Algorithm register: publishing and maintaining algorithms

The government’s Algorithm Register, at algoritmes.overheid.nl, makes it possible to find out where and how public authorities use algorithms and AI. It aims to explain algorithms more clearly and make their use and impact easier to understand. If you disagree with the use of an algorithm, the register shows you where you can lodge an objection.

The first version went live in December 2022. The register relates to our work on data governance.

According to the register, these include at least AI systems classified as high-risk under the AI Act, and algorithms that directly affect people, such as those used in decisions about fines or grants. The Algorithm Framework also considers an algorithm to be impactful if it determines how the government categorises or approaches citizens or organisations, for example in relation to risk or fraud.

If you are unsure, the Algorithm Framework says to treat the algorithm as impactful.

There is no legal obligation yet. The register states that submission is not currently mandatory for public-sector organisations, but that this requirement is coming. The Algorithm Framework from the Ministry of the Interior and Kingdom Relations describes it as established policy for public institutions to publish their impactful algorithms and high-risk AI systems, unless grounds for exclusion apply.

Your legal adviser can assess what this means for your organisation. We do not provide legal advice.

The publication standard sets out the fields. A published description includes details such as purpose and impact, considerations, human oversight, risk management, data, technical operation, supplier, status and impact assessments such as a DPIA. According to the register, the fields are not yet mandatory, but organisations are encouraged to publish as much as possible.

You can address the privacy aspects in a DPIA for AI.

Waterschap Limburg describes its discharge forecasting model for the River Roer: its inputs are measured discharge, recorded and forecast rainfall, and evaporation; its outputs are hourly forecasts. The section on human oversight states that the model supports decision-making and that an expert interprets its output. We did not write that description.

You can read more about models for water management in AI in water management.

According to the register, each organisation decides that for itself. In practice, responsibility often sits with a CISO, CIO, data specialist or data protection officer, or with the department responsible for the algorithm. An initial step is often to inventory the algorithms; for purchased applications, you can ask the supplier.

Read about how to define these roles in AI governance.

According to the register, the Dutch register will continue alongside the database established under the AI Act, because it also includes algorithms that do not learn independently and algorithms with a lower risk level. Work is under way to align the two, so that public authorities with a high-risk system in the register can more easily comply with the Act.

Read more in AI Act, or, for ministries and municipalities, in AI consultant in The Hague.

When you have just a handful of simple calculation rules whose owner understands how they work, or when the supplier already has a template for registration. The register recommends checking this. Its implementation team can then help you publish the entry.

Technical help only becomes worthwhile when models are difficult to explain or change frequently, as with Rijkswaterstaat’s AI models.