[Paper] The Runtime Dimension of Ethics in Self-Adaptive Systems

Source: arXiv - 2602.17426v1

Overview

Self‑adaptive software systems are increasingly sharing physical and virtual spaces with people—think autonomous drones, smart factories, or AI‑driven health assistants. When these systems make decisions on the fly, ethical considerations (fairness, safety, privacy, environmental impact) can no longer be hard‑wired once at design time. This paper argues for runtime ethical reasoning, treating ethics as a dynamic requirement that must be continuously elicited, represented, and negotiated among all stakeholders.

Key Contributions

• Runtime‑first ethics model – Shifts ethics from static, rule‑based constraints to runtime requirements that can be updated as contexts and stakeholder values evolve.
• Ethics‑as‑Negotiation framework – Introduces explicit, multi‑party negotiation mechanisms to resolve conflicts among competing ethical preferences while staying inside a legally mandated “hard‑ethics” envelope (e.g., safety regulations).
• Taxonomy of ethical challenges – Systematically categorizes sources of ethical uncertainty, value conflicts, and multi‑dimensional drivers (human, societal, environmental).
• Research agenda & open questions – Outlines concrete directions for building ethically self‑adaptive systems, from formal requirement languages to runtime monitoring and decision‑making algorithms.
• Bridging disciplines – Connects self‑adaptive systems engineering with moral philosophy, human‑computer interaction, and regulatory compliance, highlighting where cross‑domain collaboration is needed.

Methodology

The authors adopt a concept‑driven, interdisciplinary approach:

1. Literature synthesis – Review of existing self‑adaptive architectures, ethical AI frameworks, and requirement‑engineering techniques to pinpoint where current solutions fall short.
2. Scenario analysis – Real‑world use cases (e.g., collaborative robots, autonomous vehicles) are examined to illustrate how ethical preferences can diverge among users, regulators, and the environment.
3. Requirement‑centric modeling – Proposes a runtime requirement model that treats ethical preferences as first‑class citizens, enabling continuous elicitation and revision.
4. Negotiation abstraction – Sketches a high‑level negotiation protocol (similar to multi‑agent contract net) that can mediate trade‑offs among stakeholders while enforcing non‑negotiable safety/legal constraints.

The methodology stays deliberately high‑level, aiming to inspire concrete implementations rather than deliver a finished prototype.

Results & Findings

• Static ethics are insufficient – Fixed rule sets cannot accommodate the fluid nature of human values, leading to either over‑constrained systems or ethically blind behavior.
• Ethical uncertainty is multi‑faceted – Uncertainty stems from incomplete stakeholder input, ambiguous legal texts, and context‑dependent value interpretations.
• Conflicts are inevitable and must be negotiated – Even with a shared “hard‑ethics” envelope, trade‑offs (e.g., privacy vs. safety) arise and require systematic resolution.
• A runtime negotiation layer is feasible – By decoupling hard constraints (non‑negotiable) from soft ethical preferences (negotiable), systems can adapt decisions on the fly without violating safety or compliance.
• Research gaps identified – Formal languages for ethical requirements, scalable monitoring of ethical compliance, and user‑friendly tools for stakeholders to express preferences are still missing.

Practical Implications

In short, the paper sketches a plug‑and‑play ethical negotiation service that could sit between the adaptation engine and the execution layer, turning abstract moral concerns into actionable runtime decisions.

Limitations & Future Work

• Conceptual focus – The paper stops short of delivering a concrete prototype or empirical evaluation; the proposals remain at the architectural and theoretical level.
• Scalability concerns – Negotiation among many stakeholders and high‑frequency adaptation loops could introduce latency; performance trade‑offs are not quantified.
• Human factors – While stakeholder input is highlighted, the mechanisms for eliciting and updating preferences (e.g., UI design, consent management) are left open.

Future directions suggested by the authors include:

1. Formal language & toolchain for specifying ethical requirements and hard‑ethics envelopes.
2. Runtime negotiation engine prototypes integrated with popular self‑adaptive frameworks (e.g., MAPE‑K loops, Kubernetes operators).
3. Case‑study deployments in domains such as autonomous logistics, smart healthcare, or collaborative manufacturing to validate the approach under real‑world constraints.
4. Human‑in‑the‑loop studies to assess how non‑technical stakeholders interact with ethical preference interfaces and how trust evolves.

By tackling these next steps, the community can move from “ethical theory at design time” to truly ethically self‑adaptive systems that respect both the law and the nuanced values of the people they serve.

Authors

• Marco Autili
• Gianluca Filippone
• Mashal Afzal Memon
• Patrizio Pelliccione

Paper Information

• arXiv ID: 2602.17426v1
• Categories: cs.SE
• Published: February 19, 2026
• PDF: Download PDF