Austria's plan for AI government decisions needs a human brake

Austria is about to cross a line no democracy should cross first. This autumn, the Constitutional Affairs Committee of Austria's parliament will debate a sweeping amendment to the country's General Administrative Procedure Act that would let authorities issue fully automated administrative decisions — legally binding rulings on social assistance, building permits, business licenses, and public subsidies — with no human ever reviewing the individual outcome. If it passes, Austria becomes the cautionary tale for AI government decisions across Europe, and the generation that will live under this bureaucracy longest is ours.

The draft does not merely allow automation at the margins. It treats generative large language models the same as deterministic rule-based software, as if a system that predicts plausible text were equivalent to one that follows fixed rules. Which proceedings qualify would be left to future secondary regulation — with no statutory exclusion for sensitive areas like asylum or welfare. Citizens would get only two weeks to challenge a decision, even though errors produced by a language model are often invisible: a fluent, confident, entirely wrong paragraph looks exactly like a correct one.

Why AI government decisions are a civilizational no-go

A government decision is not a customer-service reply. It must be lawful, reasoned, and reviewable — a person whose benefits are cut or whose asylum claim is rejected has a right to know why, in terms a court can test. AI government decisions break that chain at every link. Language models hallucinate; even in legal tasks, published research has documented fabrication rates between 17 and 33 percent. They cannot explain their reasoning, because there is no reasoning to explain — only statistical pattern-matching. And they absorb the biases of their training data, which means the people already most likely to be disbelieved by bureaucracies become the most likely to be misjudged by its machines.

The United Kingdom is already showing where this road ends. This summer, an Upper Tribunal judge found that the Home Office had refused an asylum claim while citing a policy document that apparently never existed — evidence the court ruled could not be trusted, the legal equivalent of a fabricated citation. That was one human-led case with lawyers able to spot the fabrication. Multiply it across an entire administration where no human reviews the outcome, and the error does not get caught; it gets mailed, stamped, and enforced.

Speed is not the same as justice

The honest argument for automation is backlog: administrations are slow, caseworkers are overwhelmed, and software is cheap. But the purpose of administrative procedure was never speed — it was legitimacy. The reason a decision letter carries a name, a legal basis, and an appeal path is that the state's power over an individual demands accountability. Fully automated AI government decisions keep the power and delete the accountability, replacing a civil servant who can be questioned with a model that cannot be cross-examined.

There is a telling asymmetry in who bears the risk. The officials who deploy these systems will not be the ones whose disability benefits hinge on a model's output; the politicians who vote for the amendment will not be racing the short appeal window themselves. The cost of every hallucinated denial lands on the person least equipped to fight it — often young, poor, newly arrived, or navigating the system in a second language. That is not modernization of AI government decisions. It is the outsourcing of the state's duty of care to a statistical parrot.

The human-final-call rule is the minimum

Ahead of the debate, the digital-rights group epicenter.works sent an open letter to parliament laying out why fully automated rulings with no human review are incompatible with fundamental rights, according to reporting by European Digital Rights on October 8. Their demand is modest and should be non-negotiable: a human must make the final call. Not a rubber stamp on a machine's output, but a genuine review by a person who can be held responsible for the result.

Austria is not debating whether AI can assist caseworkers — it already does, and sensible guardrails for that exist. The question is whether a democracy hands the gavel to software that cannot understand what it decides. AI government decisions without human review are not an efficiency upgrade; they are a decision that the rule of law is optional when it is inconvenient. Parliament should reject the draft, and every other country watching should treat it as the warning it is. Read the full case against the plan at European Digital Rights.