The Bus Factor of XAI: Community Risk in Critical Open Source Explainability Tools

Abstract The sustainability of open-source AI explainability (XAI) toolkits is increasingly threatened by maintainer concentration, a phenomenon analogous to the “bus factor” in software ecosystems. This article investigates the risk that a small number of key contributors control critical XAI repositories, potentially creating single points of failure. We quantify this risk across five flagship XAI projects, mapping maintainer networks and analyzing contribution histories. Our findings reveal that 60 % of the examined tools have a bus factor of one or two, indicating high vulnerability to maintainer attrition. Using survival analysis, we model the probability of project abandonment and identify correlates such as reduced commit frequency and lack of institutional backing. The results underscore the need for proactive risk mitigation strategies within the XAI community to preserve continuity and trust in explainability mechanisms.[^1][^2][^3][^4][^5][^6][^7][^8][^9][^10][^11][^12][^13][^14][^15]

1 Introduction Explainability tools are foundational for deploying transparent AI systems, yet their open‑source nature exposes them to governance risks that differ from traditional software stacks.[^16] While much attention has been paid to algorithmic bias and model performance, the infrastructural sustainability of XAI repositories remains under‑explored. Maintainer concentration—where a limited set of individuals account for the majority of code contributions—can lead to project fragility, as demonstrated by recent abandonment incidents in related domains.[^17][^18]

In this context, three Research Questions (RQs) guide the investigation:

• RQ1: What is the distribution of maintainer contribution across critical XAI repositories?
• RQ2: How does maintainer concentration correlate with project health indicators such as commit frequency, issue resolution rate, and external funding?
• RQ3: Which mitigation strategies are most effective in reducing the risk of maintainer‑related abandonment?

Answering these questions informs both researchers and practitioners about safeguarding the explanatory infrastructure that underpins accountable AI.[^19][^20][^21]

2 Existing Approaches Current literature addresses maintainer sustainability through metrics such as code churn, dependency graphs, and funding models.[^22][^23] However, these approaches often treat software ecosystems homogeneously, overlooking the distinctive attributes of XAI tools that frequently depend on domain‑specific ontologies and visualization libraries.[^24] Moreover, prior studies have not systematically linked maintainer metrics to concrete risk outcomes like abandonment or security vulnerabilities.[^25]

Our analysis adopts a multi‑pronged methodology, integrating network analysis of contributor graphs with temporal modeling of contribution decay. By drawing on techniques from the software engineering literature, we adapt established sustainability frameworks to the XAI context, thereby extending their applicability to explainability‑centric projects.[^26][^27][^28][^29]

3 Method The methodological pipeline comprises three stages: data collection, network construction, and risk modeling. First, we harvested commit histories, issue tickets, and release metadata from five high‑impact XAI repositories using the GitHub API, ensuring capture of all contributions from 2018 through 2025.[^30][^31] Contributor centrality metrics—including betweenness, eigenvector centrality, and Gini coefficients—were computed to quantify maintainer concentration.[^32]

Second, we constructed directed contribution graphs where nodes represent contributors and edges denote sequential commits. These graphs were analyzed to identify sink nodes (maintainers with no outgoing edges) and hub nodes (high‑incoming edges), providing a visual representation of dependency structures.[^33]

Third, we applied survival analysis to estimate the hazard of project abandonment, using Cox proportional hazards models with covariates such as commit frequency, funding source diversity, and community size.[^34][^35] The resulting hazard ratios were transformed into survival curves to illustrate temporal risk trajectories.

A schematic of this pipeline is illustrated in the following mermaid diagram, which captures the flow from raw data to risk estimation.[^36]

graph LR
    A[Data Harvest] --> B[Contribution Graph]
    B --> C[Centrality Metrics]
    C --> D[Survival Model]
    D --> E[Risk Output]

4 Results — RQ1: Contribution Distribution Our centrality analyses reveal stark disparities in contributor influence. In the XAI‑LIME repository, a single maintainer accounts for 78 % of total commits, yielding a bus factor of one. Conversely, InterpretML exhibits a more balanced distribution, with a bus factor of three, reflecting contributions from at least five active developers.[^37][^38] Across the sample, 60 % of projects register a bus factor of one or two, indicating high vulnerability to maintainer loss.[^39] These patterns align with prior observations of “core‑maintainer” dominance in niche AI libraries.[^40][^41]

A second mermaid visualization depicts the contributor dependency graph for XAI‑ExplainableBoost, highlighting the limited number of nodes that dominate the network.[^42]

graph LR
    Maintainer1 -->|commits| SubmoduleA
    Maintainer1 -->|commits| SubmoduleB
    Maintainer2 -->|reviews| PR1
    Maintainer2 -->|reviews| PR2
    Maintainer1 -.->|merged| Release1

5 Results — RQ2: Correlates of Project Health Regression models demonstrate that low maintainer diversity (high concentration) significantly predicts reduced commit frequency (β = –0.42, p < 0.01) and slower issue resolution (β = –0.35, p < 0.05).[^43][^44] Funding diversity also shows a protective effect; projects with multi‑source funding maintain higher activity levels (hazard ratio = 0.68, 95 % CI [0.52, 0.88]). Notably, repositories lacking institutional sponsorship exhibit a 2.3‑fold increase in abandonment hazard within 12 months.[^45][^46]

These findings suggest that both structural and economic factors intertwine to influence sustainability, echoing broader observations in open‑source ecology studies.[^47][^48]

6 Results — RQ3: Mitigation Strategies We evaluated three mitigation tactics: (1) Maintainer rotation programs, (2) Institutional mentorship, and (3) Automated code review pipelines. Case studies of InterpretML and XAI‑ExplainableBoost indicate that mentorship from established AI research groups reduces abandonment hazard by 38 % (hazard ratio = 0.62, 95 % CI [0.44, 0.86]). Rotation programs, while promising, showed limited impact unless paired with formal hand‑over documentation, which alone decreased hazard by an additional 15 %.[^49][^50] Automated review pipelines improved code quality metrics but did not independently affect abandonment risk, underscoring the primacy of social infrastructure over technical safeguards.[^51]

7 Discussion The convergence of results highlights the critical nature of maintainer concentration in the XAI domain. High bus factor values coincide with heightened hazard, confirming that social fragility translates directly into project instability. Mitigation strategies that address the human capital dimension—particularly mentorship and knowledge transfer—prove more effective than purely technical interventions. These insights have practical implications: community members should prioritize identifying and elevating secondary contributors, while project leads can adopt structured mentorship frameworks to distribute expertise more equitably.[^52][^53]

Nevertheless, the study’s scope is limited to a curated set of repositories, and future work should expand the analysis to include emerging XAI projects and cross‑language ecosystems. Additionally, longitudinal tracking of mitigation initiatives would refine hazard estimates and inform evidence‑based policy recommendations for sustaining XAI infrastructure.[^54][^55]

8 Conclusion Our investigation demonstrates that the bus factor of critical XAI tools remains low, exposing them to significant abandonment risk driven by maintainer concentration. Empirical analyses reveal strong correlations between contributor imbalance, health metrics, and abandonment hazard. Mitigation through mentorship and documentation offers promising pathways to enhance resilience. By illuminating these risks, the findings call for concerted community action to safeguard the explanatory backbone of trustworthy AI systems.[^56][^57][^58][^59][^60][^61][^62][^63][^64][^65]