Second JMCE JUST-AI Summer Colloquium

The Jean Monnet Centre of Excellente JUST AI gladly invites you to its second summer colloquium: Fairness, Accuracy, Knowability and Transparency - Transparent AI, conducive to accurate explanations of AI output, fostering fair AI ecosystems that will take place on 23-24 June 2025.

JUST AI Second SummerColloquium Program FINAL 

The dialectic between transparency and accuracy in service of fairness in the field of AI: ‘getting it right ‘in assessing risks of fundamental rights violations

In Artificial Intelligence (AI), transparency and accuracy entertain an interesting, multifaceted dialectic; one that appears to be critical in fostering and enhancing fairness. On the one hand and bearing in mind the entry into force of the AI Act (Reg. No 2024/1689 - AIA), transparency appears as a mandatory requirement for (high-risk) AI systems and an epistemic precondition to knowability. The AIA carries the imprint of ‘standard’ product safety regulatory logic, since it assumes that the observance of the safety norms listed therein will suffice to prevent risks of harm (to safety, health and fundamental rights) from materialising.

The gist in terms of transparency/accuracy is that only when a system complies with the requirements of transparency, auditability and human oversight, can providers, deployers and relevant authorities effectively access (accurate) information on the system’s data, functionalities and output (or lack thereof). Indeed, when transparency of both an AI system and the information about that system is achievable, it supports the AIA’s effective implementation by allowing for the timely detection and evaluation of AI-related risks of harm - particularly, fundamental rights violations - and enabling individuals to seek judicial redress, should their rights indeed be violated.

On the other hand, we can inquire if the normative translations of the principle of transparency (safety standards, subjective rights, namely to explanations) are effective and truly conducive to (accurate and relevant) knowledge about the risks of violating fundamental rights associated with an AI system. Two reasons justify this inquiry. First, total AI transparency is not always achievable, amounting not only to an imperfect application of transparency as a safety standard, but also to altering he effective application of connected principles like human control and oversight, as well as explainability.

Second, a closer look at the EU law’s framing of the exercise of subjective rights like the right to an explanation, uncovers a selective, rather than a generalised application of those rights. Indeed, only individuals who succeed in meeting the fairly stringent legal requirements laid down in various substantive and procedural provisions seem to be able to gain relevant knowledge about a given AI system and the risk of fundamental rights violations it raises.

It can be argued that the key normative expressions of the principle of transparency - as a safety requirement, for AI systems, and a right (to information), for individuals - cannot always be effectively put into practice, which, from the perspective of accuracy, raises one critical issue linked to the AIA’s implementation: the possibility of correctly assessing risks of fundamental rights violations.

Indeed, the AIA aims to prevent risks of harm to health, safety and fundamental rights which - as Article 6 AIA states - must be ‘significant’ that is, present a level of salience or gravity. The big question is, of course, how to assess ‘significance’ or ‘gravity’ in that connection? The AIA does not include specific metrics or a checklist of criteria to be applied in such assessments. Perhaps the technical standards drafted by the CEN and CENELEC will provide more detailed criteria (or metrics) for detecting the presence and ‘significance’ of said risks.

That said, while we wait for those standards to be made available, the AIA requires that fundamental rights impact assessments be performed both pre- and post-market, under the watchful eye of market surveillance authorities. It is therefore important to reflect on the assessment criteria that providers, deployers and relevant authorities can, or should rely on to assess fundamental rights violations so that the AIA’s effective implementation is not compromised.

Transparency and accuracy as individual entitlements

No doubt, to facilitate individuals’ access to knowledge of risks of fundamental rights violations, the (otherwise safety-standardising) AIA includes, in Article 86, the right to an explanation. Echoing Article 22 GDPR, the explanations given based on that provision should uncover the impact a high-risk AI system had on a deployer’s decision, revealing the degree of the latter’s adherence to the AI’s output. In other words, the object of Article-86 (AIA) explanations is AI-assisted human decisions, which contrasts some interpretations of Article-22 (GDPR) explanations according to which, those explanations should yield an understanding of an AI’s functionalities and underlying decisional and/or predictive logic.

Expanding on the right to an explanation in Article 86 AIA, the recently enacted Product Liability Directive (PLD, Dir. 2024/2853) and the still-under-discussion AI Liability Directive (AILD, COM(2022) 496) include a right to request disclosure of evidence. However, each instrument sets out relatively stringent conditions for the exercise of the right to such disclosure, limiting the types of information for which disclosure can be sought to information about the compliance, of the relevant actors, with the obligations laid out in the AIA.

Through an overview of the ‘pantheon of rights’ that translate the principle of transparency into entitlements to informational transparency and accuracy (i.e. right to explanation, disclosure of evidence), we cannot help but wonder if the exercise of those rights will - again - be effective and conducive to the uncovering of accurate and relevant information on the impact AI systems, or AI-assisted human decisions, had on the users’ fundamental rights.

The JUST-AI Summer Colloquium: a platform for pluri- and interdisciplinary discussion

The JUST-AI Summer Colloquium aims to bring together researchers in various stages of career advancement (doc, post-doc, Assist. and Assoc. Prof) active in IT, the social sciences (sociology, political science, law and economy) and the humanities (philosophy, history, linguistics).

About the JUST-AI Jean Monnet Center of Excellence

Launched in 2023 under the European Commission’s Jean Monnet Centre of Excellence label, the JUST-AI JMCE strives to become a Belgian and European ‘hub of excellence’ on issues pertaining to procedural fairness in the field of AI. With the participation of a network of Belgian and international experts on AI, and in view of reaching out to expert and non-expert audiences around the world, the activities launched under the Centre of Excellence label aim at promoting innovative, pluri- and interdisciplinary research, creating original and interactive educational content and raising awareness on the most salient current and future issues on procedural fairness in the era of AI. The Centre’s activities will, in particular, be developed along two research axes: 1. AI liability and fair trial requirements and 2. regulation and procedural capabilities of litigants in AI-related disputes. To learn more about the actions of the JUST-AI JMCE, you are welcome to browse through our website.

If you have questions pertaining to the Center of Excellence and/or the JUST-AI Summer Colloquium, please email us at just.ai@uliege.be.

Check out our ongoing call