This paper explores the development of an algorithm for child welfare administration in Denmark. Based on an ethnographic study of the development process, we argue that scientists and IT developers enacted not one but multiple versions of "the" algorithm, and in this process, also engineered its multiple potential worlds. We conceptualize these as "algorithm-worlds"; specific sets of relations in which a version of the algorithm can exist and act. We illustrate three examples from our study: the algorithm as a docile tool within the world of child welfare casework; the algorithm as a data-connector within the world of public data infrastructure; and the algorithm as a data processor in the world of a legal assessment. We contend the merely foregrounding algorithms' multiplicity risks underestimating their world-making power, whereas construing algorithms as powerful without attending to their processes of becoming renders them seemingly singular and universal. By combining recent work on the multiplicity of algorithms in STS with actor-network-theory studies on "heterogeneous engineering," our approach, in turn, allows elucidating not only three different versions but also how each of these came with a distinct world of practice. We argue that these algorithm-worlds partially coexist, partially conflict or cascade.
In this paper, we demonstrate how a new form of ethics work emerges in the area where social work and artificial intelligence (AI) technologies converge. The paper reports on an organisational ethnography of a Scandinavian NGO, specifically comprising the efforts of social workers and data engineers to establish a fair AI Counselling Assistant (AICA) for supporting volunteer staff in their online communications with children seeking help and support. The purpose of the AICA is to retrieve relevant information and advice for the volunteer social workers' conversations with children written in real time. Drawing on science and technology studies, we analyse ethics work related to the AICA as a more-than-human endeavour. We highlight four ethical dimensions related to (1) distance, (2) agency, (3) time and (4) errors, which characterise what we term socio-technical work which continuously questions and addresses the ethicality of the AICA. We conclude that ethics work in the area where social work and AI converge requires social workers to possess a technological awareness that enables them to engage with both data ethics and the situated ethics of social work.
Multi-agent systems of large language models (LLMs) are rapidly expanding across domains, introducing dynamics not captured by single-agent evaluations. Yet, existing work has mostly contrasted the behavior of a single agent with that of a collective of fixed size, leaving open a central question: how does group size shape dynamics? Here, we move beyond this dichotomy and systematically explore outcomes across the full range of group sizes. We focus on multi-agent misalignment, building on recent evidence that interacting LLMs playing a simple coordination game can generate collective biases absent in individual models. First, we show that collective bias is a deeper phenomenon than previously assessed: interaction can amplify individual biases, introduce new ones, or override model-level preferences. Second, we demonstrate that group size affects the dynamics in a non-linear way, revealing model-dependent dynamical regimes. Finally, we develop a mean-field analytical approach and show that, above a critical population size, simulations converge to deterministic predictions that expose the basins of attraction of competing equilibria. These findings establish group size as a key driver of multi-agent dynamics and highlight the need to consider population-level effects when deploying LLM-based systems at scale.
Generative Artificial Intelligence (AI) tools are increasingly deployed across social media platforms, yet their implications for user behavior and experience remain understudied, particularly regarding two critical dimensions: (1) how AI tools affect the behaviors of content producers in a social media context, and (2) how content generated with AI assistance is perceived by users. To fill this gap, we conduct a controlled experiment with a representative sample of 680 U.S. participants in a realistic social media environment. The participants are randomly assigned to small discussion groups, each consisting of five individuals in one of five distinct experimental conditions: a control group and four treatment groups, each employing a unique AI intervention-chat assistance, conversation starters, feedback on comment drafts, and reply suggestions. Our findings highlight a complex duality: some AI-tools increase user engagement and volume of generated content, but at the same time decrease the perceived quality and authenticity of discussion, and introduce a negative spill-over effect on conversations. Based on our findings, we propose four design principles and recommendations aimed at social media platforms, policymakers, and stakeholders: ensuring transparent disclosure of AI-generated content, designing tools with user-focused personalization, incorporating context-sensitivity to account for both topic and user intent, and prioritizing intuitive user interfaces. These principles aim to guide an ethical and effective integration of generative AI into social media.
We study a generalization of the classic Global Min-Cut problem, called Global Label Min-Cut (or sometimes Global Hedge Min-Cut): the edges of the input (multi)graph are labeled (or partitioned into color classes or hedges), and removing all edges of the same label (color or from the same hedge) costs one. The problem asks to disconnect the graph at minimum cost. While the $st$-cut version of the problem is known to be NP-hard, the above global cut version is known to admit a quasi-polynomial randomized $n^{O(\log \mathrm{OPT})}$-time algorithm due to Ghaffari, Karger, and Panigrahi [SODA 2017]. They consider this as ``strong evidence that this problem is in P''. We show that this is actually not the case. We complete the study of the complexity of the Global Label Min-Cut problem by showing that the quasi-polynomial running time is probably optimal: We show that the existence of an algorithm with running time $(np)^{o(\log n/ (\log \log n)^2)}$ would contradict the Exponential Time Hypothesis, where $n$ is the number of vertices, and $p$ is the number of labels in the input. The key step for the lower bound is a proof that Global Label Min-Cut is W[1]-hard when parameterized by the number of uncut labels. In other words, the problem is difficult in the regime where almost all labels need to be cut to disconnect the graph. To turn this lower bound into a quasi-polynomial-time lower bound, we also needed to revisit the framework due to Marx [Theory Comput. 2010] of proving lower bounds assuming Exponential Time Hypothesis through the Subgraph Isomorphism problem parameterized by the number of edges of the pattern. Here, we provide an alternative simplified proof of the hardness of this problem that is more versatile with respect to the choice of the regimes of the parameters.
