Skip to main content
Praxikon

Using special category data to detect bias: Article 4a of the AI Act explained

··15 min read

To find out whether an AI system disadvantages a group, you need to know who belongs to that group. That is exactly the kind of data the GDPR protects most strictly: ethnic origin, health, religion. Article 4a of the AI Act creates a narrow exception for it. The article was inserted by Regulation (EU) 2026/1744, the Digital Omnibus on AI, and replaces the deleted Article 10(5).

This piece explains the article itself, paragraph by paragraph, for providers and deployers in every sector. For a municipality that uses a supplier's model, the practical side is in six lessons on bias testing with special category data. The examples below are fictional.

In short

Paragraph 1 allows providers of high-risk AI systems, exceptionally, to process special categories of personal data to the extent strictly necessary to ensure bias detection and correction in accordance with Article 10(2), points (f) and (g). Paragraph 2 gives a similar option, with its own threshold, to providers of other AI systems and models and to every deployer. The same six conditions apply to all of them, in addition to the GDPR. Paragraph 2 expressly creates no obligation. In our reading the article has applied since 27 July 2026.

What the article allows, and for whom

Paragraph 1 covers providers of high-risk AI systems. They may exceptionally process special categories of personal data to the extent strictly necessary to ensure bias detection and correction in relation to their systems "in accordance with Article 10(2), points (f) and (g)". Those points concern the examination of data sets for possible biases and appropriate measures to detect, prevent and mitigate them. The processing is subject to appropriate safeguards for the fundamental rights and freedoms of natural persons.

Paragraph 2 covers three other groups: providers of other AI systems and models, deployers of those systems and models, and deployers of high-risk AI systems. The processing must be strictly necessary to ensure bias detection and correction. It concerns possible biases that are likely to affect the health and safety of persons, have a negative impact on fundamental rights or lead to discrimination prohibited under Union law, "especially where data outputs influence inputs for future operations". And all the conditions and safeguards of paragraph 1 must be applied.

What "special categories of personal data" means is set out in Article 3, point (37): the categories in Article 9(1) GDPR, in Article 10 of Directive (EU) 2016/680 and in Article 10(1) of Regulation (EU) 2018/1725. Data relating to criminal convictions and offences under Article 10 GDPR are not among them. In our reading Article 4a therefore offers no legal basis for those data.

A provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark; a deployer uses an AI system under its authority (Article 3, points (3) and (4)). An organisation that builds a system and uses it under its own name is both.

The six conditions, one by one

The conditions apply in addition to the GDPR, Regulation (EU) 2018/1725 for Union institutions and Directive (EU) 2016/680 for police and criminal justice, "as applicable". All of them must be met.

(a) Other data will not do. The bias detection and correction cannot be effectively fulfilled by processing other data, including synthetic or anonymised data.

(b) Limits and pseudonymisation. The data are subject to technical limitations on re-use, and to state-of-the-art security and privacy-preserving measures, including pseudonymisation.

(c) Strict access. There are strict controls and documentation of access. Only authorised persons with appropriate confidentiality obligations have access.

(d) No other parties. The data are not transmitted, transferred or otherwise accessed by other parties. The text does not say whether a processor within the meaning of the GDPR is an "other party". The safe reading is that no external party receives or can access the data. In our reading you can share totals per group that cannot be traced back to individuals, because they are no longer personal data.

(e) Deletion. The data are deleted once the bias has been corrected or the personal data have reached the end of their retention period, whichever comes first.

(f) Records of processing. The records of processing activities include the reasons why the processing was strictly necessary and why the objective could not be achieved by processing other data. Where special categories are processed, the exemption from keeping records for organisations with fewer than 250 employees does not apply (Article 30(5) GDPR).

What the article is not

Not an obligation. Paragraph 2 states that it "does not create any obligation to conduct such bias detection and correction", and paragraph 1 says providers "may". Providers of high-risk systems do have to examine their data sets for possible biases, but that duty comes from Article 10(2), points (f) and (g). In our reading Article 10 does not require the use of special category data for it either. A duty to be able to show that you do not discriminate may follow from the GDPR; the municipal piece covers that.

Not a blank cheque. The legal basis covers processing only to the extent strictly necessary for bias detection and correction; making the model predict better or sorting people falls outside it in our reading. And the article covers AI systems and AI models only. Fixed decision rules that are not an AI system fall outside it.

