# The algorithm register examined: what governments report

[Skip to content](#lm-inhoud)Network/[NL](/en/het-algoritmeregister-doorgelicht-wat-overheden-melden)EN[Hubhub.llmnet.nlCompare models on task, language, cost and license.](https://hub.llmnet.nl/en/)[Communitycommunity.llmnet.nlPrompt techniques, patterns and system prompts.](https://community.llmnet.nl/en/)[APIapi.llmnet.nlLLMs in production: rate limits, routing, structured output.](https://api.llmnet.nl/en/)[Consultancyconsultancy.llmnet.nlRolling out AI in an organization, pilot to production.](https://consultancy.llmnet.nl/en/)[Newsnieuws.llmnet.nlAI developments, explained for the Netherlands.](https://nieuws.llmnet.nl/en/)[Benchmarkbenchmark.llmnet.nlMeasure AI quality yourself, on your own tasks.](https://benchmark.llmnet.nl/en/)[Careersvacatures.llmnet.nlAI roles, salaries and career paths in the Netherlands.](https://vacatures.llmnet.nl/en/)[Learnleren.llmnet.nlAI concepts in plain language, beginner to builder.](https://leren.llmnet.nl/en/)[Guidegids.llmnet.nlRun AI privately on your own Mac, PC, NAS or home server.](https://gids.llmnet.nl/en/)[Directorydirectory.llmnet.nlMapping the AI ecosystem: tools, models, companies.](https://directory.llmnet.nl/en/)[Radarradar.llmnet.nlSignals from X, research and communities for indie developers.](https://radar.llmnet.nl/en/)[Appsapps.llmnet.nlReviews of AI apps and open-source repos, with tips for builders.](https://apps.llmnet.nl/en/)[llmnet.nl — main site](https://llmnet.nl/en/)[](https://x.com/intent/post?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fhet-algoritmeregister-doorgelicht-wat-overheden-melden&text=The%20algorithm%20register%20examined%3A%20what%20governments%20report)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fhet-algoritmeregister-doorgelicht-wat-overheden-melden)[](https://www.reddit.com/submit?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fhet-algoritmeregister-doorgelicht-wat-overheden-melden&title=The%20algorithm%20register%20examined%3A%20what%20governments%20report)[](#)[](https://x.com/intent/post?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fhet-algoritmeregister-doorgelicht-wat-overheden-melden&text=The%20algorithm%20register%20examined%3A%20what%20governments%20report)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fhet-algoritmeregister-doorgelicht-wat-overheden-melden)[](https://www.reddit.com/submit?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fhet-algoritmeregister-doorgelicht-wat-overheden-melden&title=The%20algorithm%20register%20examined%3A%20what%20governments%20report)[](#)

 
# The algorithm register examined: what public bodies report

 By Ivo Donker — compiled with AI assistance (Claude & Gemini)

 The Dutch government's Algorithm Register was set up as the central window on automated decision-making within the public domain. After public outrage over opaque risk models and data linkages in benefit checks and allowances, this register had to restore trust. Citizens, journalists, and regulators gained insight, through one central portal, into which mathematical formulas, statistical models, and artificial intelligence support or influence government decisions. Several years on, the register has grown from a pilot with a handful of registrations into a database with hundreds of descriptions from national government bodies, executive agencies, provinces, and municipalities.

 When we take a closer look at the registered systems, however, a mixed picture emerges. On the one hand, the register shows steady growth in quantity and documentation standards. On the other hand, a considerable gap remains between simple administrative process rules that are extensively documented and heavy, high-risk systems that in practice are still too often missing or only superficially described. To understand how this register relates to broader government policy, it is worth reviewing [the state of AI within government](https://nieuws.llmnet.nl/en/ai-in-nederlandse-overheid-stand-van-zaken) which examines in detail the tension between the drive for innovation and transparency requirements.

 
## 1. The legal basis and the registration obligation

 The status of the Algorithm Register has undergone a notable development. Where the initiative began as an administrative agreement within the Digitalization action program, the pressure to register on a mandatory basis has been stepped up gradually. Initially, government bodies were free to decide for themselves what they defined as an 'algorithm' and which systems they published. This led to inconsistency: one organization would register a simple spreadsheet model for room bookings, while another kept a complex risk classification model for enforcement out of public view.

 With the introduction of formal frameworks such as the Impact Assessment Human Rights and Algorithms (IAMA) and the tightening of the Baseline Information Security for Government (BIO), standards have been sharpened. Government organizations must demonstrate that they have weighed the fundamental rights impact of their algorithms before a system goes into production. Registration in the central register serves as the final step of that accountability obligation. Still, enforcement remains a thorny issue: regulators such as the Autoriteit Persoonsgegevens can only take targeted action once a system processes personal data or when there are signals of discrimination, but a general sanction for failing to publish in the register was absent for a long time.

 
## 2. Typology of the registered algorithms

 Anyone browsing through the register's current datasets encounters a broad spectrum of technologies. To keep things organized, the entries can be grouped into four main categories: rule-based decision trees, statistical prediction models, computer vision, and modern language models. The lion's share of registrations still falls into the first category: deterministic rule systems that determine whether an application meets formal statutory criteria.

 
 
 
 
 Algorithm Type | 
 Example in the register | 
 Main risks | 
 Degree of transparency | 
 

 
 
 
 Rule-based (decision tree) | 
 Calculation of parking permits, fees | 
 Incorrect interpretation of the law, rigidity | 
 High: rules are directly traceable | 
 

 
 Statistical risk scores | 
 Selection of inspections for address quality, fraud | 
 Bias, proxy discrimination | 
 Moderate: parameters often partially shielded | 
 

 
 Computer vision (image analysis) | 
 Automatic scan cars, waste detection | 
 False positives, incorrect recognition | 
 Technically complex, documentation limited | 
 

 
 Generative AI & LLMs | 
 Summarizing public consultation responses, search assistants | 
 Hallucinations, data breaches, source fabrication | 
 Low: prompt architecture rarely shared | 
 

 
 
 

 Rule-based models rarely present major ethical dilemmas, because the logic follows directly from published laws and regulations. The complexity lies precisely in the statistical models and machine learning applications. In risk classification for oversight and enforcement, we see that government bodies are reluctant to disclose specific weighting factors and threshold values, out of fear that bad actors will 'game' or evade the system. This results in abstract descriptions that state that a model is running, but not which variables are decisive.

 
## 3. Municipalities versus the central national government

 A clear divide is emerging between large national organizations and local governments. Major executive bodies such as the UWV, the Sociale Verzekeringsbank, and the Belastingdienst have specialized teams for data ethics, privacy, and model validation. Their registrations are often extensive, include detailed appendices, and refer to formal DPIAs and IAMA reports. At the same time, these organizations operate in a politically sensitive context, causing publications to sometimes remain stuck in internal legal review processes for months before being made public.

 At municipalities, we see a very different pattern. Large cities such as Amsterdam, Rotterdam, and Utrecht are leading the way and developed their own standards early on. Medium-sized and smaller municipalities, by contrast, struggle with capacity and specialist knowledge. They often use standardized software packages from external vendors, leaving the municipality itself with hardly any insight into the underlying algorithms. To find out how local governments handle this tension between standard packages and their own responsibility, one can read about [the transition from pilots to formal service delivery at municipalities](https://nieuws.llmnet.nl/en/ai-bij-nederlandse-gemeenten), which explains the challenges of local practice.

 
## 4. Structural blind spots: what's missing?

 Despite the hundreds of published fact sheets, the Algorithm Register has clear blind spots. A major pain point is software acquired through commercial procurement. Many government bodies purchase off-the-shelf SaaS solutions for staff planning, fraud prevention, camera analysis, or document processing. Because vendors protect their intellectual property, they often refuse to be transparent about the training data and weighting factors used. As a result, government bodies do not report these tools at all, or only with brief boilerplate text.

 In addition, there is the rapid rise of generative AI and language models in the workplace. Civil servants increasingly use internal assistants to summarize memos, draft decisions, or analyze policy documents. When these systems run through external API gateways or cloud providers, they often escape central oversight. Anyone who wants to understand how governments and developers bundle external language models behind a central interface can turn to [the explanation of LLM API aggregators](https://api.llmnet.nl/en/aggregator-uitleg) to gain insight into the architecture of combined API flows and data buffering.

 {
 "algoritme_id": "GEM-DH-2026-041",
 "naam": "Meldingen Openbare Ruimte Routing",
 "organisatie": "Gemeente Den Haag",
 "status": "In gebruik",
 "type": "Natuurlijke taalverwerking & Classificatie",
 "doel": "Automatisch toewijzen van meldingen aan de juiste buitendienst",
 "impact_score": "Laag",
 "menselijke_tussenkomst": "Buitendienstmedewerker valideert werkorder vóór uitvoering",
 "toezichtkader": {
 "iama_uitgevoerd": true,
 "dpia_status": "Afgerond conform AVG art. 35",
 "datum_laatste_audit": "2026-03-12"
 }
}

 
## 5. The shift toward autonomous processes

 A notable trend in recent registrations is the shift from static classification models toward more dynamic, automated workflows. Where algorithms previously mainly generated advice that was then manually adopted by an employee, we now see systems that independently initiate steps in an administrative process. Think of automatically requesting additional data from third parties when an inconsistency is detected, or directly granting simple permits without human review.

 This development raises fundamental questions for how oversight is organized. When a system no longer follows a single fixed formula but translates goals into a series of successive actions, a one-time description in a register is no longer sufficient. For a thorough understanding of the risks surrounding systems that independently plan and carry out tasks, the background article on [the rise and architecture of agentic AI](https://nieuws.llmnet.nl/en/agentic-ai) offers further insight into how autonomy complicates oversight.

 
## 6. Quality and completeness of the metadata

 The usefulness of the Algorithm Register stands or falls with data quality. A quantitative analysis of the fields shows that government bodies are very consistent in filling in basic information such as the system's name, the responsible department, and the intended societal purpose. However, once we look at the deeper technical and ethical fields, completeness drops sharply.

 Many registrations contain vague descriptions in the 'How the algorithm works' field. Texts such as "the model uses advanced statistical techniques to recognize patterns" offer citizens no practical insight whatsoever into the criteria applied. The fields relating to data sources are also often limited to generic terms such as 'internal source systems', without specifying the historical periods from which training data originates or the cleaning methods used. As a result, it remains impossible for external researchers to check whether historical bias or selection effects are present.

 
## 7. The relationship with the European AI Act and oversight

 The national Algorithm Register does not function in a vacuum. Since the European AI Regulation came into force, strict, harmonized obligations apply to high-risk systems throughout the European Union. Public bodies that deploy so-called High-Risk AI Systems — for example in law enforcement, migration management, access to essential government services, or the judiciary — are required to register these systems in the central EU database.

 This raises the question of how the Dutch register relates to the European registration layer. To understand which legal obligations and timelines currently apply to governments and developers, one can consult [the dossier on the implementation of the AI Act](https://nieuws.llmnet.nl/en/ai-act-tijdlijn-wat-er-in-2026-daadwerkelijk-in-werking-is-getreden), which clearly sets out the phased enforcement and oversight tasks. The Autoriteit Persoonsgegevens, which has been designated in the Netherlands as the coordinating regulator for AI, explicitly uses the national register as an inspection tool to proactively oversee high-risk applications.

 
## 8. What can citizens actually do with the register?

 The original purpose of the register was citizen empowerment: if a resident suspects that a rejection or check is the result of an algorithm, the register should provide clarity. In current practice, that translation proves difficult. The register is primarily set up as a policy and technical catalog, not as a citizen portal that answers individual questions.

 A citizen can indeed use the register to find out that a municipality applies a risk model for, say, address checks, but cannot directly see whether their specific case was selected by that model. For that, the regular route of a right-of-access request under the GDPR remains necessary. The register does provide ammunition for asking targeted questions: because the system's name and legal basis are public, an objection can be substantiated much more precisely than before.

 
## 9. Toward a mature registration practice

 To transform the Algorithm Register from a static inventory into an effective instrument for public accountability, concrete improvements are needed. First, the definition of what must be registered needs to be enforced more strictly, so that commercial procurement and generative models no longer stay out of view. Second, automated validation of entries is desirable: registrations containing meaningless boilerplate terms should be rejected by the system until minimum substantive criteria are met.

 Transparency is not a goal in itself, but a precondition for democratic oversight. The register has taken an important first step by making automated decision-making discussable and visible. The next phase calls for deepening: not more registrations to polish the statistics, but better, verifiable documentation of the systems that actually influence citizens' daily lives.