Presentations A Abstracts

Session A


Marcin Rabiza         

Mechanistic Explanatory Strategy for Deep Learning



Kristian G. Barman 

Inference to the Best Explanation in Explainable AI



Ben Macintosh         

Skepticism about isolated explanations of Large Language Models

By practising on broad samples of language, ‘Large Language Models’ (LLMs) learn to predict text likely to follow a given linguistic input. With this objective, LLMs pick up on general reasoning abilities and the common roles or uses of words. This has recently invited various positions on whether such use of language is meaningful. However, none of this literature has appreciated Kripke’s skeptical challenge to the concept of meaning. We argue here that efforts to explain LLM behaviour inherit this challenge. Kripke’s point was that meaningful language use is not ‘blind’ or habitual but guided by reasons and justifications. LLMs also produce language in a non-trivial way - this is revealed by explanations of the model. We argue that if LLMs use language meaningfully then these explanations must play the role of guidance. Wittgenstein (in Philosophical Investigations) and Kripke's remarks, however, show that guidance cannot be captured in terms of a mechanistic isolated process. As an instance of this problem, current ‘inner’ kinds of explanation cannot capture all LLM behaviour. We should seek explanation methods that understand how LLMs belong to a public context of other language users, thereby taking on Kripke and Wittgenstein’s insights.


Thomas Raleigh & Aleks Knoks    

Opacity, explainability, and the merits of distorting idealizations



Session B


Jojanneke Drogt et al.

Aligning Artificial Intelligence (AI) with Medical Expertise: A Conceptual Understanding of Expert Practices to Foster Ethical AI Integration

Much technological progress has been achieved in the development of Artificial Intelligence (AI) systems, a broad range of computer technologies where ‘‘the system [has the ability] (…) to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation’’ [1]. Advances in AI have increased the hopes of what these kinds of systems could mean for medical practices and how these might positively complement the expertise of medical professionals. It is, for instance, argued that AI systems can function at an expert level [2, 3] serve as expert systems [4], increase the accessibility and opportunity to medical expertise [5] and might (partially) replace the expertise of medical professionals (as discussed in [6]).

Yet when examining how such systems relate to medical professionals’ expertise, it becomes clear that relatively few articles on medical AI explicitly mention what medical expertise entails, how it is practiced in medical contexts, and how understanding expertise helps to build AI systems that have a positive impact on the work of professionals. This is problematic because AI can also have a negative impact on how medical professionals perform their work. Severe problems with AI utilization in medicine, such as automation bias [7], a decrease in accuracy when AI systems are inaccurate [8], and increased risk to patients by following the recommendations of a flawed algorithm [9], indicate that AI can potentially harm medical practices when it is not carefully assessed how AI will affect how medical professionals perform their work.

We consider it, therefore, essential to investigate medical expertise in relation to computational systems such as AI, especially since scholars working on responsible and explainable AI have shown that the success of AI highly depends on the expertise of medical professionals and their willingness to accept AI outcomes in their decision-making processes [10-12]. Clinicians might, for instance, be reluctant to use AI systems because they do not have access to an explanation of why an output was produced from an input and thereby feel they cannot justify AI-supported decisions [13, 14]. In order to understand to which extent such issues can be resolved and how AI can best function in medical contexts, it is essential to further evaluate how AI systems should be aligned with medical expertise. Nevertheless, so far, a roadblock to gaining a more concrete understanding of AI’s possible impact on medical expertise has been that it is complicated by the pervasive, yet particularly elusive and diffuse nature of medical expertise. This is influenced by many normative ideas on what counts as expertise; moreover, expertise is a generic concept that involves a potentially infinite range of facets belonging to the practices of medical professionals [15].

In this article, we respond to the ambiguity surrounding medical expertise by offering a categorization of five facets of medical expertise that emerge as key elements in expert practices: (1) epistemic expertise (knowing that), (2) performative expertise (knowing how), (3) adaptive expertise (knowing when), (4) virtuous expertise (knowing why) and (5) collaborative expertise (knowing with whom). We argue that this categorization can be helpful in relation to the development of AI in the field of medicine, as it can clarify to which kind(s) of expertise AI aims to contribute and determine how desirable this is. By gaining a better understanding of how AI can impact medical expertise and specific medical practices, the article aims to improve the effective and responsible implementation of AI in medical expert practices.


1. Kaplan A, Haenlein M. Siri, Siri, in my hand: Who’s the fairest in the land? On the interpretations, illustrations, and implications of artificial intelligence. Business Horizons. 2019;62(1):15-25.

2. Bressem KK, Vahldiek JL, Adams L, Niehues SM, Haibel H, Rodriguez VR, et al. Deep learning for detection of radiographic sacroiliitis: achieving expert-level performance. Arthritis Research & Therapy. 2021;23(1):106.

3. Kowalewski K-F, Garrow CR, Schmidt MW, Benner L, Müller-Stich BP, Nickel F. Sensor-based machine learning for workflow detection and as key to detect expert level in laparoscopic suturing and knot-tying. Surgical endoscopy. 2019;33:3732-40.

