HMN 2026: How ‘Diversifying’ social feeds can cut exposure to toxic content and preserve enjoyment

A new study from Northwestern University and the University of Chicago offers underlying evidence that the engagement-based algorithms used by major social media platforms amplify intergroup, moralized, emotional (IME) and toxic political content—and that a relatively simple redesign can reduce that amplification without hurting users’ overall experience of the platform.

The study, published in Nature, was led by William Brady, assistant professor of Management and Organizations at Northwestern’s Kellogg School of Management.

On Bluesky Social over the eight weeks surrounding the 2024 U.S. presidential election, researchers found that engagement-based algorithms systematically amplified contentious content, increasing moral outrage and political content and reducing users’ ability to perceive social norms. In comparison, diversified extremity algorithm and reverse-chronological feeds fostered a positive platform experience and exposed users to less contentious content.

“For years, the debate over algorithmic amplification has been hampered by the fact that independent researchers don’t have full control over the algorithms they’re studying,” said Brady. “Bluesky’s open architecture let us build the algorithms ourselves, run a real-world experiment at scale, and measure what these systems actually do to political conversation.”

The experiment

The researchers recruited 2,000 U.S. citizens who identified as Democrats or Republicans and were active Bluesky users during the 2024 presidential election. Over an eight-week period, participants were randomly assigned to view content through one of three custom-built feed-ranking algorithms:

• An engagement-based algorithm, designed to mirror the kind of feed-ranking used on major platforms like X and Meta, which surfaces content predicted to maximize user engagement
• A reverse-chronological feed, the control condition, which simply shows the most recent posts first
• A “diversified extremity” algorithm, designed by the research team to reduce the outsized influence of extreme users by downranking super-posters, lowering the probability of toxic posts and increasing the probability of constructive dialogue

The team analyzed roughly 20 million posts over two months and surveyed participants weekly about their perceptions of online dialogue, partisan animosity and their experience using the platform.

“This setup allowed us to actually test theories in recent work concerning algorithmic amplification: Do social media algorithms that exploit human attention actually amplify IME content on these platforms?” Brady said.

Perception becomes reality

Compared with the reverse-chronological control, the engagement-based feed amplified IME and toxic political content—with effects that grew larger after the election. The largest amplifications were in moral outrage and political content, which increased by roughly 37% before the election and nearly 80% after it, relative to the reverse-chronological baseline.

“These algorithms are doing exactly what critics have long argued: They’re selectively pushing content that grabs attention by appealing to outrage, moral conflict and negative emotion,” Brady said.

The engagement-based feed also significantly increased perceptions of partisan animosity. According to Brady, users came to see their network as more hostile to the political outgroup, particularly after the election. It reduced the accuracy of users’ perceptions of social norms about toxic political dialogue—though in a direction the researchers did not anticipate. “Users actually underestimated how inappropriate others found toxic posts,” Brady said.

Notably, despite these shifts in perception, the engagement-based feed did not significantly change participants’ own posting, liking or sharing behavior compared with the other conditions.

According to researchers, the diversified extremity algorithm consistently reduced exposure to toxic and morally outraged content relative to the engagement-based feed, and improved the accuracy of users’ perceptions of how appropriate it was to post toxic content.

It also revealed a finding worth noting: Users in the diversified condition reported greater overall enjoyment of Bluesky as a platform after the election, compared with users on the engagement-based algorithm.

“There’s a common assumption that reducing toxic content will inevitably hurt the user experience because we like to click on it,” Brady said. “Our results push back on that. Limiting the influence of a small number of extreme users—who account for a disproportionate share of toxic posts—can meaningfully reduce the toxicity people encounter while keeping the overall platform experience comparable, and in some respects better.”

The researchers note that the study has limitations: Bluesky has a left-leaning user base, the experimental window was only two months and behavioral effects on individual engagement were modest. But the findings speak directly to ongoing policy debates about algorithmic transparency and accountability.

“Engagement-based ranking shapes the informational environment more reliably than it shapes individual actions, at least over a two-month window,” Brady said. “That still matters because the informational environment is where perceptions of our political opponents, and of what’s socially acceptable, are formed.”

He added, “Our study suggests there’s room for platforms to design feeds that reduce distortions in political discourse without necessarily sacrificing a user’s platform experience. That’s a more optimistic picture than the usual framing of an inevitable trade-off between healthy discourse and engagement.”

The content is provided for information purposes only.