Necessary storage is on. Analytics load only after consent. Learn more
The Netherlands was ahead of the curve. Years before the EU AI Act entered the picture, Dutch government organizations were already publishing their algorithms in a national register. The Algoritmeregister, launched by the Ministry of the Interior, gave citizens a window into how automated systems influenced public decisions.
It was a good start. But it was voluntary, inconsistent, and often incomplete. Now the EU AI Act is about to change the rules entirely.
The Dutch Algorithm Register was always a transparency initiative. Organizations could decide what to register, how much detail to provide, and when to update their entries. Some municipalities like Amsterdam and Utrecht took it seriously. Others treated it as a checkbox exercise, listing systems with vague descriptions and no meaningful risk assessment.
The EU AI Act transforms this voluntary practice into a binding legal framework. Under Article 6 and Annex III, any AI system used in areas like social benefits, law enforcement, migration, or access to essential services is classified as high-risk. For government organizations, this covers a remarkable range of existing systems.
Think about what a typical Dutch municipality uses automated decision-making for: social welfare assessments, fraud detection in benefit applications, parking enforcement, permit processing, recruitment screening, and predictive policing models. Under the EU AI Act, many of these systems will need to meet strict requirements that go far beyond a register entry.
In February 2020, a Dutch court struck down SyRI (Systeem Risico Indicatie), a government system that combined data from multiple agencies to detect welfare fraud. The court ruled that SyRI violated the right to privacy under the European Convention on Human Rights. The system disproportionately targeted low-income neighborhoods and lacked adequate transparency safeguards.
SyRI became an international case study in algorithmic harm. But here is the uncomfortable truth: many municipalities continued running similar risk-profiling systems after the ruling. Rotterdam's fraud detection algorithm, which flagged benefit recipients based on behavioral patterns, drew scrutiny for similar bias concerns. Amsterdam quietly scaled back several surveillance algorithms after civil society pressure.
Editorial transparency
Zahed Ashkara is a lawyer, AI governance specialist, and founder of LearnWize. Factual and legal references link to the sources below and in the article. Always check the official publication for the current legal position.
Published on April 8, 2026
Start with the 5-minute scan and see where your AI literacy is not yet provable.
The EU AI Act makes the SyRI scenario legally untenable. Systems used for social benefit eligibility assessment fall squarely under Annex III, Category 5. That means conformity assessments, technical documentation, human oversight requirements, and mandatory registration in the EU database for high-risk AI systems.
The transition from Algorithm Register to EU AI Act compliance involves several concrete obligations that most government organizations have not yet started preparing for.
System classification is the first hurdle. Every automated decision-making system needs to be evaluated against the high-risk categories in Annex III. This is not always straightforward. A simple rule-based system that routes permit applications might not qualify as AI under the Act's definition. But a machine learning model that predicts fraud risk almost certainly does. The grey zone between these two extremes is where most municipalities will struggle.
Risk management systems become mandatory. Article 9 requires providers and deployers of high-risk AI systems to establish and maintain a risk management system throughout the AI system's lifecycle. For a municipality, this means documented risk assessments, mitigation measures, and ongoing monitoring for every qualifying system. Not once, but continuously.
Data governance standards must be met. Article 10 sets requirements for training, validation, and testing data. Government organizations using historical data to train predictive models need to demonstrate that this data is representative, free from bias to the extent possible, and appropriate for the intended purpose. Given that much government data reflects existing societal inequalities, meeting this standard requires active effort.
Human oversight is not optional. Article 14 requires that high-risk AI systems are designed to be effectively overseen by humans. In practice, this means that a fraud detection system cannot simply flag cases for automatic denial. There must be qualified humans who understand the system, can interpret its outputs, and have the authority to override automated decisions.
Transparency obligations expand significantly. Beyond the Algorithm Register, deployers must inform individuals when they are subject to a high-risk AI system (Article 13). The level of documentation required, including instructions for use, technical specifications, and performance metrics, far exceeds what most register entries currently contain.
One of the biggest practical challenges for government organizations is determining which of their systems actually fall under the EU AI Act's definition of AI. The Act defines AI systems broadly, covering machine learning, logic-based approaches, and statistical methods. But it also sets a threshold: the system must have "a degree of autonomy" in its operation.
This creates genuine ambiguity. A deterministic decision tree that applies fixed rules to benefit applications might not qualify. A system that uses machine learning to assess risk profiles almost certainly does. But what about a system that uses statistical scoring combined with manually defined thresholds? What about a chatbot that directs citizens to services?
The European Commission's guidelines on the definition of AI systems, published in early 2025, provide some clarity. But municipalities will still need to make judgment calls, and those judgment calls should be documented, defensible, and conservative. When in doubt, treating a system as potentially high-risk is the safer path.
Government organizations that have not started preparing are running out of time. Even with the potential deadline extensions proposed in the Digital Omnibus, which could push some high-risk requirements to late 2027 or 2028, the Algorithm Register obligation and the fundamental risk assessment work cannot wait.
Start with a complete inventory. Not just the systems in the Algorithm Register, but every automated process that influences decisions about citizens. Include the ones that nobody thinks of as AI, the scoring models buried in legacy software, the vendor-provided tools that nobody fully understands, the Excel models with macros that have grown into decision-support systems.
Classify each system against Annex III categories. Be honest about what constitutes a high-risk system. Document your reasoning. Engage legal counsel and domain experts in the classification process, not just IT.
For systems classified as high-risk, begin the conformity assessment process. This includes technical documentation (Article 11), quality management systems (Article 17), and registration in the EU database (Article 49). These are not tasks that can be completed in a few weeks.
Invest in AI literacy across the organization. Article 4 of the EU AI Act requires that personnel involved in operating or overseeing AI systems have sufficient understanding of the technology. For government organizations, where frontline workers interact with algorithmic outputs daily, this is a particularly urgent requirement.
Compliance is often framed as a burden. But for government organizations, the EU AI Act represents something more significant: a chance to rebuild public trust in algorithmic decision-making.
The SyRI ruling showed what happens when government algorithms operate in the shadows. The Algorithm Register was a first step toward transparency. The EU AI Act provides the framework to go further, ensuring that automated systems used in public administration are fair, documented, monitored, and subject to meaningful human oversight.
The municipalities that embrace this framework proactively will not just avoid fines. They will demonstrate to citizens that their government takes algorithmic accountability seriously. In an era of declining institutional trust, that matters more than any compliance certificate.