The Ideological Turing Test for Moderation of Outgroup Affective Animosity

TL;DR

The paper tests the Ideological Turing Test, a gamified approach combining perspective-taking with structured debate or writing to reduce affective animosity and ideological polarization. In a 2x2 design across modality (debate vs writing) and perspective (Own vs Opposite), opposite-perspective interventions reduced both affective polarization and ideological distance, with writing yielding large immediate ideological shifts and debate yielding more durable affective gains. The findings reveal distinct temporal trajectories: non-adversarial writing fosters short-term empathy and shifting positions, while cognitive engagement in debate sustains affective improvements over weeks, and ideological shifts persist across arms. Winning judgments did not meaningfully moderate outcomes, and willingness to re-engage was modest and similar across conditions, highlighting the potential and limits of scalable, depth-focused polarization interventions for real-world deployment, including AI-assisted moderation.

Abstract

Rising animosity toward ideological opponents poses critical societal challenges. We introduce and test the Ideological Turing Test, a gamified framework requiring participants to adopt and defend opposing viewpoints, to reduce affective animosity and affective polarization. We conducted a mixed-design experiment ($N = 203$) with four conditions: modality (debate/writing) x perspective-taking (Own/Opposite side). Participants engaged in structured interactions defending assigned positions, with outcomes judged by peers. We measured changes in affective animosity and ideological position immediately post-intervention and at 2-6 week follow-up. Perspective-taking reduced out-group animosity and ideological polarization. However, effects differed by modality (writing vs. debate) and over time. For affective animosity, writing from the opposite perspective yielded the largest immediate reduction ($Δ=+0.45$ SD), but the effect was not detectable at the 4-6 week follow-up. In contrast, the debate modality maintained a statistically significant reduction in animosity immediately after and at follow-up ($Δ=+0.37$ SD). For ideological position, adopting the opposite perspective led to significant immediate movement across modalities (writing: $Δ=+0.91$ SD; debate: $Δ=+0.51$ SD), and these changes persisted at follow-up. Judged performance (winning) did not moderate these effects, and willingness to re-participate was similar across conditions (~20-36%). These findings challenge assumptions about adversarial methods, revealing distinct temporal patterns: non-adversarial engagement fosters short-term empathy gains, while cognitive engagement through debate sustains affective benefits. The Ideological Turing Test demonstrates potential as a scalable tool for reducing polarization, particularly when combining perspective-taking with reflective adversarial interactions.

Figures (33)

• Figure 1: Left: Intervention arms, 2x2 design including participant counts assigned to each arm. Here, the focus participant has four different possibilities once they engage in the in-person sessions. There is an issue at hand for which we have the participant's true opinion based on a pre-intervention survey. In this case, the question is: "Should pineapple be on pizza?" (only for pictorial purposes, as topics in our experiment are more serious and potentially divisive). Imagine that the participant agrees with the statement; they will be a true PRO. The participant can engage in either a debating modality against a peer (top row) or a writing-only modality (bottom row). However, the participant is randomly assigned to engage either in defending their own position (Own) or the opposite position (Opp), in which case they would be considered as "a pretender". We also note that debating and writing happen via an anonymized chat interface. Right: Judging and incentives. After the intervention occurs, a panel of judges will assess either the debate log or the written statements of paired participants. The judges assess not only who presents the best argument but also who is a pretender. An intervention participant only wins when both conditions are fulfilled. We use this future judgment as an incentivizing method for participants.
• Figure 2: Adjusted within–arm change $\Delta_{t}$ from pre to pos ($t{=}\text{pos}$) and follow ($t{=}\text{fol}$) for affective (top; higher means warmer outgroup feelings) and ideological (bottom; higher means movement toward the opposite stance or moderation when applicable) outcomes. Points are estimated marginal means (EMMs) with 95% CIs; the dashed line marks no change. Key remark: the largest immediate gains appear for (Write, Opp) at $t{=}\text{pos}$, while the sustained affective benefit at $t{=}\text{fol}$ is primarily (Debate, Opp); ideological shifts persist for both perspective–taking arms.
• Figure 3: Shares of participants who Improve, show No Change, or Worsen relative to pre for each arm and outcome. Top panel: Affective polarization, where 'Improve' corresponds to warmer feelings toward the issue's opponent. Bottom panel, ideological position where "Improve" corresponds to a movement toward the assigned opposite stance or toward moderation, when applicable. Bars show stacked proportions at post and at follow for each arm.
• Figure 4: Perspective contrasts within modality: $\delta^Y(t|m)$ which is the incremental pre–to–$t$ improvement attributable to assigning the opposite perspective rather than one's own perspective within the same modality $m$ on the EMM (response) scale. So $\delta>0$ means the opposite perspective produced a larger gain, $\delta=0$ means equal change, and $\delta<0$ means an own-perspective advantage. Points are EMMs with 95% CIs; the dashed line indicates no change. Key remarks: contrasts are generally positive, with weaker detectability for affect and clearer, more persistent advantages for ideology.
• Figure 5: Within–perspective estimated changes ($\Delta_{\text{pos}}$, $\Delta_{\text{fol}}$) by modality for Own vs. Opp. Points are EMMs with 95% CIs; the dashed line indicates no change. Key remark: the perspective advantage for Opp is driven by Own lying near zero—particularly in writing—while Opp shows clear movement, especially on ideology.