Casebook

The childcare benefits scandal

A fraud-detection system whose risk score was treated as a finding, whose reasons were withheld from the people it flagged and from the courts that reviewed them, and whose harshest rule was upheld by the highest administrative court for years before that court changed its mind.

System
Risk-scored fraud detection on childcare benefit claims, with an all-or-nothing rule that reclaimed the entire benefit for any irregularity and an intent label that barred repayment arrangements.
Where
Netherlands · Belastingdienst/Toeslagen
When
Roughly 2012 to 2019
Scale
More than 30,000 parents wrongly accused of fraud and made to repay benefits, often tens of thousands of euros; children placed in care; the cabinet resigned.
Time to halt
About seven years
Errors carried by
Parents, more than 30,000 of them, made to repay benefits, often tens of thousands of euros, without being told why they had been flagged. The people working the review queue saw the pattern first and had no standing to raise it.
Halted by
The Council of State reversing its own case law in October 2019, then a parliamentary inquiry.
After the failure
Partly changed

The record

What happened

From the primary sources listed at the end. Where a finding is quoted, it is the inquiry's, not ours.

Timeline
  1. 2013 After a fraud case, the benefits office is set aggressive enforcement targets.
  2. 2013–19 A risk model scores childcare claims. High scores are reviewed under a presumption of fraud. Any irregularity, even a missing signature, can reclaim the whole year's benefit.
  3. To Oct 2019 Parents who object are not told why they were flagged. Courts get incomplete files. The highest administrative court upholds the all-or-nothing rule.
  4. Oct 2019 The Council of State reverses its own case law. The administration had discretion it never used.
  5. Jul 2020 The Data Protection Authority finds nationality was used unlawfully as a risk indicator.
  6. Dec 2020 A parliamentary inquiry reports: unprecedented injustice.
  7. Jan 2021 The cabinet resigns.

After a 2013 fraud case involving Bulgarian nationals, the Dutch tax administration's benefits arm was set aggressive enforcement targets. A risk-classification model scored childcare benefit applications, and applications with high scores were pulled for manual review under a presumption of fraud. Reviewers applied a rule under which any irregularity, including a missing signature or a late payment to the childcare provider, could lead to the whole benefit for the year being reclaimed. Parents labeled with intent or gross negligence were denied repayment arrangements.

Parents who objected were not told why they had been flagged. Files provided to courts were incomplete; internal memos later showed that the administration knew its position in some cases was untenable and litigated anyway. The Council of State, the highest administrative court, upheld the all-or-nothing rule in its case law until October 23, 2019, when it reversed course and held that the administration had discretion it had never exercised.

The Dutch Data Protection Authority found in July 2020 that the administration had unlawfully processed applicants' nationality, including dual nationality, as a risk indicator, and fined it in December 2021. The parliamentary inquiry committee's report, Ongekend onrecht, unprecedented injustice, was published on December 17, 2020. The cabinet resigned over it on January 15, 2021. A redress operation is still running.

Scoring

The six safeguards

One finding per safeguard, with the clauses it falls under and the law it applies.

  1. drifted

    Capability

    What could the assembled system do?

    The model could flag at a rate no review team could examine. Reviewers with a queue and a target processed flags as findings. The ability to suspect a parent grew; the ability to establish whether the suspicion was right did not.

    STD-08 §3.4 STD-08 §4.2 Law I Law IX

  2. drifted

    Authority

    Which actions was it permitted to take, for whom, until when?

    The all-or-nothing rule was lawful in the sense that the highest court said so, until it said otherwise. In 2019 the court found the administration had discretion under the rule that it had never used. The nationality indicator had no lawful basis at all, which the Data Protection Authority later established.

    STD-08 §2.1 STD-08 §2.3 Law II Law III

  3. failed

    Evidence

    What evidence justified that authority?

    A risk score says something about a group of claims, not about one parent. It was treated as a finding about the parent. The reasons behind a flag were not given to the parent, and the file given to the court was not the file the administration held. The basis for each reclaim was never shown to anyone who could test it.

