What Fälldin Knew

On nuclear waste, AI systems and the governance principle we keep forgetting.

I was a child when Sweden split in two over nuclear power.

Not metaphorically. Literally. In the small community where we lived, the one closest to the plant where my father worked, you could tell where people stood by what they wore. Pins. Symbols. For and against. The adults around me carried them, and so did the children at school.

My parents had moved there because of the jobs. Engineers and economists, as people were in that kind of place. We were the newcomers, the incomers, the people whose presence was bound up with the plant itself. Many of the families who had lived in the community for generations were less certain. Some were openly opposed. The tension was not abstract. It lived in the staffroom, at the bus stop, across the fence.

I was too young to understand the policy debates. But I understood, the way children do, that something large was being decided. Something that the adults around me could not fully control, and that depended on a question no one had quite answered: what do you do with what remains?

That question has stayed with me.

Years later, when I came to work at SKB, the company responsible for managing Sweden’s nuclear waste on behalf of the nuclear industry, the threads pulled together. A senior colleague, a project manager who had been in the industry long enough to have lived its entire history, told me the story of Villkorslagen over dinner at the ironworks manor in Forsmark. The room was old. The walls held something of the industrial century. And he began, with the easy confidence of someone who has told a story many times but still means it:

We can thank Fälldin for the fact that we are all sitting here today.

It took me a moment to understand what he meant.

Torbjörn Fälldin was Sweden’s Centre Party prime minister, elected in 1976 partly on the promise of phasing out nuclear power. He had campaigned against expansion. For the industry, for the families in communities like the one I had grown up in, he was a source of frustration and uncertainty.

The Stipulation Act of 1977, a Swedish law that made new nuclear reactors conditional on proving a safe solution for waste, was seen by many inside the industry as a political victory for the anti-nuclear movement: you cannot fuel a new reactor until you can demonstrate an absolutely safe method of disposal for the waste it produces.

For the industry, it felt like an obstacle. A demand that could not be met. A way of blocking expansion under the guise of safety.

My father worked at the nuclear plant at this time. The frustration inside the organisation was real. Fälldin was seen as having won a round in a debate that had already split communities like ours. The Villkorslag was, to many inside the industry, a political instrument dressed as a technical requirement. And then came the referendum. Nothing would be decided until it was over. The wait stretched out. The uncertainty deepened.

What the project manager was telling me, thirty years later, was that the demand had been the making of the industry.

Because in trying to prove that it could handle the waste, in responding to the political ultimatum by building SKB, developing the KBS method, constructing interim storage, and eventually applying to build one of the world’s first permanent repositories for spent nuclear fuel in Forsmark bedrock, the industry did not just solve a technical problem. It built the condition for its own legitimacy.

Fälldin did not intend to save nuclear power. He intended to stop it.

But the structural demand he created produced something durable:

show us how you handle the consequences before you are allowed to scale.

I have been thinking about that structural demand ever since.

The Nuclear Waste Fund holds approximately 83 billion SEK today (i.e. 83 miljard kronor).

It is financed by a fee per kilowatt-hour of nuclear electricity, accumulated over time and regularly adjusted based on updated cost estimates by the Swedish Radiation Safety Authority.

When I worked at SKB, that process was not abstract. It was real: engineers calculating decommissioning timelines, geologists assessing bedrock, economists modelling fund returns over decades. The obligation does not end when a reactor shuts down. Under Swedish law, the liability remains until it is fully discharged.

It is the polluter who pays. Not future taxpayers.

That sentence, from the Swedish National Debt Office, is the clearest formulation I know of what producer responsibility actually means in practice. It is not a philosophical position. It is an institutional design.

The fee per kilowatt-hour is the mechanism by which a consequence that would otherwise be invisible, the long-term cost of waste that remains radioactive for tens of thousands of years, is made visible inside the economics of every unit of electricity produced.

You cannot ignore it. It is in the price.

I no longer work at SKB. But the question followed me.

It has followed me into another field.

The AI systems being deployed today are not radioactive. Their consequences are harder to see, and easier to ignore.

But they produce something that functions like waste. Something that does not appear on any balance sheet, that is treated as external, that becomes someone else’s cost.

A hiring algorithm filters out applicants from certain postcodes. It does this quietly, systematically, at scale. The people it filters out never receive a rejection letter. They simply do not hear back. The company’s financial statements show improved efficiency. The cost, a career that did not start, a family that did not stabilise, a municipality that eventually carries the weight, appears nowhere in that accounting.

