Governing Its Shadow Impacts
A 90-minute working session reframing explainable AI from a model-level feature to a resource-consuming architectural quality attribute — and building the governance vocabulary to manage it.
Current practice treats explainability as a cost-neutral model property. This session argues it is a governed architectural concern — and asks the community to define what governing it would require.
Should explainability be formalised as a first-class architectural quality attribute, alongside performance, security, and maintainability?
How should the shadow impacts of explainability be governed within the fast-paced cycles of CI/CD pipelines and automated infrastructure?
Which specific architectural tactics can effectively mitigate structural redundancy in explanation artifacts while preserving systemic accountability?
An agreed taxonomy for transparency debt, shadow impacts, and explanation lifecycle — established during the invited talks and refined in the working groups.
A completed 3×3 canvas mapping architectural pattern adaptations and lifecycle metrics to the shadow impact dimensions — published within seven days.
Prioritised research questions across the three discussion tracks, with volunteer leads and evidence requirements recorded before the session ends.
These concepts form the shared language for the session. Each is grounded in the empirical findings described in the next section.
The operational shadow of explainability is the aggregate of resource costs — storage, compute, energy, maintenance — generated by explanation artifacts across the lifecycle of an AI-intensive system, independent of the explanatory value those artifacts provide.
Explainability is a resource-consuming architectural property that introduces systemic tensions with sustainability, privacy, and maintainability — not a cost-neutral feature.
Attribution maps, feature embeddings, model checkpoints, and saliency logs produced by XAI mechanisms. They accumulate passively and are rarely governed.
The growing architectural burden created when explanation artifacts are retained beyond their marginal explanatory value, accruing storage, energy, and maintenance costs.
A model linking the marginal explanatory value of an artifact to its operational cost — the basis for promotion, demotion, and pruning lifecycle decisions.
Shadow impacts manifest in storage tiers, AIOps toolchains, and governance layers — not in model accuracy metrics. Current observability practices do not surface them.
In microservice architectures, the rationale for a decision is distributed across independently-evolving components, making a coherent audit chain architecturally unviable.
Artifacts whose marginal explanatory value exceeds their operational cost are promoted to long-term storage with full provenance metadata.
Artifacts that retain some value but are infrequently accessed are demoted to cheaper storage, reducing retrieval and compute overhead.
Artifacts whose marginal value has fallen to zero — successive near-identical checkpoints, stale attribution maps — are pruned to reduce energy and storage footprint.
Characterising and measuring the accumulation of explanation artifacts — redundancy cost functions, observability signals, and cross-dimension shadow impact interactions.
Patterns and tactics for governing explanation artifacts — explanation contracts, lifecycle policies, surrogate fidelity trade-offs in edge–cloud deployments.
Formalising XAI governance — CI/CD regression tests, ISO/IEC 25010 extension, ATAM adaptation, and the interaction with EU AI Act retention requirements.
Both findings are external to model performance. Neither appears in existing architecture quality models or CI/CD governance frameworks.
Longitudinal study · large-scale medical imaging pipelines · millions of samples · iterative augmentation experiments
Explanation artifacts — high-dimensional attribution maps, feature embeddings, and model checkpoints — accumulate passively across every iteration of the evolution cycle. These artifacts exhibit high informational redundancy without explicit architectural lifecycle evaluation. Storage, retrieval, and maintenance costs grow non-linearly relative to the system's utility. The marginal explanatory value of each successive artifact approaches zero while its operational cost remains constant.
Architectural decay through passive artifact accumulation across every CI/CD iteration.
Non-linear growth in storage, retrieval, and maintenance costs relative to utility gain.
No current lifecycle policy triggers demotion or pruning of stale explanation artifacts.
Joint research programme · Queen's University Belfast & Concordia University · generalised from Discovery 1
Embedding explanation mechanisms triggers a cascade of hidden impacts across multiple architectural dimensions simultaneously. These impacts are external to the model — they manifest in storage tiers, AIOps iterations, and governance layers, not in accuracy metrics. In a continuous software engineering context, they accumulate invisibly because current practice treats explanations as passive outputs rather than governed architectural assets.
Computational overhead, data pipeline complexity, provenance tracking cost.
Increased energy footprint from artifact generation, retention, and retrieval at every iteration.
Attribution artifacts retain training data signals; persistent retention elevates leakage risk.
Explanation overload for stakeholders; degraded decision quality from artifact saturation.
Maintenance debt accumulating invisibly across teams and release cycles.
Compliance drift as explanation chains fragment across independently-evolving agents.
Cross-cutting finding · service-oriented and microservice architectures · CI/CD-deployed AI-intensive systems
In service-oriented and microservice architectures, AI evolution leads to explanation fragmentation — the rationale for a decision is distributed across multiple autonomous agents that evolve independently. No current continuous architecture practice includes explanation regression tests in CI/CD, lifecycle policies for explanation artifacts, or cross-boundary explanation contracts between services.
Explanation regression tests — no fidelity threshold gate exists in any standard pipeline.
Cross-boundary explanation contracts — no schema specifies what one service owes another in XAI provenance.
Explainability as a lifecycle-governed attribute — ISO/IEC 25010 and ATAM offer no mechanism for transparency debt.
The invited talks establish shared vocabulary and frame the empirical evidence. The 45-minute working session converts that context into a governance canvas and a prioritised research agenda.
Facilitators frame the session in one sentence: explainability accumulates costs that architecture does not currently govern. Session structure and track overview. Participants self-assign to one of three discussion tracks.
Three parallel tracks refine the pre-seeded research questions, then converge in plenary synthesis onto the shared Governance Canvas.
Working sheets distributed. Each group nominates a scribe and a skeptic — the skeptic is required to challenge the first consensus reached.
Groups evaluate, refine, or reject the four pre-seeded research questions. May add one new question. Output: ranked list of 2–3 RQs with evidence requirements.
Each track reports in 4 minutes. Facilitator maps outputs onto the Governance Canvas. Final 3 minutes: dot vote on top cross-cutting tension and formalization position.
Daniel Balouek is a Research Scientist at Inria, the French National Research Institute for Digital Science and Technology. His research focuses on distributed computing and complex applications, with a special emphasis on next-generation Utility Computing Infrastructures—encompassing Cloud, Edge, and beyond—as well as Urgent Science. He is a co-recipient of an Outstanding Paper Award in Artificial Intelligence for Social Impact at the AAAI Conference on Artificial Intelligence. He holds a PhD in Computer Science from École Normale Supérieure de Lyon, France. His career has spanned institutions in industry and academia across France, the USA, Japan, and India.
Dr. Wahab Hamou-Lhadj is Professor and Chair of the Department of Electrical and Computer Engineering at Concordia University. He previously served as an Affiliate Researcher with the NASA Jet Propulsion Lab at the California Institute of Technology from 2022 to 2025 and continues to collaborate with the lab on advanced research initiatives. He leads a research group specializing in software observability, AI for IT Operations (AIOps), software tracing and logging, and model-driven engineering. He is a Senior Member of IEEE and a long-standing member of ACM.
Present the outputs of the working session, including the refined research questions and the completed Governance Canvas. Discuss next steps for publication and collaboration.
Full Professor · Gina Cody School of Engineering and Computer Science · Concordia University, Montréal
Specialises in architecture-level quality assurance for AI-intensive systems. ICSE NIER 2026 work formalises the redundancy cost function and the operational shadow of explainability.
Lecturer in Software Engineering · Queen's University Belfast
Leads research in cloud-native software engineering and edge intelligence, focusing on architecture design and quality assurance for AI-enabled systems in distributed environments.