International Workshop: Trust in the Age of AI

The ubiquitous integration of AI-powered systems in morally consequential decision-making procedures raises a thorny question: when such systems generate harm, who should be held responsible? A prominent regulatory response proposes that suitably designed control architectures can ensure that blame is appropriately allocated. To live up to this promise, the proposed architectures must be both normatively adequate and regarded as such, since otherwise they are at best practically useless, and at worst useless because morally problematic. In this talk, I will present some preliminary data on whether two of the most widely discussed control arrangements—loop-based oversight relations and Santoni de Sio and Van den Hoven’s track and trace framework—succeed in directing laypeople’s responsibility attributions along the pathways they prescribe, as well as whether those pathways are in fact normatively apt. 

Do labels such as “AI-generated” influence how people interpret online content? As deepfake videos and other AI-generated media spread rapidly on social media, platforms and regulators increasingly rely on transparency labels to inform users about manipulated or synthetic content. Yet such signals may do more than merely provide information: they may shape how viewers interpret the claims presented in the content itself. This is especially critical in the context of health claims, where misleading information can affect beliefs, judgments, and potentially behavior. Combining an eye-tracking experiment with an online experiment using AI-generated TikTok-style videos featuring factual and myth-based health claims, we investigate how AI-generated labels and fact-check labels guide visual attention, influence credibility judgments, and shape viewers’ interpretation of the same message. We further explore spillover effects by examining whether labeling some videos changes how users evaluate unlabeled content. Our findings aim to advance research on digital transparency, misinformation, and AI-mediated information environments by showing how labeling mechanisms shape user cognition and trust in health-related deepfake content. For practice, the study provides guidance on how to design transparency signals that inform users without unintentionally creating false reassurance or blanket skepticism.

The advancements in LLMs have opened the door to Personalized AI Assistants, which are trained on personal data to model the characteristics of a specific person and assist them with performing tasks (writing, decision-making, and communication). This innovative AI-based personalized assistance brings about unprecedented implications for personal autonomy, warranting ethical analysis and regulation. Drawing on Catriona Mackenzie’s multidimensional and relational account of personal autonomy, this paper analyses the ways in which different Personalized AI Assistants can support or undermine three distinct but intertwined dimensions of personal autonomy (self-determination, self-governance, self-authorization). By spelling out various challenges and opportunities, this analysis provides insights for designing Personalized AI Assistants to support each dimension of personal autonomy. In so doing, this paper paves the way for anticipating potential impacts and guiding the development of Personalized AI Assistants in light of personal autonomy.

AI systems are often presented as neutral tools for classifying people. Yet Automatic Gender Recognition (AGR) technologies show how fragile this assumption can be. Far from merely making technical mistakes, these systems embed problematic assumptions, namely, that gender is binary, fixed, and inferable from physical traits. As a result, they systematically misclassify gender non-conforming individuals, producing forms of algorithmic misgendering that are not only unfair but also epistemically and ethically problematic. Building on recent work on data quality dimensions for fair AI and on a feedback-based rethinking of AGR, this talk argues that trust in AGR cannot be reduced to accuracy alone. Rather, it requires a richer understanding of fairness, one that includes other quality dimensions beyond accuracy, while also recognizing the limits of automated classification itself. I will discuss why AGR systems generate ontological and epistemological errors, and propose possible strategies to mitigate these problems. Rather than offering a definitive solution, the talk aims to open a broader discussion about whether, how, and in which contexts these technologies should be used at all. In this sense, AGR serves as a case study for reflecting on what it means to trust AI in contexts involving human identity and self-determination.

Our reliance on AI systems as sources of information is steadily increasing. LLMs answer questions, explain complex topics, offer advice, and solve problems across a wide range of domains. As a result, users sometimes trust their outputs and treat them as experts or epistemic authorities. But is such trust justified? Do AI systems genuinely deserve the status of experts? While reliability is an important condition for expertise, the literature on expertise suggests that it is not sufficient. Experts are not merely reliable performers; they possess understanding of their domain. This raises a crucial question: can AI systems genuinely be said to understand? In this paper, we address this question by distinguishing between two forms of understanding. We argue that it is one thing to understand an epistemic mediator—such as a theory, model, or representational system—and another to understand phenomena or reality through the lens of such mediators. We call the former symbolic understanding and the latter noetic understanding. While symbolic understanding primarily consists in the ability to competently deploy epistemic mediators for explanation, inference, and prediction, noetic understanding requires a further element: an appropriate noetic profile, including rational commitment to the mediator as a plausible representation of reality. We argue that contemporary AI systems may exhibit some degree of symbolic understanding but lack noetic understanding. This distinction has important implications for the status of AI expertise: whether AI systems qualify as experts depends on whether expertise requires noetic understanding or whether symbolic understanding suffices.