← CHI 2026 map

CHI '26 · Best paper · full-paper review · confidence high

Narratives and Perspectives: How AI Summaries Steer Users' Opinions and Engagement on Social Media

Jarod Govers , Cherie Sew , Eduardo Velloso , Vassilis Kostakos , Jorge Goncalves

DOI PDF Program page Page JSON

This paper is strong because it tests a live platform design question with a controlled experiment and shows that AI summaries are not neutral wrappers around discussion. The percentage display increases conformity toward majority views, while narrative summaries can create false balance and dampen opinion change. The contribution is important for HCI because it identifies summary format as a causal design lever, but the claims should remain bounded to simulated Reddit-style threads rather than assumed to transfer directly to live platforms.

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

This is a strong CHI paper because it takes a seemingly mundane interface feature—AI summaries in comment threads—and shows that the feature can materially steer opinion formation. The central contribution is not just that summaries matter, but that different summary formats matter in different ways. The percentage breakdowns of agreement/disagreement amplify conformity toward majority views, while narrative summaries can make polarized discussions appear more balanced than they are, thereby reducing opinion change. That is a meaningful empirical contribution because many people, including designers, would treat summaries as neutral comprehension aids rather than as interventions that reshape social influence dynamics. The paper is also careful in how it validates the claim. It uses a 144-participant controlled experiment on simulated Reddit-style discussions, includes both civil and toxic discourse conditions, and analyzes outcomes with mixed quantitative models plus qualitative responses. That gives the work a solid causal footing for the task class it studies. Another strength is that the authors do not overstate the downstream effects: they report that the AI tools did not significantly affect willingness to engage, while discourse tone did. That narrower result improves credibility because it shows the paper is distinguishing opinion effects from engagement effects rather than collapsing them together. The main caution is ecological validity. The discussions are simulated, the comments are generated and controlled, and the interaction context is much simpler than a live platform with recommendation systems, social ties, repeated exposure, and richer behavioral incentives. The limitations section also makes clear that model performance, cultural nuance, meta-irony, and platform-specific formats could change the observed effects. So the right reading is that this paper offers strong causal evidence about a plausible and important mechanism in social-media interface design, not a final answer about every real deployment. Even with that caveat, the contribution is substantial because it identifies a concrete design lever—summary framing—that can intensify conformity or create false balance, and it translates that finding into actionable design implications for responsible AI deployment.

What Changed

Canon before

It is commonly assumed that AI-generated summaries on social media primarily serve as neutral aids to enhance understanding and reduce information overload, without significantly influencing user opinions beyond comment content itself, and that these summaries potentially encourage engagement.

Departure from common sense

Contrary to the assumption of neutrality, displaying AI-generated commenter agreement percentages significantly amplifies social conformity beyond reading comments alone, steering users toward majority viewpoints and polarising opinions further; meanwhile, AI narrative summaries cause misperceptions of debate balance, reducing opinion change, challenging the expectation that summaries simply aid understanding without shaping opinion formation.

Actual novelty

This paper empirically demonstrates how two types of AI summaries—percentage breakdown visualizations and textual narrative summaries—differentially influence user opinion change on social media; specifically, that percentage displays increase conformity toward majority views while narrative summaries moderate opinion change by creating false perceptions of balance, revealed through a controlled 144-participant experiment on simulated Reddit threads covering civil and toxic discourse, along with design implications for mitigating AI-driven polarisation effects.

Evidence

The evidence comes from a controlled 144-participant experiment on simulated Reddit-like threads with civil and toxic discourse, using mixed quantitative models and qualitative thematic analysis. The results show that AI percentages increase conformity and opinion shift magnitude, AI narratives can reduce opinion change by creating false balance perceptions, and AI tools do not significantly affect willingness to engage, while thread tone does.

Limits

Method limits

The study uses simulated Reddit threads generated via GPT-4.5, which limits ecological validity and diversity of real-world discourse. The design also fixes reading time and topic structure, and the sample is US-based, so the findings may not transfer cleanly to other populations, languages, or more naturalistic browsing behavior.

Deployment limits

Findings are based on simulated environments rather than live social media platforms; actual AI summary deployments with different interfaces, moderation regimes, or recommendation systems may alter effects. Longer-term exposure, misinformation dynamics, and cross-cultural deployment remain untested.

Boundary conditions

Effects were observed in a US English-speaking sample engaging with simulated Reddit threads on selected civil and toxic political topics. The results are most defensible for short-form comment-thread settings where summaries are shown before reading comments, and may vary with platform norms, topic stakes, and user familiarity.

Manual check: Verify exact offsets for all evidence spans against the source text if downstream tooling requires strict span validation.

Position in field

This work contributes causal evidence that AI summary tools can steer opinion formation rather than merely compress information. It extends HCI and social computing research on conformity, polarization, and AI mediation by showing that summary format matters: percentages amplify majority pull, while narrative summaries can create false balance and dampen change.

Abstract

AI summaries on social media are reshaping how users form opinions about political topics, yet their influence remains largely unexamined despite their widespread deployment. This paper investigates how two types of AI summaries affect user opinions and engagement: textual summaries of discussion narratives and percentage breakdowns of agreement/disagreement. Through a 144-participant experiment on simulated online discussion threads, we found that displaying commenter agreement percentages amplified social conformity towards the majority views beyond reading comments alone. Conversely, AI narrative summaries created misperceptions of balance in polarised threads, reducing opinion change. While these summaries did not influence participants’ willingness to engage, toxic discussions deterred participation even when participants held majority views. Based on our findings, we provide critical design interventions for industry and researchers to mitigate these tools' polarising effects, paving the way for responsible AI deployment on social media platforms.