Mauro D. Ríos (The Internet Society) has posted “Can What an AI Produces be Understood and Unraveled” on SSRN. Here is the abstract:

Over the past few decades, AI has radically transformed industries as diverse as medicine and finance, providing solutions with high levels of efficiency and accuracy that were previously unattainable (Revolutionizing healthcare, 2023)

However, the sophistication of these models, which include deep neural networks with millions of parameters and sophisticated mechanisms for producing results, has led to the perception that they operate as a “black box” whose internal logic is inaccessible to human understanding (Hyperight, 2024)

Far from being an intrinsic feature of AI, this opacity is the result of both the volume and heterogeneity of training data and the lack of adequate methodologies to record and unravel each phase of the internal calculation (Stop Explaining Black Box Models, 2022).

To overcome these myths and reveal the “why” and “how” of AI decisions, various interpretability and auditing techniques have been developed. In addition, relevance propagation methodologies, such as Layer‐Wise Relevance Propagation (LRP), make it possible to track, layer by layer, the influence of each “digital neuron” on an AI’s final decision (Montavon et al., 2019).

While these tools offer an unprecedented level of visibility, their practical application involves addressing challenges of scale and computational cost. Exhaustive logging of execution traces and parameters during training demands distributed computing infrastructures and storage systems designed for metadata versioning (Unfooling Perturbation-Based Post Hoc Explainers, 2022).

A comprehensive understanding of AI processes requires not only the use of advanced interpretability techniques, but also the establishment of governance frameworks and structured documentation. Reports from organizations such as the Centre for International Governance Innovation (CIGI) underline the need for accountability policies that require detailed records of each phase of the AI model lifecycle, from data selection to production (Explainable AI Policy, 2023). Without these mechanisms, the aspiration to full interpretability will remain limited by practical and organizational barriers, not because we can claim that we do not know why an AI does what it does, but because we have failed to implement the appropriate mechanisms to know, and thus we will be compromising transparency and trust in critical AI applications. 

Knowing what an AI does and why is then a possibility that is in our hands but requires instruments, time and resources that we must decide if they are justified in each case or we will be selective when.

Francesco Paolo Levantino (Scuola Superiore Sant’Anna di Pisa) has posted “Assessing the Risks of Emotion Recognition Technology in Domestic Security Settings: What Safeguards against the Rise of ‘Emotional Dominance’?” on SSRN. Here is the abstract:

In light of the growing interest in biometric technologies among public authorities, civil society, and international organisations, this chapter focusses on some risks associated with the use of Emotion Recognition Technology (ERT) by Law Enforcement Agencies (LEAs). In fact, despite significant attention being directed towards Facial Recognition Technology (FRT) and its uses for analogous purposes, ERT has received comparatively limited scrutiny. The chapter argues that this imbalance is reflected in the European Union’s AI Act, which inadequately addresses the potential risks associated with ERT – including its capacity to generate forms of ‘emotional dominance’. By contextualising ERT within the broader category of biometric systems, the discussion highlights the distinctive characteristics of ERT, while detailing its similarities and differences with FRT and biometric categorisation systems. The analysis shows that, beyond the intuitive perception of ERT as more intrusive due to its focus on extracting emotional states from bodily cues, its contested scientific foundations raise substantial concerns regarding its deployment in the security sector. Before some concluding remarks, the chapter draws historical parallels between the military origins of some biometric technologies and their adaptation for law enforcement, illustrating how the integration of military technologies into LEAs’ practices can affect their relationship with fundamental rights and freedoms in democratic societies. Also, it specifies how the AI Act classifies the use of ERT in law enforcement and highlights the core issues it identifies in this respect. The chapter emphasises the need for more robust regulatory frameworks to protect against the interferences posed by ERT in the context of law enforcement, asserting that existing protections under international and European human rights law should serve as a litmus test for the deployment of modern technologies by LEAs.

Jennifer Raso (McGill U Law) has posted “Interoperable AI Regulation” (Forthcoming in the Canadian Journal of Law and Technology) on SSRN. Here is the abstract:

This article explores “interoperability” as a new goal in AI regulation in Canada and beyond. Drawing on sociotechnical, computer science, and digital government literatures, it traces interoperability’s conceptual genealogy to reveal an underlying politics that prioritizes harmony over discord and consistency over plurality. This politics, the article argues, is in tension with the distinct role of statutory law (as opposed to regulation) in a democratic society. Legislation is not simply a technology through which one achieves the smooth operation of governance. Rather, legislation is better understood as a “boundary object”: an information system through which members of different communities make sense of, and communicate about, complex phenomena. This sense-making includes and even requires disagreement, the managing and resolution of which is a vital function of both law and indeed of any information system.