Austria is about to find out what happens when a member state lets large language models write legally binding government decisions, with no human required to check them first. We want to make this case as widely known as possible, not least because we're hoping to hear from readers elsewhere in Europe about how their own countries are handling, or planning to handle, the same question. More on that at the end.

This autumn, the Constitutional Affairs Committee of Austria's parliament will debate a sweeping change to the country's General Administrative Procedure Act (AVG). The AVG is the default procedural code for essentially all administrative proceedings in Austria, covering everything from social assistance and building permits to business licensing and public subsidies, unless a specific field of law sets out its own procedural rules, in which case that takes precedence. A single change to the general code can therefore affect how decisions get made across almost the entire administration, not just in one sector. Under the draft, authorities would for the first time be allowed to issue fully automated administrative decisions: legally binding rulings with direct effect on a person's rights, comparable to what other jurisdictions might call an administrative act or order, without a human ever reviewing the specific outcome before it goes out.

Different Technology, Same Law

Before getting into the legal detail, it's worth being precise about what's actually being automated, because the draft treats generative AI the same way earlier automation laws treated simple rule-based software. That's the root of almost everything that follows.

Classic automated administrative systems are deterministic: same input, same output, traceable logic. A system built on a large language model works fundamentally differently, in three ways that each map directly onto a legal guarantee:

• Hallucination. LLMs generate outputs that sound plausible but can be flatly wrong or entirely invented. Even specialised legal AI systems show hallucination rates of roughly 17 to 33%, and errors are often impossible for the affected person to spot. Hallucinated facts underpinning a binding decision touch the right to a fair trial under Article 6 ECHR: a person can't meaningfully exercise that right when the decision against them is built on evidence that doesn't exist.
• Black box. How the system arrived at a given output is structurally not fully explainable. This is not simply "not yet documented" but not explainable even in principle with current models. That sits uneasily with a basic rule-of-law demand: that the exercise of state power be transparent and reviewable, not just efficient.
• Bias. Systematic distortions are baked into how these models are trained and can't be engineered away by picking a "better" vendor. Left unchecked, this risks group-based unequal treatment that no one notices, because the system doesn't expose it, touching the principle of non-discrimination, which both EU law and the Austrian constitution guarantee explicitly.

These aren't implementation flaws. They're inherent to how current LLMs work, and every legal objection below follows directly from this technical reality.

Not hypothetical: two examples, months apart

Weeks before this autumn's parliamentary debate, Austria's own state-run AI assistant "ida" supplied a preview of the problem: it falsely described a sitting vice-chancellor as a former government member and invented a government minister who doesn't exist.

A more serious example came out of the UK in July 2026. An Upper Tribunal judge found that the Home Office had refused an asylum claim, from a woman who had fled forced child marriage and severe violence, by citing a country policy document that, on inspection, appears never to have existed. The judge noted the refusal letter bore the hallmarks of AI-generated content, suggested the reference was likely an AI hallucination, and warned that this would represent an extremely serious failing on the department's part if confirmed.

Read together, these two cases sketch an escalation from a chatbot inventing a minister's name in Austria to a fabricated document underpinning an actual refusal of protection in the UK, in exactly the kind of fundamental-rights-sensitive, factually complex proceeding that should never be handed to a system with no reliable way of checking its own claims.

Two New Powers, One Big Problem

The bill (89/ME) gives authorities two new powers that are each individually contentious:

• Chatbots (§§ 13, 13a AVG) that don't just inform citizens but can treat what someone tells the system as a formal submission in a proceeding.
• Fully automated decisions (§ 18a AVG) with real legal effect, with no human involved in the individual case. Which types of proceedings actually qualify for this is largely left to secondary legislation, a regulation, rather than being written into the law itself.

The second point runs into a structural problem that goes beyond AI risk management: under Austrian constitutional law, decisions this consequential are supposed to be made by parliament, through a law, not delegated to the administration to decide for itself. The draft hands that call to whichever "top authority" is materially responsible for the matter at hand, and that's a far more scattered group than it sounds. Depending on the subject, it could be a federal minister, a state government, or, since Austria's roughly 2,092 municipalities each administer their own local affairs, a municipal council. There is no single or small set of bodies deciding this: the power to define the scope of automation is spread across potentially thousands of separate decision-makers, each authorising the use of a technology whose risks they may have little capacity to assess, and each one also responsible for making sure their own system meets the AI Act's sector-specific requirements. Relying on thousands of potential decision-makers to each verify their own compliance strikes us as an inadequate safeguard. There's no exclusion in the law itself for sensitive areas like asylum or social welfare, either. That's left entirely to secondary legislation, case by case. 

Austrian courts have long required that a human official retain what's known as "decisive influence" over an automated decision for it to count as a genuine act of the state at all, a standard built for simple, rule-based systems decades ago. Nobody has a workable answer for what that standard even means applied to a model that cannot fully explain its own output. The draft's four safeguards, traceability, a testing phase, the ability to intervene, and ongoing checks, are standard IT governance measures for any public-sector deployment. They don't actually test whether "decisive influence" exists; they sidestep the question.

The draft also shifts risk onto the people affected by it. Citizens get two weeks to challenge an automated decision, a tight window when errors from an LLM are often invisible without close scrutiny, while the authority itself can revoke the same decision on its own initiative for up to two months, in some cases even without pointing to a specific error, simply citing a "systemic" one. The government keeps the option to fix its own mistakes later; the person on the receiving end carries the uncertainty in the meantime. Austria still has no independent AI oversight body, despite the EU AI Act calling for exactly that.

Turning an Insider Debate into a National Story

For months after the government bill was introduced in March 2026, debate around it stayed confined to the 35 submissions filed during the official consultation, a technical, insider affair. The bill has since advanced to a government bill and is scheduled for parliamentary debate this autumn and winter.

We still hope the process produces changes before it's finalised, which is part of why we published our open letter to the Constitutional Affairs Committee at the start of September. Within days, Austria's national broadcaster ORF covered the bill on television and online, and public radio interviewed the president of Austria's judges' association. This is no longer a niche debate confined to legal-tech circles.