. If we can’t tax the boom sufficiently, we have a problem.

Aside: does AI driven economic growth necessarily mean falling employment?

No. The best outcome would be one where AI drives a load of growth, but doesn’t take away people’s jobs: human augmentation rather than replacement. In 1817, the eminent economist David Ricardo argued that the industrial revolution would drive a collapse in employment for the working classes. This didn’t occur, instead machines replaced humans at the tasks they could do better than us (heavy manual labour), and human labourers found new - typically better paid - work elsewhere. If the same happens again, we won’t need a massive fiscal surplus to compensate the newly unemployed, as there won’t be many people in this category.

But we can’t be sure everything will be similarly fine this time around.

Firstly, the industrial revolution took time to sort itself out. British workers’ wages only began to consistently increase about half a century after the industrial revolution caused a step change in GDP growth. The interim period was brutal for many people, and pockets of misery lasted long after average wellbeing finally began to improve. Many suffered in Blake’s “dark Satanic mills” long after they began to put bread on the table for others. Even if the net impact of AI on employment turns out positive eventually, we may need to find our way through a difficult period of labour market churn while we get there.

Secondly, the industrial revolution involved machines that could compete only with human physical abilities. The intellectual domain remained exclusively ours. If AI increasingly competes with us intellectually too, it’s unclear whether there will be enough economic territory (embodied humanity???) left to build a high wage future on.

I don’t know what will happen, but we can’t rule out the possibility of rapid changes to people’s jobs that lead to a temporary or permanent increase in unemployment. If this might happen, we should prepare for it. It could be the single biggest political challenge of our generation.

So let’s return to a refined version of the question.

Will the UK government get more revenue from AI than it costs to support people through the transition?

I recently wrote a paper on the fiscal implications of rapid AI-driven growth for the UK, which includes a lot of detail on this question. If you have an hour or so free, you can read it below. The summary is that if employment stays high then very likely yes, we can afford it. If AI drives growth by replacing people without creating new jobs, then there is a risk we would need radical policy changes to raise enough revenue to support displaced workers, despite healthy headline GDP statistics.

The Fiscal Challenge Of Advanced AI

1.47MB ∙ PDF file

Download

Download

In this paper I identified three reasons that government wealth might not increase proportionately with AI-driven growth. If employment support is expensive, these increase the risk that we won’t have enough money protect displaced workers.

Risk 1: labour/capital tax differential

We currently tax labour at a much higher rate than capital. If an increasing proportion of our economy shifts from labour to capital, our effective tax rates will fall. This means that AI systems (a form of capital) may need to be much more productive than the people they replace just to break even on government revenue. It might be possible to fix this by simply increasing taxes on AI systems, but doing so faces a few challenges.

• How do you “increase AI taxes” in practice? Increasing all capital taxes might create problems elsewhere in the economy. A targeted AI tax increase addresses this issue, but means you need to define “AI”, (or perhaps more precisely for this purpose, “synthetic labour”) in an easy to apply, future proof way. I don’t currently know of a suitable definition, but think it’s worth trying to work out some good options.

• AI is currently at the heart of the UK’s industrial strategy, and one of our best bets for economic and productivity growth. Increasing tax rates could slow innovation right when we need it the most. If AI pushes up human wages this is good for tax revenues: it may only make sense to punitively tax AI if we are sure it is replacing people rather than augmenting them. In reality, in early stages of AI adoption it will be doing a bit of both in parallel, and working out which effect is more likely to win out long term may not be easy. This closely links to the second risk:

Risk 2: international tax and capital flows

AI is very internationally mobile, and like other digital services doesn’t always drive much local consumption. If a UK company pays a UK worker for a task, the worker pays a load of income tax, and also usually spends a lot of their wages elsewhere in the UK economy, attracting further taxes and increasing demand for other UK products. If the worker is replaced by an overseas AI system, firstly there is less headline tax (risk 1), but secondly that AI system is also not getting drunk at the weekend in UK bars, buying a new car in the UK market or paying tuition fees at a UK University. If we keep taxes on AI systems low enough, they might at least be developed and hosted in the UK, creating some domestic jobs and domestic demand. If we raise AI taxes to anything approaching the level of our current labour taxes we might not increase revenue, and instead just drive economic activity offshore.

It might be possible to solve challenge 2 by agreeing global minimum tax rates for AI systems. Something like the snappily titled OECD/G20 Global Anti-Base Erosion Model Rules (Pillar Two) would be a very good start. But the USA pulled out of these discussions in January, and without them on board, further international efforts might achieve very little beyond driving even more economic activity to the USA.

