Essay · AI, work and society

More AI should not mean fewer people

AI is changing work faster than our institutions can respond. Two ideas to make the system as ready as the technology, and why I think we should price the layoff, not the technology.

November 2026 · 7 min read

I build AI for a living. Most of my work is on power grids: forecasts, optimisation, tools that help operators decide faster. I believe in what this technology can do. That is exactly why the way it is being rolled out worries me.

The technology arrives in months. The systems meant to absorb its shocks, such as tax rules, unemployment insurance, training programmes and the political process itself, move in years. The people caught in that gap are the ones paying for it.

What the numbers already show

By the end of October, US employers had named AI as the reason for 121,401 announced job cuts this year. That is more than double the 54,836 for the whole of 2025. For several months this spring and summer, AI was the reason employers gave more often than any other.¹

These are stated reasons, and some companies surely use AI as a convenient story for cuts they would have made anyway. But the direction is hard to miss, and announced layoffs are only the visible part.

The less visible part is hiring. Researchers at Stanford’s Digital Economy Lab, using payroll data that covers millions of US workers, find that employment of 22–25 year-olds in the most AI-exposed occupations has fallen 16% relative to the least exposed.² The quietest layoff is the job that is never posted, and it removes the first rung of the career ladder.

At the same time, Amazon, Microsoft, Alphabet and Meta plan around $725 billion of capital spending in 2026, mostly for AI, up from about $410 billion last year.³ The IMF estimates that about 40% of jobs worldwide, and 60% in advanced economies, are exposed to AI.⁴

The money is moving fast. The safety net is not.

Why the system is not ready

Three things make this shock different from the ones our institutions were built for.

Speed. Unemployment insurance was designed for cyclical downturns: you lose a job in a recession and find a similar one when the economy recovers. AI is structural. The job may not come back, and the next one may need skills that take a year or two to build. Meanwhile, budgets are set once a year and elections come every four or five.

Incentives. Our tax systems quietly favour replacing people. For the United States, Acemoglu, Manera and Restrepo estimate that labour faces an effective tax rate of over 25%, while equipment and software face about 5% after the 2017 reform.⁵ Payroll taxes and employer contributions make every employee expensive; generous write-offs make software cheap. A finance team doing its job will automate more than is good for society, because the rules tell it to.

The tax base itself. Most welfare states are funded by taxes on work: income tax and social contributions. If AI shifts income from wages to profits, the money that pays for pensions, health care and unemployment support shrinks at exactly the moment more people need it.

None of this is an argument against AI. It is an argument that the system around AI has to be as ready as the technology.

Idea 1: keep people in work while jobs change

We have done this before. In 2020, Germany’s short-time work scheme, Kurzarbeit, kept almost six million people on payroll at its peak.⁶ Companies cut hours instead of jobs, and the state paid part of the lost salary. Norway’s permittering and similar schemes across Europe worked on the same principle. When demand came back, the workers were still there. One study estimates Kurzarbeit secured around 2.2 million German jobs.⁷

An AI version would look similar, with one important change. Short-time work was built for temporary shocks; AI is permanent. So the hours freed should not just be paid for. They should be used: paid retraining inside the same company, or for the next role elsewhere. Germany already has a version for structural change, the transfer short-time allowance, which supports workers for up to a year while they move on.

• For the company: keep the people, reduce hours for a period, and put the productivity gain into new work.
• For the worker: fewer hours for a while, most of the pay, and a real path into the next role.
• For the state: support that costs less than unemployment, and a tax base that stays intact.

Idea 2: price the layoff, not the technology

My first instinct was to cap how much a company can spend on AI as a share of its revenue. I have come to think that is the wrong lever. “AI spending” is hard to define: chips, cloud bills, consultants, your own engineers? A cap would punish companies that use AI to grow and hire, and it would push investment to countries without one.

South Korea’s experience is a warning. When it trimmed tax credits for automation equipment in 2017–18, robot installations fell by about 28% relative to Japan, but there was no measurable effect on employment.⁸ A blunt tax on technology slows the technology. It does not protect jobs.

The better target is the swap itself: the decision to replace people with AI while the business is doing well. Every large company already reports the numbers needed, namely revenue, headcount, wage bill and investment. So set a workforce baseline from the company’s own recent history, for example a three-year average of revenue and employees, and look at what happens next.

In plain words: more AI spending should come with fewer layoffs, not more. More revenue should mean more people, or at least the same people doing better work. A company that wants higher profit by cutting staff while AI does more of the work can still do that, but it should pay for the cost it pushes onto everyone else.

This is not as radical as it sounds. US unemployment insurance has been experience-rated for decades: employers who lay off more workers pay a higher rate. The workforce baseline applies the same logic to the AI era. And to the objection that companies will simply move: in 2021, more than 130 countries agreed a global minimum corporate tax. Coordination is hard, not impossible.

The design would need care. Thresholds, so that small firms are untouched. Room for genuine restructuring. Hiring counted alongside firing, so that freezing junior hiring does not become the loophole. A sunset clause and an honest review after a few years. But the principle fits in one sentence: AI may make a company more productive, but it should not let the company pass the cost of that productivity on to its people and the state.

The other bill

AI’s cost to society does not stop at jobs. The IEA expects data-centre electricity use to more than double, from about 415 TWh in 2024 to around 945 TWh by 2030, slightly more than Japan uses today.⁹ Rystad Energy estimates that data centres used about 222 billion litres of water for cooling in 2025, and that this could nearly triple by 2030.¹⁰

Working on power grids, I see this from the inside. New large loads need new lines, new substations and new generation, and those costs are often shared across everyone on the network. Water stress is local, and it falls on communities that never see the profits.

The principle should be the same as for jobs. Whoever profits from AI should pay its full bill: new clean supply to match its demand, water use reported and priced where water is scarce, and grid costs that follow the load that causes them.

I do not have every answer. These ideas will have holes, and people who know tax law and labour markets better than I do will find them. I would like them to. What I am sure of is the direction: the technology is ready for scale, and our systems are not. We should fix that before the next million jobs, not after.

Views are my own.