Erster Auftrag und der Grund für eine GmbH: Mein Start in die Selbstständigkeit | Umair Zaffar
Umair Zaffar November 14, 2026 0 Comments

EU AI Act Article 4a: Sensitive Data for Bias Testing, Implemented in Practice

EU AI Act Article 4a exceptionally allows special categories of personal data to be processed in order to detect and correct bias in AI systems, under six conditions. Our answer: a sealed bias-testing environment with a red-zone store, purpose binding, a pseudonymised join, an access log and deletion after the test, from which only the result record travels into the registry.

Article 4a resolves a paradox that every fairness project in a bank or an insurer knows. Art. 10(2)(f) requires high-risk systems to be examined for biases that may lead to prohibited discrimination. To measure whether a model disadvantages people by ethnic origin, health status or religion, you need exactly the data that Art. 9 GDPR prohibits by default. It addresses providers of high-risk systems (paragraph 1) and providers and deployers of other AI systems as well as deployers of high-risk systems (paragraph 2). Data governance itself is covered in the post on Article 10; this post is only about the special case of sensitive data.

The EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy duties have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 4a sits in Chapter I, was newly inserted by the Omnibus and replaces the former Art. 10(5); according to the AI Act Explorer it has applied since 27 July 2026, so before the high-risk obligations its paragraph 1 refers to. I describe the overall frame in EU AI Act in practice.

What EU AI Act Article 4a requires

Strictly speaking, the article requires nothing; it permits. Article 4a is an opening clause against the prohibition in Art. 9 GDPR, and every opening hangs on conditions that must all be met at the same time:

  • Art. 4a(1), the basic rule: Providers of high-risk systems may exceptionally process special categories to the extent strictly necessary for bias detection and correction under Art. 10(2)(f) and (g). The GDPR, Regulation (EU) 2018/1725 and Directive (EU) 2016/680 continue to apply alongside.
  • Point (a), necessity: The purpose cannot be effectively achieved with other data, including synthetic or anonymised data.
  • Point (b), technical limits: Technical limitations on re-use and state-of-the-art security and privacy-preserving measures, pseudonymisation included.
  • Point (c), access: Secured data, strictly controlled and documented access, only authorised persons under confidentiality obligations.
  • Point (d), no transfer: No transmission, no transfer, no access by other parties.
  • Point (e), deletion: Once the bias has been corrected or the retention period ends, whichever comes first.
  • Point (f), the record: The record of processing activities states why the processing was strictly necessary and why other data would not have sufficed.
  • Art. 4a(2), the extension: Providers and deployers of other AI systems and models, and deployers of high-risk systems, may do the same where (a) the processing is strictly necessary because of biases that affect health, safety, fundamental rights or a prohibition of discrimination, especially where outputs influence future inputs, and (b) all conditions of paragraph 1 are applied. The last sentence: paragraph 2 creates no obligation to conduct bias detection.

The duty to examine itself sits, for providers, in Art. 10(2)(f) and (g); Article 4a only supplies the data route. For deployers the pressure comes from Art. 26(4) and, for banks and life and health insurers under Annex III point 5, from the fundamental rights impact assessment under Art. 27.

The technical example: claims triage and credit pre-check

Two systems, one problem. An insurer runs a model that pre-sorts claims: automatic settlement, case handling or in-depth review. A bank runs a credit pre-check that sorts applications into traffic-light colours before the advisor meeting. Neither decides finally, but both decide who gets looked at critically. Both have been proxy-tested on first name, postcode and language. What those tests cannot answer is the board’s question: do people of a certain origin, with a chronic illness or of a certain religion end up in the critical lane systematically more often?

To answer it you need, per case, the model decision and the protected attribute side by side. The insurer lawfully holds health data, but for benefit assessment, not for model testing. The bank holds health data from payment protection insurance. Religion is usually held by neither. The dangerous moment is the join: decisions sit pseudonymised in the evidence log, attributes with real names in business systems, and merged in a notebook they become a table nobody in the company is allowed to have. This is where Article 4a bites, with all six conditions, and the addition in Art. 4a(2)(a) about outputs that influence future inputs fits both systems: today’s triage decision is tomorrow’s training label.

How we implement it

At SIFAMO we treat the bias test as its own registered process with a start, an end and a deletion date, not as an analysis on the side of the project. The building blocks come from our governance layer. Six steps, none is skipped.

1. Clarify the legal basis and the role

Before the first access it is settled whether the company is a provider under paragraph 1 or a deployer under paragraph 2. The data protection officer and the business owner check whether the attributes are lawfully held, for which original purpose, and whether a data protection impact assessment is needed; for special categories it usually is. We collect no new attributes for tests. Where one is missing, the result record says openly that this group could not be measured. I say this deliberately: Article 4a permits measuring, not collecting.

2. The necessity test, documented

Point (a) demands that there is no other way, point (f) demands the justification. We run the proof as an experiment, not as boilerplate. The proxy test shows whether the model reacts to signals correlated with origin, but not how strongly which group is affected. Synthetic cases probe constructed variations, not the real distribution of errors across real groups, and inherit the assumptions of their generator. For health status the attribute also sits in the free text the model reads; whether it learns from it can only be seen on real cases. The result of these preliminary stages, with numbers, is the necessity record for the record of processing activities. If the proxy test suffices, Article 4a is not invoked.