So until we can solve the international politics, the fix for risk 1 makes risk 2 worse, and the fix for risk 2 makes risk 1 worse. Maybe we can find a better balance than the current one, but it’s hard to see how we can fix the whole problem by tweaking tax rates alone.

Risk 3: government costs may rise faster than savings

But maybe this is all missing the point. Tax matters a lot, but there is (astonishingly) a policy world beyond it. Maybe if the government uses AI skilfully enough, it can reduce the cost of delivering public services to make the necessary savings to support displaced workers. After all, if we are assuming AI that’s capable enough to replace people in the wider economy, why shouldn’t it replace costly government workers too?

Unfortunately, for a given level of AI capability, the total possible savings available to government from using AI often work out smaller than the total possible costs. A savvy government may be able to chart a course where they capture the savings faster than the costs and deliver a net benefit overall, but this won’t be the default outcome.

For example, in an “AI replaces workers” scenario, the biggest cost to government may be unemployment support, while the biggest saving may be from firing government workers. Unfortunately, there are around five times as many private sector workers as public sector ones. Government would need to be innovating five times faster than the private sector, or spend five times less supporting a displaced worker than it did employing a government worker, just to break even. From what I know of public sector innovation, the former is a tall order. The latter might be doable, but not if we want to directly replace lost salaries (e.g. via a universal basic income).

Another way to look at this is to chart the extreme costs and savings of an unrealistically radical scenario (as they are easiest to calculate) to get a sense of the relative scale of potential costs or benefits. Replacing all existing unemployment benefits with a flat UBI based on the minimum wage for all currently employed people would cost about £609bn a year. (£22k per person * 34 million working adults = ~£750bn. Take away existing benefits of £141bn gives £609bn). In contrast, if we could deliver the three big hitters of health (£221bn) education (£112bn) and defence (£57bn) entirely for free with snazzy AI tools, this would only save £390bn.

These extreme scenarios are not a prediction about what will happen in a more realistic situation. I don’t believe all jobs will be automated any more than I believe you could replace the entire education sector with chatbots. GPT 5 is impressive, but I don’t think it can teach my one-year-old to share a tambourine with other children. What these extreme scenarios give is a sense of relative scale: the pool of potential costs is larger than the pool of savings, so we will have to deploy AI carefully if we want to catch more of the savings overall.

What does this mean for policy?

Rapid AI adoption does not automatically mean rich governments. We may need to actively intervene to make sure the government can capture its share of the benefits. Measures to support human employment, international coordination on capital taxation, and industrial strategy to ensure enough economic activity remains onshore are all important ingredients for success.

We are talking about a radical economic shift, and some of the policy responses we might need currently look too radical to implement. We also don’t yet know exactly how AI will develop and be deployed, so can’t be sure which interventions will be needed. By the time the impact of AI on labour markets becomes clear, it may be too late to carefully think through various options in detail: we may need to simply grab something that looks credible off the policy shelf, as we did for the Covid-19 furlough scheme. But there are some things we can do immediately, and some thinking we can start now to make sure there are good options on the “grab in case of emergency” shelf if we find we need them in a hurry.

• We can immediately move on international cooperation to increase minimum capital tax rates and tackle profit shifting from big tech companies. The OECD/G20 BEPS framework is a little clunky, but has the right goals. Continuing to support this is a good start, although unless the USA can be brought back on board, we may need to also start exploring alternatives.

• We could use better data on UK labour markets. The ONS do a decent job, but don’t have anything as granular as the US O*NET data that tracks occupation profiles at a task level. Granular understanding of what kinds of work people versus AI systems are doing would be hugely useful for policy design, and we can’t assume the UK will always closely match the US on this.

• Regulatory frameworks on advanced AI. Even if we don’t want to make strict rules now for AI providers, the ability to do so in future would be a useful option. Setting up a new regulator takes time: there is value in progressing on underlying organisational structures now so we are ready to deploy them quickly if needed.

• We need much more discussion on what to do about displaced workers. I worry UBI is too expensive (and creates too great a risk of unemployment lock-in effects), and government funded retraining is insufficiently effective. If large proportions of the population have to fall back on the existing welfare system, AI adoption will have been a failure. The options currently on the shelf are not necessarily good ones. I propose some alternatives on pages 27-29 of the paper linked above, particularly focussing around keeping employment rates high even if demand for human workers falls, and will come back to this in a future post.

I’ll be expanding on these ideas - and many more - in future posts. Subscribe to keep in the know! And comment below if there are things you are particularly keen to see me write about.