Click the star to add/remove an item to/from your individual schedule.
You need to be logged in to avail of this functionality.
Log in
- Convenors:
-
David Moats
(University of Helsinki)
Maria Eidenskog (Department of Thematic Studies)
Dorthe Kristensen (University of Southern Denmark)
Send message to Convenors
- Format:
- Traditional Open Panel
- Location:
- HG-15A16
- Sessions:
- Tuesday 16 July, -
Time zone: Europe/Amsterdam
Short Abstract:
This panel takes an empirically grounded perspective on recent attempts to align algorithmic systems with human values. Drawing on work from Valuation Studies, we are interested in how collective or shared values are mobilized and negotiated in relation to these systems, broadly conceived.
Long Abstract:
Recent work in machine learning under the heading of ‘value alignment’ seeks to align autonomous systems with ‘human values’ (Russell 2016). Some of this happens through the mathematical formalization of values like ‘fairness’, while approaches like Inverse Reinforcement Learning (IRL) seek to extract a reward function from human preferences or behaviors. Although they are discussed and operationalised in drastically different ways, values seem central to recent discussions of algorithmic systems.
How do these understandings of values, drawing from cognitive psychology and economics, correspond to anthropological (Graeber 2001) or sociological (Skeggs and Loveday 2012) theories or indeed empirical approaches like valuation studies (Helgesson and Muniesa 2013), which see values not as a driver of action but as an upshot of practices? How can individual-level data or preferences be reconciled with more complex collective, shared values? How can we agree on what values to prioritize or how to implement them in practice?
This panel takes an empirically grounded perspective on values in algorithmic systems, broadly conceived. We will explore how (collective) values are invoked, negotiated and used to settle disputes in this context and examine attempts to invest algorithmic systems with specific values. We invite contributions, including the following:
- Ethnographic and other studies of attempts to translate values into machine learning systems and Automated Decision Making (ADM) in different domains.
- Investigations of such machine learning systems being confronted with (for example, professional) value-laden practices on the ground.
- Empirical analyses of discourses or debates around values in value alignment, AI safety or Fair ML, including divergent interpretations of concepts such as ‘fairness’ and ‘bias.’
- Accounts of disagreements between academic disciplines or professional domains over the meaning of values.
- Critical and reflexive descriptions of interventions (Zuiderent-Jerak 2015) in this space, including attempts to measure or model values computationally.
Accepted papers:
Session 1 Tuesday 16 July, 2024, -Short abstract:
This paper takes an empirical approach to the term ‘value alignment,’ which refers to attempts to align AI with human values. We analyse what understanding of values which underlies these initiatives and contrast this with understandings of values from anthropology, STS and valuation studies.
Long abstract:
The term 'value alignment,' which refers to attempts to design Artificial Intelligence systems which are aligned with ‘human values,’ is becoming more popular as leading AI companies like Open AI and Anthropic attempt to reassure the public of the safety of their models. But what do AI proponents mean by ‘values’?
This paper takes an empirical approach to the term ‘value alignment’ and asks, what understanding of values is present in these attempts to make AI safe and trustworthy? Based on an analysis of press releases, blog posts and academic papers we argue that those who use the term ‘value alignment’ are generally more concerned with existential risks in the future rather than evident harms in the present.
From these initiatives we extract an understanding of values we call the ‘secret function,’ which sees values as held individually, driving action yet mysterious to the humans who hold them. Values are seen as nothing more than (or translatable into) a ‘utility function’ or a ‘target variable’ to be optimised for.
We contrast this understanding with work from anthropology, valuation studies and pragmatist philosophy which see values as collectively formed and negotiated over time, as resources for binding groups together and policing their boundaries. What would it mean if AI was aligned this more social understanding of values?
Short abstract:
In this paper, we position value as material. Drawing on ethnographic research in a transparency project of a credit scoring agency, we argue that friction - the force when two socio-technical objects touch - occurs within the temporally situated materialities of value.
Long abstract:
Given that the making, implementation, and use of Artificial Intelligence (AI) are mostly opaque, transparency of AI is increasingly demanded by political, regulative, and private actors. These actors have different interests and ideas of how a value such as transparency can and should be aligned with proprietary AI. In this paper, we trace how this controversy was taken up by a private credit agency and addressed as a value conflict between trade secrecy and transparency. Based on our long-term ethnographic research, we follow the credit agency’s attempts to align transparency with their automated scoring system, and already existing practices of trade secrecy. Their transparency solution was to design a publicly available explainable AI tool to provide information and explanation about the scoring algorithm. Interested in AI transparency as a set of human and non-human practices, this paper pays attention to moments of friction during value alignment. We discuss friction not only occurring as a conflict between values but also in the materialities of value. In particular, friction emerges in the materialities of opaque pasts, unequal presents, and uncertain futures, and creates novel relations between transparency, trade secrecy, objects, and the positions of actors in the field. Based on our findings, we conclude by highlighting the importance of such situations where such relations are re-arranged. They reveal that the closure of value alignment and the success of technical fixes can only temporarily persist.
