E-contact facilitated by conversational agents reduces interethnic prejudice and anxiety in Afghanistan

The study investigates whether conversational agent (CA)-facilitated electronic contact (E-contact) can reduce outgroup prejudice and intergroup anxiety among Afghanistan’s major ethnic groups (Pashtuns, Tajiks, Hazaras). Building on intergroup contact theory—which posits that contact under conditions such as equal status, cooperation, shared goals, and authority support can reduce bias—the authors focus on indirect, technology-mediated contact in a high-conflict, segregated, non-Western setting where face-to-face interaction can be risky. The paper highlights concerns that unstructured or negative contact can exacerbate prejudice, especially in conflictual contexts, and proposes a CA as a neutral, cost-efficient, and relatively unbiased facilitator that can structure discussion, promote cooperation, and embody a form of authority support. The central hypothesis (H1) is that online intergroup discussions facilitated by a CA will reduce mutual prejudice and anxiety more than unfacilitated online discussions. The study also explores whether different CA facilitation settings differentially affect outcomes.

The literature review outlines classic and contemporary work on the intergroup contact hypothesis (e.g., Allport; Pettigrew & Tropp), including evidence that contact—direct and indirect—can reduce prejudice and anxiety. It discusses the expansion of contact research to indirect forms (extended, imagined, E-contact) and notes that while positive contact is more common, negative contact can have stronger effects on prejudice. The review emphasizes gaps regarding optimal conditions for E-contact and the scarcity of studies in volatile, non-Western contexts like Afghanistan. Prior work cautions that unsupervised contact may backfire in conflict settings, underscoring the value of structured facilitation. The authors position conversational agents as a modern instantiation of authority support that can provide structure, neutrality, and scalability in online intergroup dialogues.

Design: A randomized controlled trial with two conditions (CA facilitation vs. no CA facilitation) and three measurement time points: T1 (pre-intervention), T2 (after a 2-hour synchronous discussion), and T3 (after a 3-day asynchronous discussion). A secondary analysis compared three CA facilitation styles. Participants: 128 Afghan adults (Pashtun, Tajik, Hazara), aged 23–37 with Bachelor’s or Master’s education, recruited online via an Afghan agency. From 3021 registrants (March 1–14, 2022), exclusions were made for incomplete entries, duplicates, residence outside Afghanistan, non-target ethnicities, and inadequate English. Final selection used stratified random sampling by ethnicity and sex to form 32 discussion groups of four (two Pashtun and two non-Pashtun). Groups were single-sex. Participants received $30/AFN 3000. Pre-intervention checks indicated no statistically significant baseline differences between treatment and control on prejudice, anxiety, English proficiency, or general knowledge; randomization across CA styles also showed no significant baseline differences. Intervention platform and CA: Discussions occurred on D-Agree, a text-based discussion support system structured by the Issue-Based Information System (IBIS), with automated facilitation available. The CA (chatbot) used predefined rules, patterns, and facilitation policies, organizing comments into IBIS nodes (issues, ideas, pros, cons) via argumentation mining (BILSTM classifier) and posting targeted prompts with an agent-to-user post ratio threshold of approximately 1:3. Three CA settings were used: (1) normal CA facilitating issues, ideas, and arguments; (2) CA prompting for issues (emphasizing ideas, pros, cons in response to issues); and (3) CA prompting for ideas (targeting issues and prompting idea generation). The CA was explicitly labeled “AI facilitator,” unaffiliated with any ethnic group, and designed as a neutral moderator. Tasks and procedure: The experiment ran May 23–June 9, 2022, in four batches (8 groups/batch; 2 per condition). After informed consent and pre-discussion surveys (T1), participants engaged in a two-hour synchronous discussion to collaboratively prepare a proposal regarding Afghanistan’s frozen assets (to instantiate cooperative, goal-oriented contact). Immediately after, they completed a post-synchronous survey (T2). The next day, they joined a three-day asynchronous discussion to collaboratively draft a policy proposal for upgrading informal settlements; participants rotated summarizer roles, while the CA only facilitated. A final post-discussion survey (T3) followed. Identities were anonymized but labeled by ethnicity; transparency measures ensured participants recognized the CA as non-human. Measures: Primary outcomes were intergroup prejudice and intergroup anxiety. Prejudice was assessed with the Bogardus social distance scale (1–7; lower scores indicate closer social distance/willingness for closer relations). Intergroup anxiety was measured with a 6-item bipolar adjective scale adapted from Stephan & Stephan (Swart et al.’s adaptation). Social engagement metrics included number of ideas generated and total word count of contributions on the platform. Statistical analysis: Repeated measures ANOVAs with Condition (between) × Time (within) for prejudice and anxiety tested H1. Robust alternatives were applied where assumptions might be violated: Pillai’s trace for multivariate tests; Welch’s t-tests and Games-Howell for unequal variances; and Bayesian analyses with default JASP priors to corroborate null findings. Confidence intervals around multivariate effect sizes were bootstrapped (5000 resamples, percentile method). Software: IBM SPSS 28.0.1.1, JASP 0.17.1, and R. Significance set at 0.05, two-tailed. Ethics: Approved by the Ethics Committee of the Graduate School of Informatics, Kyoto University (KUIS-EAR-2021-020). Informed consent obtained at multiple points. The study emphasized transparency about the CA and protected participant anonymity.

