The European Union’s regulations for artificial intelligence creates complex challenges for organizations deploying AI using special category, or sensitive, personal data for detecting and correcting bias.
With the EU AI Act regulating the technology, and the established General Data Protection Regulation regulating personal data, legal practitioners and business leaders face the task of ensuring dual compliance, particularly with the increase of AI processing sensitive personal data.
Regulatory challenges often hinge on the fundamental tension between algorithmic fairness—ensuring AI systems treat everyone equally free from discrimination—and protecting sensitive personal data.
To address this, Article 10(5) of the AI Act permits processing special categories of personal data when “strictly necessary” for detecting and correcting bias in high-risk AI systems. But this seems to contradict the GDPR’s general prohibition in Article 9 on processing such data without explicit consent or another specific legal basis, such as a substantial public interest.
Instead, Article 10(5) appears to create what some interpret as a new legal basis for processing sensitive data, by expanding the Article 5 fair processing principles to cover bias detection or correction.
Yet the GDPR’s Article 9 legal basis exception options contain no explicit exception for bias detection or correction. And it’s unclear whether bias detection or correction can be considered a recognised substantial public interest.
This regulatory gap creates uncertainty for organizations striving to ensure their AI systems are both fair and compliant with EU data protection and AI regulatory requirements.
Processing Pathways
While the AI Act acknowledges the GDPR’s supremacy in cases of conflict, organizations must identify the appropriate legal bases under Article 9 when processing sensitive data for bias detection and correction.
In a recent article, the European Parliament Research Service noted the interplay between these regulations, suggesting that legislators may need to intervene. But organizations will need to contend with this balancing act in the meantime.
One practical solution may involve a more nuanced approach to interpreting “substantial public interest” under Article 9(2)(g) of the GDPR. The European Data Protection Supervisor elaborated on how this article might allow processing under “substantial public interest,” with the AI Act potentially serving as the legal basis.
The European Data Protection Supervisor suggesting this pathway is a positive move, but this interpretation requires supervisory authorities to reach this conclusion and to provide reliable legal certainty.
A second option follows a Belgian supervisory authority perspective noting that correcting bias is consistent with GDPR’s fair processing principle. But fair processing isn’t a recognized legal basis under Article 9. So, as with the first option, broader consensus is required among supervisory authorities to help organizations have regulatory certainty.
Dual Compliance
In the absence of definitive regulatory guidance, organizations should consider a comprehensive approach addressing both regulatory frameworks.
Those implementing high-risk AI systems must continue to conduct thorough risk assessments that consider both AI Act and GDPR requirements. This includes identifying high-risk classification, determining if special categories data processing is needed for bias detection, conducting data protection impact assessments, and documenting decision-making processes and risk mitigation strategies.
Technical and organizational measures, important under GDPR, are vital when processing sensitive personal data through an AI platform. It’s fair to conclude that supervisory authorities—the individual data protection regulators located in each EU country who interpret and apply EU law—and the European Parliament will continue to emphasize the importance of strong cybersecurity safeguards.
Organizations should apply state-of-the-art security measures, ensure robust access controls for processing sensitive data, implement strong data minimization principles, delete special category data promptly after bias correction, and explore anonymization techniques where feasible.
When considering appropriate legal bases for processing special category personal data within an AI platform, organizations may need a hybrid approach. This involves obtaining explicit consent where feasible and exploring the “substantial public interest” exception where consent is impractical for documenting how bias detection serves important societal interests.
In parallel, it may be wise to also document that correcting bias is consistent with GDPR’s fair processing principle following the Belgian supervisory authority perspective.
None of this analysis is helpful if documentation and transparency fail to demonstrate compliance. Organizations must maintain detailed records of processing activities involving special categories data, document necessity assessments and legal basis analyses, communicate clearly and transparently about how they use data, and develop governance structures overseeing AI systems processing sensitive data.
The lack of regulatory clarity for using sensitive personal data for bias mitigation in AI is a challenge for organizations striving to comply with both the EU AI Act and GDPR. While organizations await much-needed guidance from lawmakers and supervisory authorities, they must adopt a proactive and well-documented approach to risk assessment, data minimization, and transparency.
This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law and Bloomberg Tax, or its owners.
Author Information
Cavan Fabris is the legal advisory lead partner for data protection, cybersecurity and telecoms at Deloitte Legal.
Write for Us: Author Guidelines
