Netherlands Algorithm Register: transparency and AI in municipalities

In the Netherlands, a municipality that uses an algorithm to sort applications, detect anomalies or plan inspections can describe it publicly in a national register. This level of transparency is rare enough in Europe to deserve a closer look, especially now that generative AI is arriving in public sector offices.

This article looks at how the Algoritmeregister works, what it tells us about responsible use of AI in municipalities, and why transparency about algorithms is not enough to protect the data that staff type into AI assistants every day.

A public register launched in December 2022

The first version of the Dutch government's algorithm register went live in December 2022 at algoritmes.overheid.nl. Its purpose is to make public sector algorithms easier to find, explain how they work and what they are used for, and allow the public to discuss them. The register started with 109 algorithms and listed 291 by mid-February 2024, according to figures reported by the specialist local government press.

At first, registration was voluntary. In a letter to parliament in December 2022, the government announced its intention to make it a legal obligation. The Autoriteit Persoonsgegevens (AP), the Dutch Data Protection Authority, coordinates the supervision of algorithms.

What municipal entries show

Municipalities are among the most active contributors. In early 2024, Amsterdam listed 40 algorithms, Utrecht 38, The Hague 29 and Rotterdam 11, while around forty other municipalities listed between one and five each. Above all, these figures show that transparency spreads by example, one municipality at a time, and that many local authorities had not yet published their high impact algorithms.

For a researcher or a security lead, the register is a valuable source. It shows which kinds of automated processing affect residents, what data they use and who is responsible. It also offers a documentation model that other organisations, public or private, can adopt.

The tools of the Dutch framework

Around the register, the government has built several tools for the responsible use of algorithms:

• the Algoritmekader (Algorithm Framework), which brings together the standards and measures to follow, developed by the Ministry of the Interior;

• the IAMA (Impact Assessment Mensenrechten en Algoritmes), a fundamental rights impact assessment, updated to reflect practice and regulation;

• the data protection impact assessment (DPIA) required by the GDPR for high risk processing.

These tools share one thing: they force you to write down why a system is used, with which data and under whose responsibility. That documentation discipline is exactly what is missing, in many organisations, from everyday use of generative AI.

The EU AI Act adds another layer: public authorities that deploy high risk AI systems will have to register them in an EU database. Our EU AI Act and GDPR Article 32 checklist sets out these requirements.

Generative AI, the register's blind spot

The register describes algorithms deployed by the organisation. It does not see an individual employee using an AI assistant to rephrase a letter, summarise a case file or translate a request. Yet that is where residents' personal data flows most easily: names, addresses, citizen service numbers (BSN), family or social circumstances.

The Dutch government has taken a position on this. In a provisional position presented to parliament on 11 December 2023, it said that non-contracted generative AI tools generally do not demonstrably comply with privacy and copyright law, and that central government organisations should in principle not use them where there is a risk of breaching the law. Each use case requires a DPIA and an algorithm impact assessment such as the IAMA. The text applies to central government; the association of Dutch municipalities (VNG) said that municipalities were likely to face comparable legal risks.

What municipalities can do in practice

1. Take stock of the AI assistants staff actually use, including personal accounts.

2. Document approved uses the way you document an algorithm: purpose, data, owner.

3. Carry out a DPIA for each use case involving residents' data.

4. Protect at the source: detect a BSN, an address or a name before it is sent to the AI and replace it with a fictitious value.

5. Keep evidence of incidents prevented, useful if the AP comes knocking.

Our first article on the Netherlands looked at AI chatbot data breaches reported to the AP. Those incidents show exactly what the register does not capture. To understand how they happen, see also the anatomy of an AI data leak and the real cost of shadow AI.

In summary

The Algoritmeregister makes the Netherlands a useful case study for any organisation that wants its algorithms to be understandable. But transparency about deployed systems must go hand in hand with control over everyday use of generative AI, where residents' data can slip out quietly.

For a municipality, the right question is therefore not only "which algorithms do we publish?" but also "what data do our staff hand to external tools every day?".

ZeusLock detects and anonymizes personal data in prompts and files before they are sent to AI assistants, and provides an incident log for your impact assessments. Request a demo for your organisation.