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

This CHI 2026 experimental study tests whether brief AI-mediated dialogue can shift people's recognition of ableist microaggressions, and whether the direction of AI coaching (biased, inclusive, or absent) changes the nature of that shift. The authors built a custom web platform in which participants converse with a simulated person with a disability (GPT-4o, with DALL-E-generated avatars) across everyday scenarios (a farewell party and a work office). In three of four conditions a 'coach' pane visible only to the participant generates one-way suggestions before each turn; in the fourth condition participants read a 7-page informational module instead. The four between-subjects conditions: Bias-Directed (coach nudges toward ableist framings — helplessness, minimization, denial of personhood, otherization, the four domains of the Ableist Microaggressions Scale by Conover et al.), Neutral-Directed (coach nudges toward inclusive framings), Self-Directed (no coach, unguided dialogue), and Reading (non-dialogue control). 302 participants were recruited through Prolific; after exclusions, 160 completed the two-session (Day 1 pre-test, Day 6 intervention + post-test) protocol with n=40 per condition. Participants rated 40 validated vignettes (20 ableist + 20 neutral, developed via AMS adaptation and review by three individuals with disability expertise and lived experience) on two 7-point Likert items: Q1 Standard Social Experience and Q2 Emotional Impact. Analysis combined change scores (Δ = post − pre), contrast scores (neutral − ableist, to measure differentiation), ANOVA with Tukey HSD, and Cohen's d. Dialogue-condition participants also provided open-ended reflections analysed via reflexive thematic analysis across three prompts: general reflections, coach perceptions, and unguided experiences. The paper's contributions are a validated vignette corpus (released as supplementary materials), an AI-mediated intervention platform, empirical evidence of differential effects of coaching direction, and design implications for socially-aware AI.

All three dialogue conditions outperformed Reading on recognition, but trajectories diverged sharply by coaching direction. For ableist scenarios, Bias-Directed produced the strongest sensitivity gains (Q1 Δ = −0.75, Q2 Δ = −0.74), significantly outperforming Self-Directed on Q1 (p = .040) and both Neutral-Directed (p = .019) and Self-Directed (p = .009) on Q2. For neutral scenarios, Neutral-Directed and Self-Directed preserved balanced positive judgments (Neutral-Directed Q1 Δ = +0.24, Q2 Δ = +0.20; Self-Directed Q2 Δ = +0.15), while Reading showed declines (Q1 Δ = −0.26, Q2 Δ = −0.20) — participants who only read became less likely to affirm neutral interactions. Bias-Directed sharpened contrast scores most (neutral − ableist Q1 Δ = +0.85) but at the cost of a 'negative halo' that dampened positive readings of neutral scenes. Combined-scenario change scores revealed the paradox: Bias-Directed (Q1 Δ = −0.32, Q2 Δ = −0.41) and Reading (Q1 Δ = −0.28, Q2 Δ = −0.29) both produced net-negative social judgments overall, while Neutral-Directed (Q1 Δ = +0.02, Q2 Δ = −0.04) and Self-Directed (Q1 Δ = −0.04, Q2 Δ = −0.05) preserved balance. Qualitative findings (209 coded instances across 120 responses) surfaced active resistance as a learning mechanism. 34 of 80 coached participants (42.5%, concentrated in Bias-Directed) explicitly rejected coach suggestions — one wrote 'The coach was offering rude and offensive topics so I ignored them'. 32 participants (42%, mostly Neutral-Directed) selectively followed the coach as 'scaffolding'. 56 (70%) described the coach as directive or steering. Naturalness/typicality was the most common theme (82 participants, 68.3%). Moderator effects: participants with a close disability family connection (n=28) showed stronger Q2 gains on ableist scenarios (p=.034, d=0.36). Prior chatbot experience (84%) did not moderate outcomes. Setting (party vs. office) produced no main effect.

For HCI researchers, AI designers, and DEI training developers, this paper is a rigorous experimental demonstration that 'nudges' in AI-mediated dialogue are never neutral: the same architecture that can scaffold inclusive framing can also entrench biased ones, and users are not passive recipients — many actively resist suggestions that feel wrong. The striking finding that biased nudges sharpen bias differentiation through the friction of rejection has significant implications for training-system design, but the 'negative halo' cost (reduced affirmation of neutral, safe interactions) is a real trade-off that teams building conversational AI for social-learning contexts must weigh. Concrete design takeaways: (1) Guidance is not neutral — treat coach prompts as value-laden design decisions and disclose them. (2) Prefer scaffolding over prescription — generate multiple alternative suggestions users can adopt, modify, or ignore. (3) Balance sensitivity with positivity — pair examples of harm with bias-aware alternatives. (4) 'Critical friction' exercises must be explicitly labelled, consented to, and contextualised — never emitted in production. (5) AI-mediated dialogue should complement, not replace, disability-led education. Essential reading alongside Johnson et al.'s intersectional GenAI study (10.1145/3772318.3790652) and the broader literature on ableism in LLMs (Phutane, Venkit, Sap). Limitations: US-based English speakers aged 18–44 dominated (71%); a single brief exposure leaves durability and behaviour-transfer open; text-only chat excludes the multimodal richness of face-to-face interaction; LLM social intelligence remains limited. The released 40-vignette corpus and the AI-mediated intervention platform are reusable assets for future HCI work on bias recognition.