On July 2, 2026, researchers Bailey Flanigan and Michelle Si submitted a paper titled Internal Pluralism and the Limits of Pairwise Comparisons to arXiv. The study critiques the assumptions behind local pairwise comparisons used in artificial intelligence (AI) for decision-making processes, particularly in participatory design and alignment.
Understanding Internal Pluralism in AI
The concept of internal pluralism suggests that individuals hold multiple, sometimes conflicting, priorities when evaluating decision rules. This poses significant challenges for AI systems that rely on local comparisons. The authors argue that these comparisons often overlook the broader implications of decision-making criteria, such as proportionality, egalitarianism, and equal treatment.
By relying solely on local pairwise comparisons, AI systems may fail to accurately capture these global priorities. The authors provide a formal model that illustrates how individuals may struggle with decisiveness due to conflicting values, leading to potential distortions in reported preferences.
Failures of Local Pairwise Comparisons
Flanigan and Si identify two main failures of forced local pairwise comparisons:
- Global Priorities: Local comparisons may not adequately reflect how priorities interact across different contexts, thus missing essential global perspectives.
- Internal Conflicts: Individuals may experience tension between strongly-held priorities, resulting in indecision and inaccuracies in preference reporting.
Their findings indicate that allowing individuals to express indecision can significantly enhance the accuracy of preference learning, reducing the number of required queries.
Implications for Preference-Learning Methods
The model proposed by Flanigan and Si suggests new avenues for developing preference-learning methods that directly elicit priorities from individuals. This approach aims to yield more accurate and interpretable insights into what individuals value in AI-driven decision-making.
By addressing the limitations of pairwise comparisons, researchers and practitioners can create more effective AI systems that align with human values and priorities.
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv AI. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