3. Red-zone store, purpose binding, pseudonymised join

The attributes are extracted from the business systems into a dedicated red-zone store that belongs to exactly one bias test. The policy gateway knows it under a purpose ID, only the registered test process can query it, and no training run and no feature store has a route to it. That is our reading of point (b): not a policy, but a missing route. The deletion date is set at creation. The model decisions come from the evidence log with pseudonymous ID, model version and output; the attributes receive the same stable IDs from the pseudonymisation service, inside the test environment, so the mapping table never leaves it. The environment has no network egress and no export except an output gate for aggregates above a minimum group size set by the business owner. What comes out is a table of group rates, not a row with a real name.

4. No transfer, not even to us

Point (d) is where vendor models fail. The model provider does not get the attributes, the cloud provider does not, and neither do we as the service provider. We build the environment, define the metrics and review the record; the test is run by named employees of the company. A bias test that the vendor runs with the client’s sensitive data is, to me, not a borderline case but a breach.

5. Every access in the log, deletion after the test

The test environment writes into the same append-only pattern as the evidence log: who, when, which query, how many rows, which result exported. The data protection officer reads the log before approval; we turn point (c) into a condition for the signature. After correction and a passed repeat test, the environment is destroyed together with the red-zone store, and the deletion record is referenced in the registry. Point (e) names two triggers, correction or the end of the retention period; the earlier one applies. The next test, say after retraining, is a new process with a new necessity record. That is effort, and it is intended.

6. The result record goes into the registry, the data does not

What survives the test is a record without personal data. It hangs off the model’s registry entry in the bias-check field and is part of the quality gate before deployment. Simplified:

{
  "bias_test_id": "BT-2026-017",
  "model_id": "DE-V-2026-0088 (claims triage), version 2026-10-r1",
  "legal_route": "Art. 4a(1) (provider)",
  "attributes": ["health status (benefit files)", "religion: not held, not measured"],
  "necessity_record": "NR-2026-017 (proxy and synthetic test: insufficient)",
  "population": "cases 2026-04 to 2026-09, pseudonymised, minimum group size 200",
  "metrics": {"review_rate": "per group", "false_review_rate": "per group", "max_gap": "threshold set by business owner"},
  "finding": "threshold exceeded in one group, cause: free-text feature from expert reports",
  "remediation": "feature removed, retraining, repeat test BT-2026-019 passed",
  "access_log": "47 accesses, 3 authorised persons, reviewed by the DPO",
  "deletion": "environment destroyed 2026-10-21, deletion record LP-2026-017",
  "sign_off": ["business owner", "risk and data protection", "IT"]
}

An auditor sees the necessity record, the entry in the record of processing activities, the access log, the deletion record and this result with three signatures. The auditor does not see the sensitive data, and does not need it. Further posts in the series: all articles at a glance.

Typical mistakes

  • The attribute ends up in the feature store. Health status is attached to the training table for the test, and a year later it is a model feature.
  • The necessity record is boilerplate. “Synthetic data was insufficient” without numbers convinces no supervisor.
  • The vendor tests with the client’s data. Convenient, and a clear breach of point (d).
  • The notebook outlives the test. CSV exports on shared drives are the most common reason point (e) is violated.

Article 4a checklist

  • Role clarified: provider under paragraph 1 or deployer under paragraph 2.
  • Necessity record with results of proxy and synthetic tests, in the record of processing activities before the first access.
  • Dedicated red-zone store per test process, with a purpose ID in the gateway and no route to training or the feature store.
  • Join only via pseudonymous IDs in an environment without network egress; export only as aggregates above a minimum group size.
  • Named access list, append-only access log, no transfer to vendors, cloud providers or consultants.
  • Deletion date set at creation, deletion record referenced in the registry.
  • Result record with group metrics, finding and remediation on the registry entry, without personal data, with three signatures.

Open questions

  • How do you prove that synthetic or anonymised data would not suffice? Point (a) makes this the first condition, and an accepted method is missing. Our experiment is practice, not a standard.
  • Does Article 4a also permit collecting new attributes? The text speaks of processing, which in data protection law includes collection. We read it narrowly.
  • When is a bias “corrected”? Point (e) ties deletion to it, but models get retrained. Whether supervisors will accept separately registered processes or expect continuous processing is open.

Frequently asked questions

Does Article 4a also apply to us as a deployer?

Yes, via paragraph 2. Whoever deploys a purchased high-risk system, or provides or deploys another AI system, may process special categories for bias tests under the same six conditions. No obligation arises from it; for banks and insurers under Annex III, Art. 27 nevertheless supplies a good reason.

Do we have to run bias tests with sensitive data?

No. Providers of high-risk systems must examine biases under Art. 10(2)(f), but not necessarily with special categories; the route via Article 4a is open only where there is no other way. For everyone else, paragraph 2 says expressly that no obligation arises.

May we keep the data for the next test?

No. Point (e) requires deletion once the bias has been corrected or the retention period ends, whichever comes first. The next test is a new process; what remains is the result record in the registry.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment