How Do Communities of ML-Enabled Systems Smell? A Cross-Sectional Study on the Prevalence of Community Smells

Giusy Annunziata, Stefano Lambiase, Fabio Palomba, Gemma Catolino, Filomena Ferrucci · 2025

Read How Do Communities of ML-Enabled Systems Smell? A Cross-Sectional Study on the Prevalence of Community Smells by Giusy Annunziata, Stefano Lambiase, Fabio Palomba, Gemma Catolino, Filomena Ferrucci (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-17419.txt.

Introduction

Giusy Annunziata and four co-authors measured how common team dysfunction is in open source machine learning projects, in a 2025 paper. Software people call these patterns community smells: recognizable social problems, like a prima donna who will not share control, that corrode collaboration and pile up technical debt. ML teams are interdisciplinary, yet their social dynamics are barely studied. The team ran a detection tool over 188 repositories.

Certain smells, notably Prima Donna Effects and Sharing Villainy, appeared often and waxed and waned over a project's life. Others, like Radio Silence, were rarer. Smells also traveled together in patterns the analysis mapped. The authors suggest project managers watch these signals the way they watch code metrics. Fixing the social faults early, they argue, is cheaper than untangling the technical debt those faults leave behind.

Catalog record

CREATORGiusy Annunziata, Stefano Lambiase, Fabio Palomba, Gemma Catolino, Filomena Ferrucci
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-17419

