Ethical Deliberation for Competing Values in AI Systems

This direction is merged into another direction: ethical implementation gap. Its last version stays readable and citable here.

Version 1 of 114 papers23 conceptsfrozen 2026-10-04

Most AI ethics guidelines present a list of virtues—privacy, transparency, fairness, and robustness—as if they were a set of independent goals to be checked off. However, researchers are increasingly recognizing that these principles are often in direct opposition. In practice, building a "responsible" system is not a matter of maximizing every virtue, but of navigating a complex web of trade-offs where improving one metric inevitably degrades another.

The Myth of the Ethical Checklist

Traditional "Codes of Conduct" often fail because they are too abstract for the day-to-day reality of software engineering. Alexander Pretschner argues that these static checklists cannot help a developer decide what to do when, for example, increasing system transparency might inadvertently leak sensitive user data.

Instead of treating ethics as a post-hoc compliance task, Pretschner proposes Ethical Deliberation in Agile Processes (EDAP). This framework treats ethical reasoning as a practical skill integrated into the development workflow. The core insight is that technical decisions are often normative: they involve choosing a specific point on an "ethical frontier" where no single principle can be fully satisfied without compromising another.

Why Principles Conflict

The tension between ethical goals is not just a matter of poor design; it is often a structural reality of the data and algorithms themselves. Fosca Giannotti and colleagues at the KDD Lab highlight how the pillars of trustworthy AI frequently clash. For instance, providing a highly detailed explanation for a model's prediction (Interpretability) can provide an adversary with enough information to reconstruct the training data (Privacy).

This fragmentation is a major hurdle for practitioners. Arif Ali Khan and Muhammad Azeem Akbar found that both software developers and lawmakers view these systemic trade-offs as the primary obstacle to long-term ethical compliance. When principles like transparency and privacy are at odds, developers lack a clear roadmap for which value to prioritize.

Moving Toward Holistic Frameworks

To move beyond isolated optimizations, researchers are developing frameworks that manage these trade-offs simultaneously. Hao He and Wanyu Lin proposed the PAIR²Struct workshop to address these issues in complex data formats like graphs and sequences. Their goal is to move "Responsible AI" research away from silos and toward a unified understanding of how privacy, accountability, and robustness interact.

Similarly, Weishan Zhang and Vincenzo Piuri emphasize that implementing responsible AI in social computing requires a technical roadmap for balancing competing requirements. They argue that because no single principle can be fully satisfied, the focus must shift to creating systems that are "trustworthy" through a deliberate balance of values.

Transparency as the Final Arbiter

If trade-offs are inevitable, how do we decide which values to sacrifice? The emerging consensus is that the "correct" balance is often context-dependent and cultural. Jessie J. Smith and Maarten Sap suggest that instead of enforcing a single global standard, developers should focus on explicit documentation.

In their Generative AI Ethics Playbook, they advocate for documenting the specific value judgments and compromises made during the design process. This approach acknowledges that a medical AI in one country might prioritize patient autonomy, while another might prioritize collective well-being.

The difficulty of this task is underscored by new evaluation methods like MedEthicEval. Haoan Jin and Mengyue Wu found that while large language models are becoming more capable, they still struggle with "gray area" scenarios where moral principles conflict. Measuring a model's "ethical integrity" requires more than checking for blatant violations; it requires assessing how well it navigates the nuanced trade-offs that define real-world ethics.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_ethical_deliberation_for_competing_value_79fe3849,
  title = {Ethical Deliberation for Competing Values in AI Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ethical-deliberation-for-competing-values-in-ai-systems/rdr_79fe3849ab064c60be0e05cfb34762c7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_ethical_deliberation_for_competing_value_79fe3849_v1,
  title = {Ethical Deliberation for Competing Values in AI Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ethical-deliberation-for-competing-values-in-ai-systems/rdr_79fe3849ab064c60be0e05cfb34762c7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}