A content ranking system learns that certain kinds of political content produce more engagement. It amplifies it, not because anyone decided to, but because that is what it was optimised for. Over months, over years, the quality of what people believe about each other shifts. There is no moment when the harm becomes visible. There is no quarterly report that registers it.

A logistics company replaces a layer of workers with automated scheduling. The workers are offered retraining for roles that are, eighteen months later, also automated. The company reports productivity gains. The cost of the transition, extended unemployment, strained families, a town that loses its economic centre, is absorbed by the welfare state, by the municipality, by the individuals themselves.

The system performed exactly as designed.

The question is only what it was designed to ignore.

There is a word missing from most conversations about AI accountability.

Not product. Not tool. Not technology.

Infrastructure.

We have a clear intuition about how infrastructure is governed. Roads, electricity grids, water systems, telephone networks. We did not leave these to the market alone, not because we distrusted markets, but because we understood that infrastructure shapes the conditions under which everything else happens. Who gets access. On what terms. At what cost. Who is left outside.

In Sweden, we built state-owned energy companies. We invested in fibre networks in communities the market would not reach. We treated certain kinds of connectivity as a public good, not a private service. The outcome, better coverage, lower prices, more equitable access, was not an accident. It was a design choice.

AI is not a product in the way a smartphone is a product. It is becoming the operating environment for hiring, credit, healthcare, education, political information, and the delivery of public services. The systems that run on AI infrastructure determine who is seen and who is not. Who is routed toward opportunity and who is filtered away from it. Who counts as a valid participant in the economy and who, for reasons no one has clearly articulated, does not.

That is not a product.

That is infrastructure.

And we have never, in any previous technology transition, been comfortable leaving infrastructure entirely unaccountable to the people it shapes.

The nuclear waste fund exists because we understood that an industry operating at civilisational scale cannot be allowed to externalise its consequences. The fee per kilowatt-hour is not a tax. It is a recognition that the cost of production includes what the production leaves behind.

The same logic applies here. The cost of deploying AI at scale includes what the deployment leaves behind. The question is only whether we build that recognition into the system, or wait until the leaving behind becomes impossible to ignore.

I want the structural demand.

Not the ethics committee. Not the voluntary framework. Not the annual report that acknowledges risk without carrying it.

The kind of demand that Fälldin made.

Show me, before you scale, how you intend to carry what you leave behind.

In nuclear waste management, we call that the condition for a licence. In packaging, we call it producer responsibility. In Sweden’s deposit return system for bottles and cans, we built it so deeply into everyday life that most people no longer think of it as regulation.

We have done this before. Phased out substances before their full cost had unfolded.

We can call it whatever we want for AI.

The name matters less than the architecture: that the cost of consequence is built into the cost of production, that it scales with deployment, and that it does not disappear when the model is deprecated or the company restructures or the political winds shift.

The most honest objection is also the most practical one. Nuclear waste is physical. It can be measured, contained, and traced. Algorithmic harm is diffuse. A content ranking system does not produce a measurable byproduct in the way a reactor does.

But we measure complex, diffuse harm in other contexts. Carbon pricing attempts to quantify harm distributed across the atmosphere. Extended producer responsibility for packaging calculates fees based on material tonnage introduced into the market.

The metric will be imperfect. The architecture is what matters.

I remember how we used to say it at SKB, to politicians, to anyone who asked why we could not wait:

we do not pass this burden to the next generation.

Not as a slogan. As a constraint.

The cost does not disappear. It is carried.

That is what Fälldin understood.

And we have understood it before. In other industries. With other technologies. At precisely this moment: when the damage was still in the future, when the industry insisted the demand was unreasonable, when the political pressure to simply move forward felt irresistible.

We held the line then.

The question is whether we are willing to hold it now.

21/4 2026 :

Since writing this, I keep returning to the same question.

Not just what we build,

but what carries the consequences over time.

I explore that further here:

What Carries the Consequences

And I keep coming back to a simple test.

Not whether a system works,

but whether someone is still there to take responsibility for what it leaves behind.

---

Sara Eson writes about AI, accountability, and democratic governance. She has worked in nuclear waste management and regulated industries. She is developing Civic Human Value Accounting, a framework for making the human costs of AI systems visible and someone’s responsibility.

saraeson.substack.com