The pitch for artificial intelligence usually arrives with a promise attached. Feed the system enough data and it will find the fraud, spot the illness, sort the applications, and do it faster and more fairly than any tired human could. It is a seductive story, and the Nordic countries are a good place to test it, because they are further down the road of digital government than almost anywhere else and have been feeding real decisions to algorithms for years.

What that experience shows is less a triumph than a caution. AI is a genuinely useful tool. It is not a magic bullet, and the difference matters most exactly where the promise is loudest.

Denmark automated a hard question, and did not make it easier

Consider welfare fraud. Denmark’s benefits authority, Udbetaling Danmark, uses up to 60 algorithmic models to flag people for possible fraud investigation, a system built with the pension body ATP and private contractors.

In November 2024, Amnesty International published a report called Coded Injustice, based on partial access to four of those algorithms. Its conclusion was blunt. According to Amnesty, the system risks discriminating against people with disabilities, the low-paid, migrants, refugees, and marginalised racial groups, and amounts to a form of mass surveillance that leaves claimants, in one interviewee’s phrase, feeling as though a gun is permanently pointed at them.

The detail is where the point lives. One tool, the “Really Single” algorithm, tries to predict whether a benefits claimant is genuinely living alone, using parameters that include “unusual” living arrangements. But plenty of people live in unusual arrangements for entirely ordinary reasons: couples with disabilities who live apart, older partners who keep separate homes, multi-generational migrant households. Another model, “Model Abroad”, flags people with stronger ties to non-EEA countries. Amnesty argues this discriminates on national origin, and that the wider system may qualify as a prohibited “social scoring” tool under the European Union’s new AI Act.

Udbetaling Danmark disputes this. It says its data use is legally grounded and rejects the claim that its system is social scoring, though Amnesty notes it declined a collaborative audit. The dispute is unresolved, and worth reporting as a dispute rather than a verdict.

But the underlying lesson does not depend on who wins it. Deciding who is “really single” is a hard, human judgement full of edge cases. Automating it did not dissolve the difficulty. It buried the difficulty inside a model, applied it to millions of people at once, and made the resulting bias harder to see and harder to challenge.

Why the bullet keeps missing

This is the recurring shape of the disappointment. The hardest part of most problems AI is sold to fix is not the calculation. It is the messy, contested, human definition underneath: what counts as fraud, as risk, as a good candidate, as a suspicious pattern.

An algorithm does not resolve that question. It inherits whatever answer its designers and its training data already contained, including their blind spots, and then repeats it at scale and at speed. When the underlying judgement is sound, that is powerful. When it is biased or simply wrong, the machine does not correct the error. It industrialises it.

None of this makes the technology useless. It makes it a tool whose value depends entirely on the quality of the thinking around it, which is precisely what the phrase “magic bullet” invites people to skip.

The more useful Nordic export

If there is a hopeful counterpart to the Danish story, it comes from Finland, and it is notably unglamorous.

In 2018 the University of Helsinki and the technology company Reaktor launched a free online course called Elements of AI, with the initial aim of teaching one per cent of Finns, roughly 54,000 people, the basics of how the technology actually works. Within four months nearly 90,000 people from 80 countries had enrolled, and it later expanded across the European Union.

The instructive part is the goal. Teemu Roos, the lead instructor, described the aim as wanting to “demystify technology and dispel unnecessary fears through education”. Not to sell AI, and not to fear it, but to understand it well enough to ask sensible questions of it. That is the opposite of magic-bullet thinking, which depends on the audience not looking too closely.

Treating it as infrastructure, not alchemy

Put the two Nordic examples side by side and a reasonable position emerges. Denmark shows what happens when a powerful automated tool is pointed at a sensitive human judgement without enough transparency or scrutiny. Finland shows the quieter work that makes such tools less dangerous, which is a population that understands roughly what they are and are not.

The honest framing is not that AI is a con or that it cannot help. It plainly can, in narrow, well-defined tasks with good data and human oversight. The framing to resist is the one that treats it as a substitute for judgement rather than an amplifier of it. A bullet is a thing you fire and forget. The systems now being wired into hospitals, banks, and welfare offices are closer to infrastructure, and infrastructure has to be inspected, maintained, and held to account, especially when it is deciding something about a person’s life.