Paper Breakdown: The Feed Shapes What We Think Others Think

Executive summary

• The paper: Brady, Doyle, Elnakouri and colleagues (2026), "Redesigning algorithms to intervene on social norm misperceptions during a national election," Nature.
• The question: do engagement-based feeds distort how people perceive political norms, and can a different ranking rule correct that?
• The answer: yes on both counts. Engagement ranking amplified outrage and toxic content and made users misjudge what others consider acceptable. An alternative feed reduced both effects without making the platform less enjoyable.
• The lesson for leaders: algorithms shape the perceived climate more reliably than individual behavior, and the perceived climate is what drives decisions in a crisis.

Why this study matters

Last week, in our notes on the spiral of silence, we discussed how people read the "opinion climate" before deciding whether to speak. This paper, led by William Brady at Northwestern's Kellogg School of Management, tests a direct question that follows: what happens when the climate people read is curated by a ranking algorithm built to maximize engagement?

Most previous work on this question faced a major obstacle. Researchers rarely control the feed. They observe outcomes on platforms whose algorithms are proprietary. Brady and his colleagues solved that problem by building the algorithms themselves.

The design

Setting and sample

• Platform: Bluesky, whose open architecture allows third parties to build and serve custom feeds.
• Participants: 2,000 active U.S. users who identified as Democrats or Republicans.
• Period: eight weeks around the November 2024 presidential election, covering the weeks before and after the vote.
• Scale: roughly 20 million posts analyzed.

Three experimental feeds

1. Reverse-chronological: the control condition. Posts appear in the order they were published.
2. Engagement-based: a reconstruction of the ranking logic used by large platforms, which promotes content most likely to attract interaction.
3. Diversified extremity: a new design that reduces the outsized weight of "super-posters" and lowers the likelihood of toxic posts, while raising the likelihood of constructive exchanges.

Measurement

The researchers combined two layers of data. The first was content: what each feed actually showed, classified by whether it was intergroup, moralized, emotional or toxic. The second was perception: weekly surveys on how participants saw online dialogue, how much animosity they perceived between parties, how acceptable they believed others found toxic posting, and how much they enjoyed the platform.

The findings

• Engagement ranking amplified outrage. Compared with the chronological control, the engagement feed increased moral outrage and political content by about 37% before the election and nearly 80% after it.
• It distorted perceived norms. Users on the engagement feed underestimated how inappropriate others found toxic posts, and they perceived more partisan animosity, especially after the election.
• It did not change behavior. Participants' own posting, liking and sharing remained largely stable across conditions.
• The alternative design worked. The diversified extremity feed reduced exposure to toxic and outraged content, improved the accuracy of norm perceptions, and after the election users reported enjoying the platform more than with the engagement feed.

Brady summarized the core result in a sentence worth remembering: engagement-based ranking shapes the informational environment more reliably than it shapes individual actions.

The theoretical frame

The study sits at the intersection of two lines of research.

The first is Brady's own work on moral contagion. The MAD model (Brady, Crockett & Van Bavel, 2020) argues that moral and emotional content captures attention and spreads more readily online, and that platform design can amplify that tendency. The new experiment moves from describing that dynamic to manipulating it.

The second is the classic literature on pluralistic ignorance. Prentice and Miller (1993) showed that students systematically overestimated how comfortable their peers were with heavy drinking, and adjusted their own attitudes toward that imagined norm. The Bluesky experiment documents a digital version of the same error: a feed that over-represents the most extreme voices teaches users that extremity is more accepted than it is.

Limitations worth noting

The authors acknowledge several constraints:

• Bluesky's user base leans left, which limits generalization to the broader population.
• The experimental window lasted only two months.
• Effects on individual behavior were modest.

There is also a practical caveat. Bluesky's open design made the experiment possible. The platforms where most political conversation happens do not allow outside researchers to replace their ranking algorithms.

Lessons for senior leadership

1. Do not read the feed as a poll. What appears in your timeline, or your organization's monitoring dashboard, is filtered by ranking systems that favor outrage. Treat it as a signal, not a census.
2. Correct the perceived norm, not just the message. In a crisis, supporters may stay silent because they believe hostility is the majority view. Making the real distribution of opinion visible can be more effective than another rebuttal.
3. Separate volume from prevalence. A small number of highly active accounts can dominate what people see. Measure how many distinct people are speaking, not just how many posts exist.
4. Expect the effect to peak after the vote. The amplification nearly doubled after the election. Post-election periods and crisis aftermaths deserve more monitoring, not less.

The study's most useful message for leaders is a modest one. Algorithms may not change what people do, but they change what people believe everyone else thinks. In politics and in crisis management, that belief is often what decides the next move.

References

Brady, W. J., Crockett, M. J., & Van Bavel, J. J. (2020). The MAD model of moral contagion: The role of motivation, attention, and design in the spread of moralized content online. Perspectives on Psychological Science, 15(4), 978–1010. https://doi.org/10.1177/1745691620917336

Brady, W. J., Doyle, M., Elnakouri, A., et al. (2026). Redesigning algorithms to intervene on social norm misperceptions during a national election. Nature. https://doi.org/10.1038/s41586-026-10536-1

Harris, S. (2026, May 27). 'Diversifying' social feeds cuts exposure to toxic content, preserves enjoyment. Northwestern Now. https://news.northwestern.edu/stories/2026/05/researchers-found-that-diversifying-social-feeds-cut-exposure-to-toxic-content-while-preserving-platform-enjoyment

Prentice, D. A., & Miller, D. T. (1993). Pluralistic ignorance and alcohol use on campus: Some consequences of misperceiving the social norm. Journal of Personality and Social Psychology, 64(2), 243–256. https://doi.org/10.1037/0022-3514.64.2.243