Jacob Dexe is asking the right question. He just stops too soon.
In his piece in Helsingborgs Dagblad, Dexe documents how Gothenburg placed 450 children in schools that violated the regulatory framework, a consequence of an algorithm that had failed to account for the fact that a river runs through the city. He documents how the municipality then denied that the algorithm had made the decision, leaving the parents to prove something they could not prove. He documents how the Swedish Public Employment Service's head of development, when caseworkers protested against the algorithm's matching recommendations, suggested that the caseworkers ought to "question things a little less."
These are not anecdotes. This is a pattern. And Dexe, to his credit, is clear about that.
But his conclusion, that we need to answer how the AI-augmented administration should be built, treats what happened in Gothenburg as a design problem. Something that better policy, clearer guidelines, and more transparency tools could fix.
It's not a design problem. It's an incentive problem. And that distinction changes everything.
The Opacity Is Not Accidental
Think about what Gothenburg actually discovered. By positioning the algorithm as a "decision support system" rather than a decision-maker, the municipality achieved something valuable: it rendered accountability legally inoperative. No information had to be disclosed. No reasoning had to be explained. The parents bore the burden of proof for something the institution could simply deny.
Gothenburg most likely acted out of confusion, not calculation. What's interesting is what institutions witnessing the outcome learn from it.
This wasn't a failure of the system. It was the system working the way institutions, under pressure, tend to make it work. Opacity, once it proves to have legal utility, does not remain accidental. It becomes a function.
The Public Employment Service case makes this explicit. The proposed solution to the caseworkers' dissatisfaction was not to improve the algorithm, audit its edge cases, or build better feedback mechanisms. It was to reduce the humans' ability to contradict it. The algorithm was not being used to augment judgment. It was being used to replace judgment, while maintaining the fiction that judgment still existed.
This is the direction we are headed. Dexe describes the early stages. He doesn't name the endpoint.
Why "Explainability" Is Not the Answer
Dexe notes that you can ask an AI to explain itself, document the model's logic, publish training data, even release the code, but that these "do not yield reliable answers." That's correct, and it is the most important sentence in his text. But it deserves more than a paragraph.
Explainability tools are not limited to population-level statistics. SHAP, for example, assigns feature contributions to an individual prediction. That is not automatically a causal account of an inference chain or a legally sufficient justification for a decision. The difference is enormous legally, because administrative decisions are not statistical, they are individual. A citizen does not have the right to know that "distance was weighted at 0.34" in the model. They have the right to know why their child was sent to that school.
That is a gap better explainability tools cannot close. It is architectural. A system that produces outcomes through statistical pattern-matching across millions of parameters cannot, by construction, produce the kind of traceable, step-by-step logic that administrative law presupposes. You can wrap it in documentation. You cannot make it reason the way the law requires.
That means the choice is not between transparent and opaque AI in the state. It is between AI systems architected for accountability from the ground up, where inference chains are inspectable in real time, not reconstructed after the fact, and AI systems that, however well documented, are structurally incompatible with the rule of law as it currently stands.
Three Things the Strategy Refuses to Say
Sweden's AI strategy aims to be "best in the world at using AI to increase cost efficiency and raise the quality and service of public administration." Dexe quotes it. Neither he nor the strategy explains what "best" means when accountability breaks down.
In practice, it would have to mean three things.
First: a requirement of traceable reasoning as a condition of deployment. Not explainability as an after-the-fact add-on, but architectures where the inference chain is auditable at the level of the individual. Systems that fail that requirement do not belong in processes that affect citizens' rights, however efficient they may be.
Second: disclosure as default, with reversed burden of proof. No institution should be able to deny an algorithm's role in a decision that affects a citizen. The moment a system participates in a decision process, that participation shall be disclosed, documented, and subject to audit. The citizen does not prove the algorithm existed. The institution proves the human actually decided.
Third: a prohibition on using AI to shrink discretion rather than expand capacity. The Public Employment Service case is the archetype of the wrong direction. Human discretion in public administration is not inefficiency, it is the mechanism through which rules meet reality. Replacing it with a model trained on historical data does not improve judgment. It encodes the past's patterns as the future's constraint and calls it progress.
Dexe closes his piece by asking on what principles the AI-augmented administration should be built. That is the right question.
The answer is not more guidelines, more reporting requirements, or better data sharing between agencies. Those are answers to a different problem, the adoption problem. The problem Sweden actually has is that the fastest path to AI in the state is also the path that systematically tears down the accountability infrastructure it took generations to build.
Best in the world cannot just mean getting there first. It has to mean getting there right.
: The description of individual-prediction explanations has been corrected.