How the Dutch benefits scandal happened.

The short version.

From about 2013 to 2019 the benefits arm of the Dutch tax office treated tens of thousands of parents who claimed childcare allowance as suspected fraudsters. Allowances were stopped, and parents were told to repay everything they had received, often for a missing document or a small unpaid contribution. Many fell into debts they could not escape. By June 2026, 44,045 parents had been confirmed as affected.¹³

A parliamentary committee concluded in December 2020 that "fundamental principles of the rule of law were violated", by the executive, by parliament as legislator and by the administrative courts.⁷ The cabinet resigned the following month.²⁰ The data protection authority found that the way the tax office used applicants' nationality was unlawful and discriminatory, and the cabinet later acknowledged institutional racism.¹²¹ The recovery operation's budget stood at EUR 11.8 billion in September 2026.¹³

A self-learning risk model was part of it. It ranked applications so staff could check the ones it scored highest, and it used nationality as one of its signals. It did not decide who was a fraudster. Most of the financial ruin came from a rule that cancelled a whole year's entitlement for any shortcoming, from manual decisions about whole groups of parents, and from a law with no room for hardship.⁷¹² The case matters for AI because the model sat inside that system and made it faster.

Status as of 12 October 2026. Recovery figures are at 30 June 2026, the latest the government has published.

How the scheme was built.

The childcare allowance system began in 2005 and 2006. Parents were paid in advance, before any check, and the amount was settled afterwards against their actual costs and income. The data protection authority describes paying first and checking later as "a political choice".¹ The framework law had no general hardship clause, so officials had no legal room to soften a recovery that would ruin a family. One was added only in July 2020.¹¹

In April 2013 a television report exposed organised fraud in which a criminal group registered people from Bulgaria at Dutch addresses and claimed allowances in their names.⁸ The response was fast. Within days a self-learning risk model, commissioned from Deloitte, was built, tested and put into use, without a functional design or a privacy impact assessment.⁸ An anti-fraud team known as CAF was set up the same year to pursue organised abuse. The 2020 parliamentary committee describes the period as one of an "overheated" political demand for fraud enforcement, and an extra-staff business case that had to pay for itself every year through money recovered.⁷

How the risk model worked.

The model was trained on examples of correct and incorrect applications and used several dozen indicators to give every new application or change a score between 0 and 1. Each month about 100 of the highest scores went to staff, who held back the advance payment and asked the parent for documents.¹ The State Secretary told parliament in 2021 that the model "made no automatic decisions on entitlement"; staff always did.¹²

One indicator was whether the applicant held Dutch nationality, recorded as yes or no. The data protection authority documented it in the model's scorecards from March 2016 to October 2018, when it was removed for having "no predictive value".¹ Dual nationality was not an indicator, despite how the case is often reported. It was a separate breach: the tax office kept the second nationality of 1.4 million people on file without a legal basis, and fraud analysts pulled it in their own queries.¹²

Staff could not see why a case had been picked. The tax office's own description, quoted by the authority, is that "the case handler does not know why the system selects a case".¹ The 2024 inquiry found that the indicators were never tested for validity or reliability, and that without statistical correction for bias the model fed on its own selections, for example around single parents. Deloitte had warned in 2016 that the model was not retrained regularly.⁸

In one month of 2019, the people with the 1,000 highest scores were far more often non-Dutch and far more often on low incomes than everyone else.⁸ The State Secretary wrote that "it cannot be ruled out that a certain bias occurred".¹² A separate rule, "high allowance, high risk", flagged anyone receiving large amounts, which in childcare allowance means the lowest earners.⁸

The model was one part of four.

Parents came under suspicion by several routes, and the model was only one. What turned suspicion into ruin was the way the law was applied.

The anti-fraud team did not use the model to choose its cases, though it used model scores to add to the signals it received.¹¹² It worked on groups. In January 2014 it queried every applicant with Ghanaian nationality since the start of 2013 and got 6,047 names.¹ Senior officials wrote in 2015 about the "dilemma" of a group approach that assumed 80 per cent of a group was wrong and 20 per cent right. The parliamentary committee found that ratio was an assumption with no quantitative basis, and that running allowances had been stopped for groups unlawfully.⁷

The tax office also kept a fraud signals list, known as FSV, that held about 270,000 people, including anonymous tips that had never been checked. It was used across the tax office, not only for benefits, and being on it could mean intensive checks and a refused payment plan.⁵⁶ It was a database. It is often confused with the model.

How a small error became a large debt.