4. Fieschi M. Artificial intelligence in medicine: Expert systems: Springer; 2013.

5. Currie G, Rohren E, editors. Social asymmetry, artificial intelligence and the medical imaging landscape. Seminars in Nuclear Medicine; 2022: Elsevier.

6. Ahuja AS. The impact of artificial intelligence in medicine on the future role of the physician. PeerJ. 2019;7:e7702.

7. Bond RR, Novotny T, Andrsova I, Koc L, Sisakova M, Finlay D, et al. Automation bias in medicine: The influence of automated diagnoses on interpreter accuracy and uncertainty when reading electrocardiograms. Journal of Electrocardiology. 2018;51(6, Supplement):S6-S11.

8. Gaube S, Suresh H, Raue M, Merritt A, Berkowitz SJ, Lermer E, et al. Do as AI say: susceptibility in deployment of clinical decision-aids. npj Digital Medicine. 2021;4(1):31.

9. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nature medicine. 2019;25(1):44-56.

10. Sand M, Durán JM, Jongsma KR. Responsibility beyond design: Physicians’ requirements for ethical medical AI. Bioethics. 2022;36(2):162-9.

11. Durán JM, Jongsma KR. Who is afraid of black box algorithms? On the epistemological and ethical basis of trust in medical AI. Journal of Medical Ethics. 2021;47(5):329-35.

12. Reddy S. Explainability and artificial intelligence in medicine. The Lancet Digital Health. 2022;4(4):e214-e5.

13. Holm S. On the Justified Use of AI Decision Support in Evidence-Based Medicine: Validity, Explainability, and Responsibility. Cambridge Quarterly of Healthcare Ethics. 2023:1-7.

14. Tonekaboni S, Joshi S, McCradden MD, Goldenberg A, editors. What clinicians want: contextualizing explainable machine learning for clinical end use. Machine learning for healthcare conference; 2019: PMLR.

15. Klein E. Toward a Definition of Expertise in Medicine. AMA Journal of Ethics. 2006;8(2):69-70.


Alexander Tolbert

Algorithms, Justice, and the Urban Ghetto



Cristina Voinea

Of death and griefbots: from ethical problems to political solutions



Pia-Zoe Hahne

Invisible Labour: Who Keeps the Algorithm Running?

What is work? There seems to be a collective understanding of what work should mean: “It is something that, whatever its status, is hard […], yet we have to do it. […].” (Daniels, 1987, p. 403). However, while some jobs are easily visible to the public due to certain attire, status, or characteristics, others remain invisible.

The concept of invisible labour is not new. Historically, unrecognised labour focussed on the unpaid, reproductive labour of women in the household (Daniels, 1987; Molyneux, 1979). The concept expanded to focus on all work that either goes unnoticed or is actively hidden from view, especially in connection to the work of marginalised groups such as people with disabilities, people of colour, etc. (Gilbert, 2023; Hatton, 2017; Huws, 2020). The history of computing is rife with not just invisible labour, but especially the invisible labour of women (Light, 1999). This trend takes on a new meaning in AI through data generation and the intentional hiding of human involvement (Newlands, 2021). For example, before a dataset can be used to train an AI, someone must sort through the data and classify what is depicted for the AI to be trained accordingly. In the case of AI image generation, this means placing images in subjective categories such as ‘beautiful’, ‘ugly’ or ‘aesthetic’, therefore reproducing the classifier’s own values into the training data (Crawford & Paglen, 2021; Donnarumma, 2022). This is invisible and unpaid labour.

While invisible labour connected to AI depends on the development of a new technology, it fits into previously established labour relations. For example, a worker who is tasked with sorting images into different categories without acknowledgement and under time pressure becomes alienated from the product of their work (Crawford, 2021, p. 87). In a way, this is no different from the assembly line where the worker is assigned to one specific task. The worker lacks an overview of what they are doing; in AI, the person sampling data might only have a vague idea of what the algorithm will actually be capable of in the end. What do these labour relations mean for our understanding of AI-generated images?

With the advent of artificial intelligence, invisible labour extends to large-scale operations where billions of images and text samples are needed. Not only does this raise questions regarding the legality of scraping data from the internet, but also concerning the labour of artists and internet users alike. Discussing AI-generated images in terms of AI solely being responsible for the picture obstructs the human interventions needed. The images that are used as data were made by a human, the data was sorted by humans, and the algorithm was programmed by a human. These are forms of invisible labour.

The concept of invisible labour is not new; however, from its original use in feminist discussions on the labour of women in the household, the advent of AI technology necessitates a re-evaluation of the current conceptualisation of invisible labour. I argue that while AI technologies do not result in a complete restructuring of labour relations, the ethical implications of large-scale invisible labour go further than previous conceptualisations of invisible labour. To support my claim, I present both a conceptual analysis of invisible labour, its different characteristics and the unique character and scale of invisible labour in AI as well as semi-structured qualitative interviews with AI developers, computer scientists, artists, and other relevant stakeholders. This allows a more nuanced look at the frequently abstract discussions on labour, its conceptualisations, and role in AI technology.