Research Talks

07 December 2024, 16:00 - 17:30  

Venue: Institute for Ethics in Technology, TUHH. Am Schwarzenberg Campus 3, 21073 Hamburg. Building E, 1st floor.

Wakanyi Hoffman, Research Fellow at The New Institute Hamburg, explores the transformative concept of Ubuntu and its pivotal role in shaping our approach to the climate and moral crises facing our world. In a world intricately woven with diverse narratives, Africa's rich heritage presents the profound concept of "Ubuntu" - a philosophy emphasising our interconnected humanity. Ubuntu, a term that resonates beyond mere words, is encapsulated in the African ethos as "I am because we are." This concept highlights the interconnectedness of all life and the belief that our individual and collective well-being are inextricably linked.

Popularised by Archbishop Desmond Tutu in post-apartheid South Africa, Ubuntu serves as a unifying cry across various African cultures. Its essence lies in the ethical principles of survival, solidarity, compassion, respect, dignity, and the pivotal concept of reciprocity. Reciprocity, or treating all life as we wish to be treated, is a universal principle found in numerous disciplines and indigenous worldviews.

Hoffman delves into the three levels of inner development under Ubuntu: Independence, Interdependence, and Interconnectedness. These stages guide us toward a deeper understanding of our role in the natural world and emphasise the need for a new moral framework in addressing the current climate and inner climate crises.

This talk was not just a presentation of an idea; it's an invitation to embrace a new operating system, a logic of the heart, that aligns with the shared principles of Ubuntu and reciprocity.

08 May 2024: 16:00-17:30 CEST (hybrid)
Room: E - 1.022

Abstract. Critiques of opaque machine learning models, used to guide consequential decisions, are getting traction in moral philosophy. According to the received view, the legitimacy of algorithmic decisions is threatened on the grounds that they undermine the rights of decision-subjects to informed self-advocacy (Vredenburgh, 2022). The appropriate mitigation strategy, in turn, is to grant decision-subjects a right to explanation (via explanation of the model output). This paper challenges the received view. More precisely, we have two objectives. The first is a critical one: we argue that existing accounts of the right to explanation prove unsatisfactory to ameliorate concerns about the moral illegitimacy of algorithmic decision-making. This is in particular due to their individualist framing, overburdening decision-subjects in a two-fold way: (i) the relevant explanations are likely to be epistemically over-demanding since their correct interpretation will require a combination of domain-knowledge and statistical proficiency that cannot be presumed for laypersons; and (ii) shifting the task to detect inadequacies to decision-subjects makes scrutinizing explanations very costly for them. Weakening the epistemic requirements of the right to explanation also lays the ground for our positive contribution. If providing explanations to decision-subjects turns out to be an inadequate amelioration strategy for opaque algorithmic decision-making, alternative moral guardrails are necessary. We outline the basic features of our proposal by discussing literature on model auditing in machine learning.

Thomas Grote is a research fellow at the Cluster of Excellence: “Machine Learning: New Perspectives for Science” at the University of Tübingen. He is also Co-PI in a project on certification and safety of ML models in healthcare, funded by the Carl-Zeiss Stiftung.

28 February 2024, 16:00 - 17:30 (CET) (online)

Abstract: Do Large Language Models (LLMs) have credences or degrees of belief? This question matters because a growing body of empirical research aims at quantifying LLM confidence in propositions with downstream implications for calibrating user trust in LLM assertions and combatting LLM-generated misinformation. Here an important question is whether techniques for quantifying confidence in LLMs are measuring degrees of belief on the part of the LLM; and if not, what is being measured and how it relates to credences. These questions are especially significant in relation to empirical studies which compare LLM confidence scores with human degrees of belief. In this paper we argue against the view that LLMs have credences. We consider three plausible accounts of what makes it the case that an LLM has a credence in a proposition: the reported confidence view, the output probabilities view, and the logits view. We argue that each account fails to adequately capture what it means to have a credence. The upshot is to clarify the interpretation of quantitative metrics for LLM confidence by providing a philosophical basis for denying that LLMs have credences. In doing so we not only put question to empirical comparisons between measurements of LLM confidence and reported degrees of belief in humans, but also orient discourse on confidence measurement in LLMs towards a non-mentalistic interpretation of confidence measures.

Bio: Geoff Keeling is a senior research scientist at Google, specialising in machine learning ethics. Prior to this role, Geoff served as a bioethicist at Google Health. His academic background includes a postdoctoral position at Stanford University, where he was part of the Institute for Human-Centered AI and the McCoy Family Center for Ethics in Society, and at the Leverhulme Centre for the Future of Intelligence at the University of Cambridge.