Under the all-or-nothing practice, any shortcoming, such as an unpaid part of the parent's own contribution to the childcare costs or a missing document, cancelled the whole year's entitlement and triggered recovery of everything paid. The parliamentary committee found that this "does not follow directly from the law". The State Attorney had called the strict reading only "arguable" in 2009; the benefits office chose it anyway, and the Council of State upheld it.⁷

A second label made it worse. Parents judged to have acted with intent or gross negligence were refused a personal payment plan, and the committee found that label was applied far too widely.⁷ Up to 16,500 people may be entitled to redress for being wrongly labelled.¹³ The Council of State reversed its own case law on 23 October 2019, holding that proportionality applies and that partial entitlement is possible.¹⁴ By then the practice had run for a decade.

The warnings.

In 2014 the anti-fraud team stopped the allowances of some 300 parents who used one childminding agency, in a case known as CAF 11.⁹ In March 2017 the Council of State ruled the stoppage unlawful.¹⁵ In August 2017 the National Ombudsman called the treatment of 232 of these families "disproportionately harsh" and recommended compensation, which was not adopted.¹⁶⁷

The wider pattern came into view two years later. In November 2019 an advisory commission led by Piet Hein Donner found "institutional bias" and a "zero tolerance" approach in the CAF 11 case, and that ordinary legal protection had not been adequate.⁹ The committee that reported in 2020 found that information to parliament had been "late, incomplete and incorrect".⁷ As in many failures of this kind, the signal arrived one case at a time, and the organisation treated each case as an exception.

What the data protection authority found.

In July 2020 the Autoriteit Persoonsgegevens reported that the benefits office had processed applicants' nationality unlawfully and in a discriminatory way: in the risk model, in the anti-fraud team's queries and by keeping dual nationality on file.¹ Its chair was careful about what the report could show: "what concrete consequences this had for individual applicants, we cannot see in this investigation".² The State Secretary apologised the same day.²⁹

In December 2021 the authority fined the Minister of Finance EUR 2.75 million. The legal basis was the GDPR's requirement that processing be lawful and fair, read with the test of necessity for a public task. Discrimination was treated as a breach of fairness.³⁴ In April 2022 it imposed a further EUR 3.7 million for the fraud signals list, at the time its highest fine.⁶ In the same period the Venice Commission, the Council of Europe's advisory body on constitutional law, described the model as "based on Artificial Intelligence" and found that "discriminatory practices were systematised through algorithms".¹¹

The political reckoning.

The parliamentary interrogation committee chaired by Chris van Dam reported on 17 December 2020 under the title Ongekend onrecht, "unprecedented injustice".⁷ It left nationality profiling and blacklists outside its scope; its subject was how the state had treated parents. Four weeks later Prime Minister Mark Rutte announced the cabinet's resignation, saying that "the rule of law must protect citizens against an all-powerful state, and here that went terribly wrong".²⁰

In May 2022 the cabinet concluded there had been "institutional racism" in the intensive supervision carried out by the tax office and the benefits office.²¹ In February 2024 a full parliamentary inquiry into fraud policy, Blind voor mens en recht, found that the powers of the state had been "blind to people and to rights" and made 19 recommendations.⁸ No one has been prosecuted. The government decided in March 2021 not to order the prosecution of ministers, and in February 2024 the Hague Court of Appeal rejected a complaint against individual officials as time-barred.¹⁸¹⁹

The record is still incomplete. In April 2026 the government told parliament that a closed tax-office file store, set up in 2019 to deal with a backlog in deleting old files under privacy law, held at least 64 million unsorted files and had not been searched when the parliamentary inquiries asked for documents.³⁶ An external report commissioned by the ministry set out how the store was built and used, and drew no conclusions about anyone's conduct.³⁷ In August the cabinet said parliament should have been told earlier and that there had been no intent. A committee is due to begin investigating in October 2026.³⁸ An organisation that cannot find its own records cannot fully account for its decisions, which is why a register of data stores, including archives set up to meet privacy rules, belongs in any response to a regulator or inquiry.

The human cost.

The recovery figures count parents and money. The hardest question is about children. Statistics Netherlands counted 2,090 children of affected parents placed out of home between 2015 and June 2022.²⁵ That is a count, and the statisticians could not say how many of those placements the scandal caused. A 2022 study at group level found no direct effect, while warning that this did not rule out individual cases.²⁴ In March 2025 a commission chaired by Mariëtte Hamer found that, for the children it examined, the scandal had played a large role.²⁶

The cabinet acknowledged the extra harm done to these children in June 2025.³⁴ In September 2026 it said it would apologise to them in January 2027, and that it would start reassessing the files of children placed out of home, at the request of parents and young people, with a limited first group.³⁵

The recovery operation.