The GDPR still applies. Article 2(7) leaves the GDPR unaffected "without prejudice to Articles 4a and 59". In our reading you also need a lawful basis under Article 6(1) GDPR, the principles of Article 5 GDPR and the duty to inform still apply, and large-scale processing requires a data protection impact assessment (DPIA) (Article 35(3)(b) GDPR). Article 4a also does not say how you obtain the data; that must be lawful in its own right.

Not an exception to a prohibition. Article 4a does not lift any prohibition in Article 5. Biometric categorisation systems that categorise individual natural persons to deduce or infer, among other things, their race or religious beliefs remain prohibited (Article 5(1), point (g)). That point itself excludes the labelling or filtering of lawfully acquired biometric datasets. In our reading calling it a bias test does not make a prohibited system permitted; whether labelling an existing dataset falls under that exclusion has to be assessed case by case.

Relationship with Article 9 GDPR

Article 9(1) GDPR prohibits the processing of special categories of personal data. Point (g) of paragraph 2 makes an exception for processing necessary for reasons of substantial public interest, on the basis of Union or Member State law which is proportionate to the aim pursued, respects the essence of the right to data protection and provides for suitable and specific measures.

Recital 9 of Regulation (EU) 2026/1744 makes the link. Bias detection and correction constitute a substantial public interest, because they protect natural persons from the adverse effects of biases, including discrimination. The legal basis is subject to the same limitations, conditions and safeguards as the then existing, now deleted Article 10(5), "thereby ensuring compliance with Article 9(2), point (g) of Regulation (EU) 2016/679". In our reading Article 4a is thus the Union law that Article 9(2)(g) GDPR requires, and the six conditions are the suitable and specific measures.

The deleted Article 10(5)

The original AI Act gave this legal basis to providers of high-risk systems only, in the since deleted Article 10(5) (see also recital 70 of the AI Act). Regulation (EU) 2026/1744 deleted that paragraph (Article 1, point 9) and inserted Article 4a in Chapter I (Article 1, point 6). Recital 9 gives the reason: biases could also result from the actions of deployers of high-risk AI systems, and could arise in other AI systems or models. The six conditions are the same in substance.

Article 10 now refers to Article 4a(1) itself (Article 10(1) and (6)). If your records, DPIA or policy still refer to the deleted Article 10(5), replace that with Article 4a.

Article 59 is a separate route, available only within an AI regulatory sandbox.

Since when it applies

Regulation (EU) 2026/1744 entered into force on 27 July 2026 and sets no separate date of application. Article 4a sits in Chapter I, which under Article 113 has applied since 2 February 2025. Recital 9 says the legal basis should apply from the date of entry into application of the AI Act. In our reading an inserted article cannot apply before the day it entered into force; we work from 27 July 2026.

The requirements for high-risk systems under Annex III, including Article 10, apply from 2 December 2027 (Article 113, point (c)). A provider can therefore already use Article 4a to prepare its data sets.

One caveat. Article 2(8) provides that the AI Act does not apply to any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service; such activities are to be conducted in accordance with applicable Union law. The text does not say whether Article 4a can still serve as the legal basis in that phase. In our reading it can, because recital 9 of Regulation (EU) 2026/1744 aims to enable providers to lawfully undertake bias detection and correction in preparation for compliance with Article 10. That is not certain; record in the DPIA which legal basis you rely on during development.

Example: a public agency that builds its own system

A national benefits agency builds its own model that ranks benefit applications for an extra check, and uses it under its own name. An AI system intended to be used by or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits, or to reduce, revoke or reclaim them, is in principle high-risk (Annex III, point 5(a)), unless Article 6(3) provides an exception. In our reading this model falls under it, and because it ranks applicants on the basis of their data it performs profiling; the exception in Article 6(3) then does not apply (third subparagraph). In our reading no prohibition in Article 5 applies, as long as the model does not produce a social score as referred to in Article 5(1), point (c), and does not predict the risk of a person committing a criminal offence based solely on profiling (Article 5(1), point (d)).

The agency is provider and deployer at the same time. As provider it can rely on paragraph 1 to examine its training, validation and testing data sets. If the outcomes of checks come back as training data, that is the case Article 10(2), point (f), names, and it can rely on paragraph 1 as provider for that too. If it only measures during use whether the model selects one group more often, it can in our reading also rely on paragraph 2 as deployer. The six conditions are the same. An external consultancy cannot analyse the data on ethnic origin itself (condition (d)).

The agency also registers as provider before putting the system into service (Article 49(1)), and as a public authority that uses it (Article 49(3) and Article 26(8)). From 2 December 2027 it also performs, as deployer, the fundamental rights impact assessment of Article 27. If the system was placed on the market or put into service before that date, the AI Act applies to it only once it is subject to significant changes in its design. Providers and deployers of systems intended to be used by public authorities must comply by 2 August 2030 in any case (Article 111(2)).

