Emergent, not Immanent: A Baradian Reading of Explainable AI

Paper Title

Emergent, not Immanent: A Baradian Reading of Explainable AI

Publication Info

• Topic area: Onto-epistemological critique and reframing of Explainable AI (XAI) practices.
• Keywords: Explainable AI, agential realism, diffraction, interpretability, material-discursive performance, ethics, situated knowledge, entanglement, HCI, feminist epistemology.

Background and Problem

• Problem / challenge: Mainstream XAI assumes that explanations are pre-existing entities within AI models, waiting to be uncovered by external observers. This perspective overlooks the socio-technical and relational processes involved in generating interpretations.
• Significance: Reframing XAI as a relational and emergent process addresses limitations in current methods, which often fail to account for the ethical, situated, and performative dimensions of interpretability.
• Motivation and related work: Prior work has critiqued the "black-box" metaphor and the positivist assumptions underlying XAI but has not systematically applied alternative onto-epistemologies like Barad’s agential realism. This paper builds on feminist epistemology and HCI research to propose a new framework for understanding interpretability.

Solution

• Proposed approach: A Baradian onto-epistemological framework for XAI, treating interpretability as a material-discursive performance emerging from entanglements of models, humans, tools, and contexts.
• Novelty:
  1. Reframes interpretability as emergent rather than immanent, rejecting the notion of pre-existing explanations.
  2. Introduces Barad’s agential realism and diffraction as tools to critique and reimagine XAI practices.
  3. Highlights the ethical stakes of configuring explanatory entanglements and foregrounds responsibility and accountability.
  4. Proposes design directions for XAI interfaces that support emergent, situated, and plural interpretations.
• Procedure and key techniques:
  1. Critique of existing XAI methods using Baradian optics (refraction, reflection, diffraction).
  2. Analysis of how interpretability emerges through intra-actions among models, tools, and observers.
  3. Development of speculative design directions for diffractive XAI interfaces, illustrated with a text-to-music case study.

Results

• Concrete findings:
  • Mainstream XAI methods often assume an immanent ontology, treating explanations as pre-existing and discoverable.
  • Diffractive optics reveal that interpretations are co-constituted by human and non-human entanglements and are inherently situated and plural.
  • Ethical responsibility in XAI involves making agential cuts visible and addressing patterns of exclusion in explanatory practices.
• Advantage over baselines:
  • Moves beyond the limitations of refractive and reflective approaches by embracing the relational and emergent nature of interpretability.
  • Provides a framework for addressing ethical and epistemological issues that mainstream XAI overlooks.
• Experiments / evaluation:
  • Conceptual analysis of XAI methods using Baradian optics.
  • Speculative design case study for a text-to-music interface, illustrating practical applications of the proposed framework.
• Limitations and future work:
  • The framework is primarily conceptual and lacks empirical validation.
  • Design directions are not fully developed into operational guidelines or evaluated in real-world scenarios.
  • Future work should translate the framework into concrete design patterns and conduct empirical studies.

Summary

This paper critiques the dominant assumptions in Explainable AI (XAI), which treat explanations as pre-existing entities within models. Drawing on Barad’s agential realism, it reframes interpretability as an emergent, material-discursive performance arising from entanglements of models, tools, and observers. The authors highlight the ethical implications of these entanglements and propose design directions for XAI interfaces that foreground plural, situated, and evolving interpretations. While primarily conceptual, the framework offers a novel lens for advancing XAI research and practice, with future work needed to operationalize and evaluate the proposed design principles.