“I followed what felt right, not what I was told”: Autonomy, Coaching, and Recognizing Bias Through AI-Mediated Dialogue

Paper Title

“I followed what felt right, not what I was told”: Autonomy, Coaching, and Recognizing Bias Through AI-Mediated Dialogue

Publication Info

• Topic area: AI-mediated dialogue and bias recognition in social interactions.
• Keywords: AI-mediated dialogue, ableism, microaggressions, bias recognition, coaching, conversational AI, inclusion, framing effects, nudges, accessibility.

Background and Problem

• Problem / challenge: Ableist microaggressions are pervasive, yet interventions to help people recognize them are limited. Existing methods are often passive, labor-intensive, or fail to address conversational dynamics where microaggressions typically occur.
• Significance: Addressing ableist microaggressions is critical for fostering inclusion and reducing stigma against disabled individuals, a group comprising 16% of the global population. Effective interventions can mitigate psychological harm and structural exclusion.
• Motivation and related work: Prior research has documented microaggressions’ prevalence and effects but has largely focused on race, gender, and sexuality. Studies on ableism are fewer and primarily qualitative. Conversational AI systems, which increasingly mediate sensitive interactions, can encode biases and influence user judgments. This paper investigates whether AI-mediated dialogue can shift recognition of ableism and how coaching direction shapes these shifts.

Solution

• Proposed approach: An AI-mediated dialogue platform that embeds recognition of ableist interactions within simulated conversations, using coaching prompts to model biased or inclusive framings.
• Novelty:
  1. Introduction of an AI-mediated dialogue platform for studying ableism in situ.
  2. Development of a validated vignette corpus of disability-related interactions spanning four microaggression domains.
  3. Empirical evidence showing that dialogue-based interventions outperform passive reading modules in bias recognition.
  4. Design implications for socially-aware AI systems, including safeguards against biased steering and guidelines for inclusive scaffolding.
• Procedure and key techniques:
  • Participants engaged in text-based dialogues with an AI character representing a disabled person.
  • Four conditions: Bias-Directed (biased prompts), Neutral-Directed (inclusive prompts), Self-Directed (no prompts), and Reading (non-dialogue control).
  • Pre- and post-test vignette surveys measured recognition of ableist and neutral interactions.
  • Qualitative reflections captured participants’ reasoning and perceptions of the intervention.

Results

• Concrete findings:
  • Dialogue-based conditions improved recognition of ableist and neutral interactions more than the Reading condition.
  • Bias-Directed coaching heightened sensitivity to harm but led to a negative halo effect on neutral scenarios.
  • Neutral-Directed coaching fostered balanced judgments and affirmation of inclusion.
  • Self-Directed dialogue supported authentic, context-sensitive engagement.
• Advantage over baselines:
  • Dialogue-based interventions outperformed Reading in both recognition and differentiation of ableist versus neutral interactions.
  • Bias-Directed coaching produced the strongest differentiation but at the cost of overall negativity.
• Experiments / evaluation:
  • 160 participants completed pre-test, intervention, and post-test surveys.
  • Vignettes validated as reliable instruments for distinguishing ableist and neutral interactions.
  • ANOVA and thematic analysis revealed significant effects of coaching direction on recognition and emotional impact.
• Limitations and future work:
  • Lack of multimodal richness (e.g., non-verbal cues).
  • Short-term exposure limits assessment of durability and behavioral transfer.
  • Sample demographics skewed toward younger, English-speaking participants.
  • Future studies could explore immersive environments, longitudinal effects, and broader applications to other forms of bias.

Summary

This study demonstrates that AI-mediated dialogue can shift recognition of ableist microaggressions, with coaching direction playing a critical role. Bias-Directed prompts sharpened differentiation between ableist and neutral interactions through active resistance, while Neutral-Directed prompts fostered balanced, inclusive judgments. Dialogue-based interventions outperformed passive reading modules, highlighting the importance of active engagement over static instruction. The findings suggest AI-mediated dialogue as a scalable complement to disability-led education, providing a practice space for reflection and learning. Future work should explore multimodal interactions, long-term effects, and applications to other forms of social bias.