Example: a bank with a credit model

An AI system intended to evaluate the creditworthiness of natural persons is in principle high-risk (Annex III, point 5(b)), except where it is used for detecting financial fraud. If the bank builds the model itself and uses it under its own name, it is provider and deployer. If it buys the model from a supplier, the supplier is provider (paragraph 1) and the bank is deployer (paragraph 2).

Suppose the bank suspects that the model rejects applicants with a chronic illness more often. To test that, it needs health data. First it establishes that a test without those data is not enough (condition (a)) and records the reasons in the records of processing (condition (f)). The supplier does not receive the health data (condition (d)), but can receive totals per group. If the bank retrains the model itself to correct the bias and that is a substantial modification (Article 3, point (23)), it becomes a provider (Article 25(1)(b)). As deployer of such a system, the bank also performs a fundamental rights impact assessment before first use from 2 December 2027 (Article 27(1)).

In brief: HR

An employer that uses a supplier's AI system to filter job applications or evaluate candidates is in principle using a high-risk system (Annex III, point 4(a)). The employer is covered as deployer by paragraph 2, the supplier as provider by paragraph 1. If the employer wants to know whether candidates of a certain ethnic origin score lower, the same six conditions apply, and the supplier does not receive the special category data. If the system infers emotions from video interviews, that falls in our reading under the prohibition of emotion inference in the workplace, except where it is intended to be put in place for medical or safety reasons (Article 5(1), point (f)).

What you record

  1. Role and paragraph. Provider, deployer or both, high-risk or not, and so paragraph 1, paragraph 2 or both.
  2. The bias. Which possible bias you examine, and why it affects health, safety, fundamental rights or the prohibition of discrimination.
  3. Why other data will not do. What you tried without special category data, and why it was not effective.
  4. The safeguards. Pseudonymisation, limits on re-use, who has access and the access log.
  5. No other parties. Where the data sit, that supplier and external consultants cannot receive or access them, and what you do share.
  6. The deletion moment. Correction or end of retention, whichever comes first, and who monitors deletion.
  7. Records and DPIA. The reasons for condition (f) in the records of processing, the GDPR lawful basis and the DPIA.

To see which duties and which role apply to your use case, try the case check.

Frequently asked questions about Article 4a of the AI Act

Sources

EUR-Lex: Regulation (EU) 2024/1689 (AI Act), original text, the later deleted Article 10(5) and recital 70 (13 June 2024, Official Journal 12 July 2024, accessed 20 November 2026)
EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 1, points 6 and 9, Article 4 and recital 9 (8 July 2026, Official Journal 24 July 2026, accessed 20 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 5, 6, 9, 10, 30 and 35 (27 April 2016, accessed 20 November 2026)

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

Does the AI Act apply to AI in cars, aircraft, trains and ships? Articles 102 to 110 and Section B of Annex I

For AI that is a safety component of a car, a tractor, a drone, marine equipment, a train or, since the Omnibus, a machine, only a handful of articles of the AI Act itself apply. The requirements for high-risk AI systems come in through the sectoral legislation: the Commission has to take them into account when it adopts delegated and implementing acts there. What Article 2(2), Section B of Annex I and Articles 102 to 110 provide, what the Omnibus changed, what still applies, and what a municipality, a leasing company, a university of applied sciences and a transport operator need to do.

EU AI ActArticle 102Article 109

Certificates and the emergency derogation: what Articles 44, 45 and 46 of the AI Act regulate

A certificate from a notified body is valid for four or five years at most and can be suspended or withdrawn. And very occasionally a high-risk AI system may be used while the conformity assessment is still under way. What Articles 44, 45 and 46 of the AI Act regulate, who will decide on such an authorisation in the Netherlands, why the GDPR simply continues to apply, and what a safety region, a municipality and a bank need to do.

EU AI ActArticle 44Article 45

Codes of conduct under Article 95 of the AI Act: voluntary rules for AI that is not high-risk

Article 95 of the AI Act lets organisations draw up codes of conduct on a voluntary basis: to apply the requirements for high-risk systems to AI that is not high-risk, and to set their own objectives on ethics, the environment, AI literacy, inclusive design and vulnerable groups. What such a code does and does not do, who may draw it up, how it differs from the codes of practice under Articles 50 and 56 and from codes of conduct under the GDPR. With a partnership of municipalities as the main example.

EU AI ActArticle 95Codes of conduct