In December 2020 the cabinet agreed to pay every affected parent at least EUR 30,000. The first 470 families were paid in January 2021.²²²³ A recovery organisation had been set up in February 2020, and a recovery act took effect in November 2022.²⁷³⁰ By 30 June 2026, 69,679 parents had registered, 44,045 had been confirmed as affected and the average award was EUR 40,664. 7,650 objections were still open.¹³

In April 2020 the operation was expected to run "at least through 2021". The budget is now EUR 11.8 billion over several years, of which EUR 5.5 billion was spent from 2020 to 2024.¹³²⁷ In 2025 the cabinet aimed to settle all additional damage claims by the end of 2027.²⁷ By June 2026, 20,314 parents had applied for compensation for additional damage and 3,973 had received it. The cabinet also plans to close access to municipal support for new applicants after 31 August 2027.¹³

In June 2026 the newspaper NRC reported that at least 20,000 parents had been wrongly recognised as affected and compensated. The ministry said no well-founded conclusion was possible without reassessing every file, and the State Secretary told parliament the account was incorrect or incomplete. Money already paid is not being reclaimed.⁴⁰²⁸ The claim remains disputed. A remedy designed to be fast and generous has become a second large administrative system, with its own disputes.

Why it matters for AI.

The model in this case did what many AI systems in government and business are now asked to do: rank people by predicted risk so that limited staff can focus on the highest scores. It was trained on past decisions, weighted dozens of signals that no one could separate, and gave reviewers a score without a reason. Every one of those features is common in AI deployments today.

The case also shows the limit of blaming the model. A ranking only becomes harm when the organisation acts on it, and here it acted through a rule that turned a query into the loss of a year's income and a label that blocked a payment plan. The same model inside a proportionate process would have produced extra checks and some wrong suspicions. Inside this process it produced debts. Governing an AI system means governing what is done with its output.

The law has moved since. The EU AI Act, in force from August 2024, lists AI systems used by public authorities to evaluate eligibility for public benefits, or to reduce, revoke or reclaim them, as high-risk.³² A system like this one would likely fall there, though no court or regulator has ruled on it. That brings a fundamental rights impact assessment before use, human oversight, registration and a right to an explanation of the system's role in a decision. After a 2026 amendment, those obligations apply from 2 December 2027.³³ The Netherlands opened a public algorithm register in December 2022; it listed 1,587 entries in October 2026.³¹ None of this would have reached the all-or-nothing rule, which sat in law and practice.

What the 2024 inquiry asked for on risk models.

Blind voor mens en recht sets out its safeguards for data and risk models in plain terms.⁸

• Use risk models sparingly.
• Selecting is giving reasons: a person selected for checks must be told why.
• Enforcement must never be fully risk-driven; keep checks that do not depend on a risk score, such as random samples.
• The greater the potential intrusion, the greater the safeguard.
• Build knowledge of the risks of data "from top civil servant to programmer".
• Test risk indicators for validity and reliability, and correct for bias.

The inquiry also recommended a general proportionality clause in administrative law, so that government action can never have disproportionately harmful effects, and a right to personal contact with the agency.⁸ That clause is not yet law. In September 2026 the cabinet said it would decide on a wider proportionality test, which it now describes as a safety valve for exceptional cases, only after new cost checks. Implementing agencies had put the cost of the full 2024 draft at EUR 300 million a year or more, against EUR 10 million a year set aside from 2030.³⁹ The Venice Commission added that when AI systems are introduced, anonymised training data should be kept so bias can be detected later.¹¹

What boards and agencies can take from it.

1. Govern the decision the score feeds. The model only ranked cases. The harm came from what followed. Before deploying any risk score, write down what action each score band triggers and whether that action is proportionate to an error.
2. Give reviewers the reason for every selection. Staff did not know why a case had been picked, so they could not challenge it. A reviewer who sees only a score will usually defer to it.
3. Test indicators before and during use. The indicators were never tested for validity or bias, and nationality stayed in for years. Review every input for what it stands in for, and drop any that act as a proxy for a protected group.
4. Keep a random sample beside the risk sample. A model that only checks the cases it selects confirms its own assumptions. Random checks give the error rate the model cannot see.
5. Monitor outcomes by group. The over-selection of non-Dutch and low-income applicants was measurable. Someone should be required to look at the distribution every month and report it upward.
6. Assess impact before go-live. The model went into use within days, with no privacy impact assessment. A fundamental rights or algorithmic impact assessment should be a condition of launch, whatever the political pressure.
7. Build in a hardship route. No official had legal room to stop a ruinous recovery. Any automated or AI-assisted process that can take money or services from people needs a named person who can override it.
8. Treat adverse rulings and complaints as a pattern. The Council of State, the Ombudsman and parents' lawyers all raised the alarm years before the system changed. Report them in aggregate to the board or the minister, with what they imply about error rates.

To check a proposed risk-scoring or AI decision system against these points, place it on the agent risk assessment matrix, and estimate the review capacity it needs with the oversight ceiling calculator.

Full timeline.

Common questions about the scandal.