Active and Passive Decisions: How Ethical Choices Are Made (and Missed) in NLP Research

Paper Title

Active and Passive Decisions: How Ethical Choices Are Made (and Missed) in NLP Research

Publication Info

• Topic area: Ethical decision-making in academic NLP research
• Keywords: AI ethics, decision-making, NLP research, infraethics, research practice, academic infrastructure, decision recognition, human-infrastructural processes, ethical deliberation, sociotechnical systems

Background and Problem

• Problem / challenge: Many consequential choices in NLP research are not recognized as decisions due to defaults, norms, and institutional structures. Existing ethics frameworks assume researchers can identify decision points, which is often not the case.
• Significance: Unrecognized decisions can embed ethical and methodological issues into research outcomes, impacting fairness, accountability, and broader AI practices.
• Motivation and related work: Prior work in AI ethics has focused on decision outcomes and frameworks but has largely overlooked how decisions become recognizable. This study addresses this gap by examining the conditions under which alternatives surface in academic NLP research.

Solution

• Proposed approach: A conceptual framework distinguishing between "active" and "passive" decisions, based on whether alternatives are visible, viable, and voiced (VVV).
• Novelty:
  1. Introduces the concept of "decision moments" and the VVV framework to analyze decision recognition.
  2. Empirically examines decision-making dynamics across four NLP research projects.
  3. Proposes a shift from compliance-based to recognition-based ethics in research practice.
• Procedure and key techniques:
  • Conducted decision-tracing interviews with eight researchers across four NLP projects.
  • Developed a spectrum for classifying decisions as active or passive based on VVV criteria.
  • Analyzed how institutional structures, time pressure, resource constraints, and distributed expertise influence decision recognition.

Results

• Concrete findings:
  • Many decisions in NLP research are passive, shaped by defaults, norms, and constraints rather than explicit deliberation.
  • Recognition of alternatives depends on crossing a threshold where they are visible, viable, and voiced.
  • Time pressure, institutional structures, resource constraints, and distributed expertise are key factors influencing decision recognition.
• Advantage over baselines: Unlike prior studies focusing on decision outcomes, this work highlights the conditions under which decisions become recognizable, offering a new lens for ethical analysis in research.
• Experiments / evaluation:
  • Four case studies: Privacy detection, Cobweb model revival, Writing Assistance tool evaluation, and Geolocation privacy risks.
  • Methods included semi-structured interviews and retrospective analysis of decision moments.
  • Findings were classified along a passive-active spectrum using the VVV framework.
• Limitations and future work:
  • Limited to four projects at a single institution; findings may not generalize across all NLP or AI research contexts.
  • Retrospective interviews may miss unrecognized decision moments.
  • Future work could explore longitudinal studies, decision dynamics in industry, and interventions to enhance decision recognition.

Summary

This study examines how academic NLP research teams recognize (or fail to recognize) decision moments, introducing a framework that classifies decisions as active or passive based on the visibility, viability, and voicedness of alternatives. Findings from four case studies reveal that many decisions are shaped by institutional structures, time pressure, and resource constraints, often rendering alternatives invisible or non-viable. The study advocates for a shift from compliance-based to recognition-based ethics, emphasizing the need to create conditions that surface hidden alternatives. This work provides a foundation for improving ethical deliberation in research by addressing the human-infrastructural processes that precede decision-making.