Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. B. F., Lee, J., Mann, M., Merhout, F., & Volfovsky, A. (2018). Exposure to opposing views on social media can increase political polarization. Proceedings of the National Academy of Sciences, 115, 9216–9221. https://doi.org/10.1073/pnas.1804840115
Summary
This preregistered field experiment tests a widely held optimistic assumption: that exposing partisans to opposing views on social media will reduce political polarization. Bail and colleagues recruited active Democratic and Republican Twitter users and paid them to follow a bot that retweeted a month’s worth of messages from elected officials, opinion leaders, media, and nonprofits of the opposing ideology. Rather than fostering moderation, cross-partisan exposure produced backfire effects — most notably a substantial and significant conservative shift among Republicans. The paper thus pits intergroup contact theory against motivated reasoning, concluding that exposure-based interventions may be ineffective or counterproductive.
Key Contributions
- Experimental, real-world evidence of asymmetric partisan backfire from prolonged cross-partisan exposure on social media.
- Methodological advance in computational social science: combining survey research, bot technology, and digital trace data with behaviorally verified compliance measures.
- A bot-based intervention that disrupts selective exposure over a month in a naturalistic setting, rather than a one-shot lab or survey manipulation.
- Practical caution for depolarization efforts — exposure to the “other side” may deepen, not bridge, partisan divides.
Methods
- Preregistered field experiment run separately for Democrats (n=901) and Republicans (n=751) who use Twitter at least three times weekly, recruited via a professional survey firm (1,652 pretreatment respondents) through an “ostensibly unrelated” design.
- Block randomization stratified by party attachment, interest in current events, and Twitter use frequency.
- Treatment: participants paid $11 to follow a bot retweeting 24 messages/day for one month from opposing-ideology accounts, sampled from 4,176 political accounts scored on a liberal–conservative dimension via correspondence analysis of following patterns.
- Ideology measured with a 10-item, seven-point policy attitude scale (α = .91) pre- and posttreatment.
- Compliance monitored through weekly incentivized surveys (up to $18) including content questions and identification of a daily animal picture, defining three compliance tiers.
- Analysis used multivariate models predicting posttreatment ideology controlling for pretreatment score and 12 covariates, reporting Intent-to-Treat (ITT) and Complier Average Causal Effects (CACE).
Findings
- Treated Republicans became significantly more conservative (ITT = 0.12, p = 0.008), with effects rising by compliance level.
- Fully compliant Republicans shifted 0.60 points more conservative (CACE, p < 0.01), roughly 0.11–0.59 SD.
- Treated Democrats trended slightly more liberal with higher compliance, but no effect reached statistical significance.
- Compliance was moderate: ~65% of Democrats and ~57% of Republicans accepted the bot; ~62% answered all weekly content questions.
- Supplementary analyses rule out Hawthorne effects, partisan learning, message-extremity variation, and age-based differences as drivers.
Connections
This is a foundational counterpoint to the echo-chamber literature, which often assumes exposure diversity is corrective; it complements empirical work mapping actual cross-cutting exposure and its limits Bakshy2015-rn, Barbera2015-fw, Eady2023-xg, and studies of platform-driven exposure and ideological segregation Gonzalez-Bailon2023-uy, Gonzalez-Bailon2024-rq. Its asymmetric-polarization finding speaks directly to work on the affective and elite dimensions of partisan animosity Iyengar2019-jj and to debates on whether social media meaningfully alters attitudes Guess2023-ai, Allcott2025-jb. The motivated-reasoning mechanism it invokes also links it to research on backfire and belief updating relevant to misinformation correction Nyhan2023-gb, Pennycook2021-jq.