Short abstract:
This paper investigates the relationship between openness and secrecy in algorithmic systems by drawing on our study of Koronavilkku, an information system devised in Finland during the COVID-19 pandemic, aiming to streamline breaking of the chain of coronavirus infection.
Long abstract:
This paper investigates the relationship between openness and secrecy in algorithmic systems. These values are usually considered as contradictory: openness refers to the circulation of information, which is exactly what secrecy obstructs (Bok 1989). In the contemporary debates about emerging technologies, openness and secrecy predominantly manifest in the form of transparency and privacy. Since both transparency and privacy are posited to be essential for people’s usage of algorithmic systems, this often leads to the discussions about trade-offs. Approaching openness and secrecy in terms of gradations rather than dichotomies (Vermeir 2012), however, might cast better light on the role that the dynamics between access and restriction to information plays in the development and deployment of algorithmic systems.
To investigate such dynamics we draw on our study of Koronavilkku, an information system comprising a mobile application and an elaborated back-end system devised in Finland during the COVID-19 pandemic, aiming to streamline breaking of the chain of coronavirus infection. Through the interviews with key actors participating in the development and deployment of Koronavilkku, we discovered that a complex interplay between the access and the restriction to information is deemed central both for the system’s high uptake and its high churn rates. To understand such diverging outcomes, we ask what kinds of openness and secrecy were valued by different actors and why? What was the purpose of putting certain kinds of information in circulation and obstructing access to others? How did negotiations of information flows lead to disparate evaluations of Koronavilkku as a whole?
Short abstract:
This paper explores the performative re-purposing of in silico models into digital twins. We investigate how and why this distinction is made, who determines it, and why we should care. The aim is to demystify digital twins, make the values at stake visible and how this technology is made real.
Long abstract:
While we are still figuring out what it means to be living in a datafied world, a new cornerstone seems to emerge, this time, through so-called “medical digital twins”. Yet, what a medical digital twin means and constitutes of is still fuzzy, even for those involved in its development. Experts of this field have sought to clarify the issue by highlighting the technical differences between an in silico model and a digital twin, often associating the distinction with the maturity level or the aim of a model. However, in real life practices, in silico models and digital twins often overlap, making the boundaries between the two blurry. Hence, the technical approach overlooks the socio-political factors that influence distinction dynamics, and, importantly, when an in silico model(s) transitions into a digital twin. This incites reflections on how and why the distinction is made and by whom? What is the purpose of this (discursive) move? How are values arranged, prioritized and negotiated in this move? And, why should we care?
This paper explores the performative re-purposing of in silico models into digital twins. We study the field of in silico medicine from the inside, through an extensive ethnographic study where we examine the in silico community's discourses, the network of experts alongside specific practices working towards digital twins. This is undertaken with the objective of demystifying digital twins, making visible the values at stake and how this technology is made real.
Short abstract:
In this talk, I will systematically analyse the discourses that have taken place in the public forums of the Effective Altruism philosophical and social movement and its surrounding ecosystem (Weiss-Blatt 2023). I will also analyse its influence on the popularization of the value alignment problem.
Long abstract:
Nowadays and because of Stuart Russell (2019), the value alignment problem is mainly interpreted as a technical problem in the area of ML. That said, its more theoretical conception can be traced back to the reflections of Norbert Wiener (1960) and, in its most popular version, to the ideas presented by Nick Bostrom (2014) in his popular work “Superintelligence: Paths, Dangers, Strategies”. This approach to the issue, based on the orthogonality thesis and the idea of singularity—and thus superintelligence—has played a key role in forming the narrative of the existential risks of AI (Center for AI Safety 2023). However, presumably, its popularization in Silicon Valley circles, the AI scientist community and the media would not have been possible without the important conceptual activist role played by Effective Altruism. This philosophical and social movement, utilitarian in spirit, has even been accused of playing a key role in the failed dismissal of Sam Altman, CEO of OpenAI (Broughel 2023). In this paper, I will systematically analyse the discourses and arguments that have taken place in the public forums of this community and its surrounding ecosystem (Weiss-Blatt 2023), as well as their influence on the idea of value alignment.