- Baseline ethnic differences: MANOVA showed no significant differences across Pashtun, Tajik, and Hazara groups on combined prejudice and anxiety (Wilks’ Λ = 0.949, F(4, 248) = 1.652, p = 0.162). - Manipulation check (engagement): CA-facilitated groups contributed more. • Number of ideas: CA M = 26.28 (SD = 15.54) vs. control M = 19.81 (SD = 14.01), t(126) = 2.088, p = 0.039, d = 0.426, 95% CI [0.022, 0.829]. • Word count: CA M = 1908.56 (SD = 956.15) vs. control M = 1503.69 (SD = 829.44), t(126) = 2.141, p = 0.034, d = 0.437, 95% CI [0.032, 0.840]. - Primary outcomes (H1): Significant Condition × Time interactions indicated CA facilitation reduced prejudice and anxiety relative to control. • Intergroup prejudice: Wilks’ Λ = 0.932, F(2, 125) = 4.560, p = 0.012, partial η² = 0.068. Within CA condition, T1 → T3 decreased significantly, t(95) = 2.129, p = 0.036, d = 0.217. Control showed a non-significant increase, t(31) = -1.882, p = 0.069, BF10 = 0.903. • Intergroup anxiety: Pillai’s Trace = 0.050, F(2, 125) = 3.309, p = 0.040, partial η² = 0.050. CA condition decreased from T1 → T2, t(95) = 3.665, p < 0.001, d = 0.374, and T1 → T3, t(95) = 3.729, p < 0.001, d = 0.381. Control showed no significant changes (e.g., T1 → T3, t(31) = 0.070, p = 0.944, BF10 = 0.189). • Descriptive means (Table 2): Prejudice means (CA: 3.04 → 2.50; Control: 2.50 → 3.00) and anxiety means (CA: 1.34 → 1.17; Control: 1.20 → 1.20) from T1 to T3 are consistent with these effects. - CA facilitation styles: No significant differences across the three CA policies in trajectories of prejudice (Pillai’s Trace = 0.009, F(4, 186) = 0.206, p = 0.935) or anxiety (Pillai’s Trace = 0.020, F(4, 186) = 0.468, p = 0.759).

Findings support the hypothesis that CA-facilitated E-contact can improve intergroup attitudes in a highly conflictual, segregated setting. The CA not only reduced intergroup prejudice and anxiety over time but also increased participants’ engagement, suggesting that structured, neutral facilitation may promote cooperative, respectful exchanges—aligning with the authority support component of contact theory. Effect sizes were small to medium, comparable to meta-analytic estimates for contact and E-contact, which is notable given Afghanistan’s challenging context where contact effects can be attenuated. The lack of differential effects across CA facilitation styles implies that the CA’s neutral, nonjudgmental presence and discussion structuring may be the active ingredients, more so than specific content-targeting policies. Practically, CA facilitation offers a scalable, cost-efficient approach to support healthier online intergroup interactions and may inform moderation or intervention strategies, though real-world feasibility outside controlled settings requires further testing.

This study demonstrates that a neutral, structured conversational agent can facilitate online intergroup discussions that lead to increased engagement and reductions in outgroup prejudice and intergroup anxiety among Afghan ethnic groups. Contributions include empirical evidence from a non-Western, conflict-affected context; validation that CA facilitation can function as a form of authority support within intergroup contact; and the finding that different content-focused CA policies yield similar benefits. Future research should (a) compare CA versus human facilitators to examine mechanisms such as empathy and adaptability, (b) develop and evaluate bias-detection and response capabilities within CA systems, (c) test external validity and sustainability of effects in naturalistic online environments and diverse populations, and (d) assess long-term outcomes and transfer to offline attitudes and behaviors.

- Unequal group sizes when comparing the combined CA conditions to the control condition may affect precision and generalizability; future work should balance sample sizes across conditions. - The sample skewed toward higher education, English proficiency, and reliable internet access to ensure equal-status contact, which may limit generalizability to the broader Afghan population. - The intervention involved introducing a CA in a controlled experimental setting; adoption, effectiveness, and ethics in real-world, unsupervised environments may differ and warrant cautious extrapolation.