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

How Do Communities of ML-Enabled Systems Smell? A Cross-Sectional Study on the Prevalence of Community Smells Giusy Annunziata, Stefano Lambiase, Fabio Palomba, Gemma Catolino, Filomena Ferrucci Abstract . Effective software development relies on managing both collaboration and technology, but socio-technical challenges can harm team dynamics and increase technical debt. Although teams working on ML-enabled systems are interdisciplinary, research has largely focused on technical issues, leaving their socio-technical dynamics underexplored. This study aims to address this gap by examining the prevalence, evolution, and interrelations of “community smells”, in open-source ML projects. We conducted an empirical study on 188 repositories from the NICHE dataset using the CADOCS tool to identify and analyze community smells. Our analysis focused on their prevalence, interrelations, and temporal variations. We found that certain smells—such as Prima Donna Effects and Sharing Villainy—are more prevalent and fluctuate over time compared to others like Radio Silence or Organizational Skirmish. These insights might provide valuable support for ML project managers in addressing socio-technical issues and improving team coordination. DOI: XXXXXXX.XXXXXXX Conference: The 29th International Conference on Evaluation and Assessment in Software Engineering; 17–20 June, 2025; Istanbul, Türkiye ISBN: 978-1-4503-XXXX-X/18/06 CCS: Software and its engineering Open source model 1. Introduction The success of a software development project depends not only on the technical aspects of the product but also on effective collaboration and communication within the development team ( Tamburri et al., 2015 ; Martini and Bosch, 2017 ) . Social dynamics can directly impact technical processes, influencing both productivity and software quality ( Palomba et al., 2021 ) . However, such social challenges, being abstract factors rooted in communication and collaboration, are inherently complex to detect and quantify ( Tamburri et al., 2016 ) . The literature has investigated “ Community Smells ”, which are measurable patterns of organizational problems that hinder cooperation and affect project health ( Tamburri et al., 2021 ) . These smells often lead to increased social debt, where unresolved social issues accumulate over time, degrading software quality and team cohesion ( Palomba et al., 2021 ; Catolino et al., 2021 ) . Machine Learning (ML)-enabled systems introduce additional complexity when analyzing social dynamics. By their nature, these systems rely on multidisciplinary teams composed of developers, software engineers, and data scientists, each bringing distinct educational background, perspectives, and expertise ( Ozkaya, 2020 ) . While this diversity can foster innovation and improve problem-solving, it can also create social tensions, misalignments, and communication gaps that may exacerbate the emergence of community smells ( Busquim et al., 2024 ) . Despite the critical role of social aspects in the shaping of ML-enabled systems, they remain largely underexplored compared to technical aspects. Most research to date has primarily focused on issues such as technical debt, suboptimal design decisions, or neglected maintenance tasks ( Foidl et al., 2022 ; Recupito et al., 2024 ) , often leaving the social dimension insufficiently explored ( Mailach and Siegmund, 2023 ) . An initial investigation in this direction was conducted by Mailach and Siegmund ( Mailach and Siegmund, 2023 ) , who explored socio-technical anti-patterns identified by the communities of ML-enabled systems, analyzing their causes and potential mitigation strategies. Building on this, Annunziata et al. ( Annunziata et al., 2025b ) examined the correlation between these anti-patterns and known community smells, providing the first insight into community smells within communities of ML-enabled systems. Their findings highlight that the interaction between diverse teams and the complex workflows characterizing those projects can exacerbate community smells and social issues, leading to increased social debt and weakening collaboration and productivity. Therefore, analyzing these socio-technical issues is critical to developing strategies that improve team cohesion, mitigate social tensions, and improve the overall health and results of software projects. Our study aims to deepen the understanding and the examination of the socio-technical well-being of communities developing ML-enabled systems, providing a first investigation of the prevalence and evolution of community smells in real open-source ML-enabled systems. We conducted our study on 188 GitHub ML-enabled open-source projects from the NICHE dataset ( Widyasari et al., 2023 ) , identifying and detecting 10 distinct community smells, e.g., Prima Donna Effect and Toxic Communication show the adoption of unhealthy and rude communication, superiority, and uncooperative behavior, using the CADOCS tool ( Voria et al., 2022 ) . Next, we performed a Cross-Sectional Study to examine the prevalence of each community smell in the analyzed projects and explore correlations between them. Finally, we analyzed the evolution of community smells over time through a Longitudinal Study, examining each project in 3-month intervals from its inception to 2 years of age. The results obtained show that Prima Donna Effect (PDE)—which refers to a superiority, disagreement, and uncooperative behavior—is the most prevalent smell, whit a prevalence of 92.6%. Furthermore, the longitudinal analysis reveals that PDE remains consistently high over time (95-96%). In contrast, Radio Silence (RS)—where formal communication barriers between sub-communities cause time loss —is the least frequent, observed in 18,6% of the cases. Meanwhile, other community smells tend to fluctuate over time, for example, Unhealthy Interaction (UI)—the adoption of unhealthy behavior among team members—decreased by 24% to 6%, suggesting improvements in communication practices. Finally, the results suggest a strong correlation between PDE and other community smells, in particular the Organizational Silo Effect (OSE)—which describes isolated communities that communicate with each other only through a few members. Our findings offer valuable insights for project managers in ML-enabled projects, equipping them with a deeper understanding of the most common social challenges and practical strategies to address them. Additionally, these results serve as a strong foundation for researchers, providing an initial exploration and quantitative analysis of community smells in the ML-enabled context. By highlighting their evolution during the early stages of software projects, this work sets the stage for future investigations into the interplay between human dynamics and their impact on the technical success of ML-enabled projects. Structure of the paper. Section 2 presents background and related work on ML-enabled systems and community smells. Section 3 outlines the design of the study and the methodology for each research question. Section 4 reports the results of the study. Section 5 discusses the insights of the paper and practical implications, while Section 6 the threats to validity. Finally, Section 7 concludes the paper and outlines future investigations. 2. Background and Related Work In this section, we discuss the background and related work that underpin our study. 2.1. ML-Enabled Systems Machine Learning-enabled systems (ML-Enabled Systems) are a specialized category of software systems that integrate machine learning models to automate and enhance decision-making processes ( Villamizar et al., 2022 ; Martínez-Fernández et al., 2022 ) . These systems are increasingly deployed across various domains, including autonomous vehicles and healthcare, leveraging ML algorithms to analyze data, recognize patterns, and continuously improve through learning. Given the complexity of developing and integrating ML components into a system, both technical and social challenges arise throughout the development of an ML-enabled system. Technical debt is a clear example of a technical challenge, referring to the accumulation of suboptimal design decisions or deferred maintenance tasks that ultimately increase the cost of future system improvements and maintenance ( Zhang et al., 2022 ) . A significant aspect of technical debt is

Read the complete text