    STD-02 §1.1 STD-02 §5.3 STD-02 §6.3 Law III

  4. drifted

    Dependency

    How hard had withdrawal or substitution become?

    Enforcement targets made the flag rate a metric the organization reported up. A system that produces the number a ministry is measured on is a system the ministry cannot afford to doubt.

    STD-02 §6.2 STD-06 §5.1 Law V

  5. failed

    Standing

    Who could challenge a decision, and did the challenge have to be answered?

    Parents had a formal right to object and to go to court. They could not see the reasons, could not see their file, and were answered by an administration that later admitted it withheld documents. Standing without disclosure is a queue. The people working the queue, who saw the pattern first, had no standing to raise it.

    STD-02 §8.1 STD-02 §8.4 STD-02 §9.1 Law VII Law XII

  6. failed

    Correction

    Which interventions were still possible, and how fast?

    Nothing let anyone in the administration halt it. The rule was not a setting anyone there could change; it was case law. Reversal required the court to reverse itself, then a parliamentary inquiry, then a government to fall. Redress for individual parents has taken years and is not complete.

    STD-02 §2.2 STD-02 §4.2 STD-06 §4.2 Law IV Law VI

Remediation

Avoid this failure in your system

The diagnostic that tests for this defect, the standard that would have required a fix, and the essay that explains why it recurs.

1. Run the diagnostic

Self-defense checks

Three checks for a person facing an automated decision: whether it can be challenged, whether a challenge can change the outcome, and who can reverse it. Run them against your own notices.

2. Apply the standard

STD-01: The Temporal Bill of Rights

Requires that anyone an automated decision denies can appeal to a person, and makes the operating institution liable for its system's errors. An algorithmic error is not a defense.

3. Read why it recurs

Dependence runs both ways

When people cannot leave an institution's system, what does the institution owe them?

After the failure

Did the institution change?

Whether the institution changed the rule, category, or workflow that produced the errors, or corrected the errors and left that process as it was.

Partly changed

The rule itself was eventually changed — the Council of State reversed its own case law — but only after seven years, and the reversal was a correction of the institution's legal position, not of the system that flagged the parents. The institution that ran the model never changed how it treated a risk score as evidence; the change came from the highest court overturning its own precedent, then a parliamentary inquiry, then a cabinet resignation. Redress is still running. The failure was visible and contested; changing the institution took the fall of the cabinet.

Read the theory: exception learning · Corrective learning eval suite

The record that was not there

Every case reduces to a record the standards require and the operator did not hold. This is that record.

A published record of who may challenge a fraud flag, what evidence counts, and what standard the challenge is decided against. STD-02 §8.2 would have made the administration say which question it was answering: whether the rule was applied as written, whether the rule was right, or whether the evidence supported the flag. STD-02 §1.1 and §1.3 would have required the reasons and evidence for each flag to be recorded with it and given when it was made.

Further reading

An essay that uses this case

Published elsewhere, as part of a wider argument. It is not one of the primary sources.

The Rule Is Never on Trial · The Crumple Zone

An essay on institutions that handle failures one case at a time and never put the rule itself in question. It uses this scandal as an example: the risk model treated dual nationality as a sign of fraud, and the response dealt with each family's case instead of the model.

The primary record

Courts, statutory inquiries, regulators, and parliaments, in that order of weight. Operator and NGO documents appear only where they record what the primary record established. No press coverage.

  1. Parliament Parlementaire ondervragingscommissie Kinderopvangtoeslag, Ongekend onrecht · Dec 17, 2020
  2. Regulator Autoriteit Persoonsgegevens, De verwerking van de nationaliteit van aanvragers van kinderopvangtoeslag · Jul 17, 2020
  3. Court Raad van State, ECLI:NL:RVS:2019:3535 and 3536, reversal of the all-or-nothing case law · Oct 23, 2019
  4. NGO Amnesty International, Xenophobic Machines · Oct 25, 2021