The European Data Protection Board (“EDPB”) has recently published a checklist for Artificial Intelligence auditing (“AI Auditing”), representing an important step toward enhanced transparency and compliance in the use of AI technologies. The AI Auditing project aims to develop and pilot tools that help evaluate whether AI systems and applications comply with Regulation (EU) 2016/679 (“GDPR”). This project supports all stakeholders, including regulators, society, providers, and deployers of AI systems, in understanding and assessing data protection safeguards within the context of the Regulation (EU) 2024/1689, known as the artificial intelligence regulation (“AI Act”). Indeed, this EDPB tool aims to mitigate bias, enhance accountability, and ensure fairness throughout the entire AI system’s lifecycle. The outcome, documented in three detailed audit reports, ensures ongoing vigilance and adherence to legal and ethical standards, thereby fostering public trust and responsible AI deployment. But what is the purpose of AI Auditing and what aspects should be audited using the checklist? The scope of AI Auditing AI Auditing includes a methodology in the form of a checklist to audit AI algorithms based on machine learning (“ML”) to ensure legal compliance and foster accountability and transparency throughout the entire life cycle of the AI System. Specifically, AI Auditing examines a system’s actual implementation, processing activity, and operational context, focusing on the specific data used and the data subjects affected. This end-to-end and socio-technical approach recognizes that algorithmic systems interact with data generated by complex and imperfect individuals and societies and operate within intricate social and organizational contexts. The EDPB notes that this end-to-end socio-technical algorithmic audit (“E2EST/AA”) is designed to inspect algorithmic systems used in ranking, image recognition, and natural language processing, covering most systems used in both the private and public sectors. Although the primary focus is on bias assessment, the methodology also addresses broader social impact and desirability, the involvement of end-users in the design process, and the availability of recourse mechanisms for those affected by algorithmic systems. To pass such an audit, issues of impact, proportionality, participation, and resource allocation must be addressed. The AI Auditing process The E2EST/AA can be divided into four – and sometimes five – stages: Model card Model cards gather all the documentation on the AI systems – including but not limited to data protection impact assessments (DPIA) or any data processing agreement (DPA) – along with information about the training and testing of AI models and their features. This documentation is crucial for the auditor to obtain an initial picture of the system, determine which issues need further exploration, and question system developers on the origin of the information provided. System map The system map outlines the relationships and interactions between an algorithmic model (i.e., the trained algorithm which corresponds to the rules adapted to a particular domain which constitute the foundation of the technology being audited), the technical system (i.e., the entire technology), and the decision-making process (i.e., the entire lifecycle of any unit of work). Specifically, during this second stage, the EDPB suggests recording: Moments and sources of bias In AI accountability, “bias” refers to the lack of fairness and discrimination in data processes that results in individual and/or collective harms. The E2EST/AA distinguishes between moments and sources of bias, establishes the necessary documents and tests to assess compliance with legal and social requirements, provides opportunities to address and mitigate inefficiencies and harms, and offers measures for overall system fairness and impact. This means that during AI Auditing, it should be recorded the existence of: Bias testing Bias testing involves a documentation and literature review, interviews with developers/implementors, an understanding of who is impacted by AI systems (i.e., individuals, groups, society, or even the efficient functioning of an AI system itself) and how, as well as statistical analysis and checking. In some cases, bias testing requires even reaching out to end users or those affected by the systems. However, since it may not be feasible to go through all moments and sources of bias, the EDPB highlights the following main steps for measuring bias: Adversarial audit (optional) Even the most accurate audit can overlook issues such as omitted variables or proxies that become visible only when the AI system is operational. To that end, adversarial auditing can reveal additional sources of bias. This involves gathering impact data at scale, such as scraping web sources for web-based AI systems, interviewing end users, crowdsourcing end-user data, or sockpuppeting an AI system by creating fake profiles or input data with specific characteristics to trigger and analyze outcomes. The EDPB strongly recommends carrying out adversarial audits for high-risk and unsupervised ML systems. The outcome of the AI Auditing: the audit reports A crucial aspect of auditing is documentation. The EDPB lists three main audit reports: Internal E2EST/AA report with mitigation measures and annexes This internal report, not for publication, describes the audit process followed, issues identified, and mitigation measures applied or proposed. To that end, it provides solutions, monitors their implementation, and reports on the final results. Public E2EST/AA report This is the final version of the audit process. It describes the system auditing methodology, implemented mitigation and improvement measures, and any further recommendations. The public audit report must also propose periodicity and methodology/metrics for follow-up audits. Periodic E2EST/AA reports (follow-up audit reports) These reports must always refer to and provide access to the initial audit report if still relevant and ensure that the system developers keep testing for bias, implementing mitigation measures, and controlling for impact. Each periodic audit report may be produced in either an internal or public version. Conclusion Although an audit methodology does not prompt reflection on whether an AI system should exist in the first place, the EDPB’s new tool certainly represents a crucial and critical step towards ensuring that technological development aligns with the fundamental rights and freedoms of individuals. This end-to-end and socio-technical approach outlined by the EDPB provides a comprehensive framework to audit AI systems, ensuring their compliance with the principles set forth in the GDPR such as transparency, accountability, and fairness. Notably, the document frequently highlights “bias” as one of the most significant risks associated with contemporary AI technologies. To that end, the recommendation for high-risk and unsupervised ML systems shows the EDPB’s commitment to identifying and addressing even the most elusive biases. This is essential since AI systems and technologies are increasingly being used across different sectors. Moreover, the EDPB recalls the importance of ensuring compliance throughout the entire life cycle of AI systems. Indeed, the requirement for follow-up audit reports guarantees that AI systems remain compliant over time, adapting to evolving legal standards and regulations. Ultimately, the AI Auditing checklist is an important step toward responsible AI development and deployment. This is essential to foster an environment where AI can be considered responsible and suitable, thus meeting the principle of a trustworthy AI.
