Compared to What?
The fears about AI are familiar. Jobs disappear. Power concentrates. We hand over decisions we no longer understand. Eventually, we lose control.
I take these seriously. I've written about some of them: how AI quietly takes over inside companies one draft at a time.
But every argument about risk contains a hidden comparison. When we ask whether AI is dangerous, we're asking: dangerous compared to what? And the usual answer, never stated, is the present. Today, held still. As if the alternative to AI were the world we already have, continuing as it is.
That world isn't on offer.
The future we're already in
I live in Greece. If you want to see the default future of a wealthy democracy, you don't need a model. Look around.
We have one of the oldest populations in Europe and a fertility rate far below replacement. Our population is already shrinking, and it is projected to keep shrinking for decades. Public debt is well above the size of the whole economy. A generation of young, educated people left during the crisis, and many didn't come back. Every year there are more people who need care and fewer people of working age to provide it, or to pay for it.
Greece is not an outlier. It's early. Italy, Spain, Portugal and much of Central and Eastern Europe are on the same path, some on an even steeper one.
The OECD estimates that demographic pressure alone could cut growth in GDP per person across its members from about 1% a year in 2006–2019 to 0.6% in 2024–2060. That sounds small. It isn't. Compounded over those 36 years, it leaves GDP per person roughly an eighth lower than the old trend would have. And that's the average. It includes younger, faster-growing members like Mexico, Turkey and Colombia. Greece starts from a worse position.
This creates a particular problem for rich countries. Poorer countries are trying to reach a better life. We're trying to keep one. A standard of living built on pensions, public health systems and decades of rising expectations has to be paid for every year, by fewer and fewer people.
You feel it in ordinary ways. The appointment three months out. Higher taxes and worse services at the same time. A 30-year-old wondering how many promises they're expected to fund.
I don't see a path for countries like Greece to keep what they have without AI. Without a step change in what each person can produce, and in what machines can do, we are more or less doomed to a slow decline: smaller pensions, longer waits, worse care, and children who leave for somewhere that still works.
The obvious reply is Japan. It has been aging, indebted and slow-growing for thirty years, and it is still one of the safest, cleanest, most functional countries in the world. If that's decline, plenty of people would take it.
Look closer. Japan has drawn more women and older people into work, and now foreign care workers it once kept out. Families fill the rest, often by someone quitting their job to look after a parent. Even so, it will need about 570,000 more care workers by 2040 than it had in 2022, just as its care workforce has started to shrink. What keeps Japan steady is a cushion: its own currency, debt owed mostly to its own citizens, decades of savings abroad. But a cushion buys time, not escape. Japan isn't the exception that disproves the decline. It's further down the same road, with more padding. Greece is on that road with far less room to manoeuvre.
Standing still is also a bet. We just never debate it, because it doesn't feel like a decision.
When capacity comes loose from headcount
The standard responses to aging all help. Raise the retirement age. Encourage immigration. Borrow. Reform. But none of them closes a gap this size, and the countries that send us workers are aging too. An aging society runs out of people.
Machines have been stretching what each worker can do for two centuries. AI extends that to work machines could never touch before. For most of history, expanding what a society can do meant finding more people to do it. More nurses, more inspectors, more engineers, more hands. AI starts to loosen that link.
I see it in my own work. Small teams at our company now ship what used to need departments. In my evenings, I build games that once would have taken a team and two years. Capacity is coming loose from headcount.
That's the easy half. It's happening on screens, and screens are not enough. The work an aging society needs most is physical: lifting, washing, feeding, turning someone in bed at three in the morning. No amount of software does that.
This is where robotics stops being a side story and becomes the main one. Japan bet on care robots twenty years ago, with subsidies and national programmes, and the results are instructive. Adoption has been slow. Where robots were introduced, they didn't replace staff; homes that used them kept workers longer and, by some measures, delivered better care. That's real progress. But it's assistance, not substitution: the robot helps the carer, it doesn't do the care.
Helping a frail person out of bed is far harder than a repetitive task in a controlled factory. Closing that gap is the most important engineering problem of the next thirty years, and we talk about it far less than chatbots.
The objection I owe myself
I've argued before that more AI can mean less output: individuals get faster, the bottleneck moves, and the organisation absorbs the gain. Across an economy, the same thing happens by default. AI added on top of unchanged hospitals and ministries produces more paperwork, not more care.
There's a second, harder problem. Economists call it Baumol's cost disease. Care is the sector where productivity barely moves. A string quartet needs four players today, just as it did two hundred years ago, and bathing an old man takes as long as it did in 1950. So while factories get more efficient, care gets relatively more expensive every decade.
The usual answer is to make carers more productive. Let AI handle the charting, the scheduling, the paperwork, and give nurses their shifts back. That's worth doing, and some of it adds real capacity: a lift that lets one carer do what used to take two is capacity. But Japan's evidence so far shows what assistance mostly buys: better care and staff who stay longer.
Now look at the size of the problem. The number of very old people keeps growing faster than the number of people of working age, year after year. The test is whether productivity can keep pace with that, not once but every year for decades. Better paperwork won't. You cannot paperwork your way out of a missing generation.
Nor does growth elsewhere. AI in factories, labs and offices will make us richer, and that matters. Money can raise pay, improve training and buy equipment. It can't conjure people who don't exist. What closes the gap is a large, sustained fall in the human hours it takes to care for one person. In practice, that means machines taking over whole tasks: lifting, bathing, watching over someone through the night. The machine doesn't have to work alone. It has to take the work.
I know how that sounds. The idea of a machine bathing a parent makes many people uncomfortable. But that isn't the real choice. The real choice is between a machine and no one, or between a machine and a daughter who gives up her career to do it. In Japan, around 100,000 people a year already leave their jobs to care for family.
That's the uncomfortable shape of the optimistic case. It doesn't rest on the AI that already works, the kind that writes, summarises and codes. It rests on the AI that doesn't work yet: robots safe and capable enough to be trusted with a human body. If that arrives in time, an aging society can care for its people. If it doesn't, the rest will make us richer, but it won't make us cared for.
Who owns the machines
There's one more way this goes wrong, even if everything above works.
Care robots, if they come, will be expensive. If only private buyers can afford them, wealthy families get help at home while public care systems keep the waiting lists. The same holds across the economy. If a few companies own the productive machines and keep most of the returns, output rises and households don't feel it. GDP grows. The waiting list stays the same. People are told to celebrate productivity they never see.
So the test shouldn't be growth. It should be whether ordinary things get better: the appointment comes sooner, care at home becomes affordable, a 55-year-old in a physical job doesn't have to wreck their body to reach retirement. If the gains don't show up there, through competition, prices, wages or public services, then it didn't go right. It just went well for someone.
Compared to what
The risks of AI are real. Most of them are about how we build and deploy it: who controls it, who it replaces, what we let it decide. We may fail to govern them; I've written about how quietly that failure can happen. But they are risks. Decline is different. It isn't a risk we might run. It's the path we're already on, and it arrives whether we choose it or not. I'd rather take on the hard problem of governing a powerful technology than accept a loss we've already stopped arguing about.
The alternative was never safety. It was a slow decline we've already priced in and stopped noticing. The other responses help, and we should use all of them. I don't believe they're enough. Not building this is the bigger bet.