How Working in Groups Helps Republicans and Democrats Agree on Content Moderation Online
Online content moderation has become one of the most contentious challenges facing technology companies today. Social media platforms are under constant pressure to remove violent, hateful, or offensive material, yet there is often deep disagreement over what exactly crosses the line. A new academic study suggests there may be a surprisingly effective solution: having moderators work together in structured groups rather than alone.
According to data from the Pew Research Center, more than half of Americans believe technology companies should take active steps to restrict extremely violent content on their platforms. Despite this broad public support, real-world moderation decisions are far from straightforward. Even trained content moderators frequently disagree on how to classify content, especially when it involves hate speech, graphic imagery, or morally controversial material.
A new study conducted by Damon Centola, a professor at the Annenberg School for Communication, and Douglas Guilbeault, an assistant professor at Stanford University, explores why disagreement is so common — and how it can be reduced. Their findings were published in the Journal of Social Computing and provide compelling evidence that collaboration dramatically improves agreement, even among people with opposing political views.
Why Content Moderation Is So Difficult
Content moderation often involves moral judgment, not just rule enforcement. Images depicting bullying, domestic violence, armed conflict, or terrorism can be interpreted in multiple ways depending on a person’s values, experiences, and political worldview. This is one reason why moderation decisions can vary so widely, even among professionals.
Previous large-scale studies of social media platforms have shown something intriguing, though. Despite strong political polarization in society, Democrats and Republicans sometimes reach surprisingly high levels of agreement when categorizing content like fake news or hate speech. Centola and Guilbeault wanted to understand how this agreement happens and whether it could be intentionally reproduced.
Their hypothesis focused on a concept called structural synchronization.
What Is Structural Synchronization?
Structural synchronization refers to a process where patterns of interaction within social networks help filter out individual differences. When people repeatedly interact under structured conditions, their judgments begin to align, even if they start with very different perspectives.
In simple terms, the network itself shapes consensus. Instead of relying on personal instincts alone, participants adjust their decisions based on shared signals emerging from interaction. The researchers believed this mechanism could explain how politically divided groups sometimes end up agreeing on controversial issues.
Inside the Experiment
To test this idea, the researchers designed a large-scale experiment involving 620 participants, all of whom had prior experience as content moderators. The participant pool was politically diverse:
- 49.6% identified as Democrats
- 28.3% identified as Independents
- 20.7% identified as Republicans
- A small fraction identified with another political party
Participants were shown a curated set of controversial social media images. These images included depictions of interpersonal conflict such as bullying and domestic violence, as well as images involving militaristic violence, including armed conflict and terrorism. Participants were told the images came from Facebook posts, and that Facebook had asked for help deciding whether the images should be removed or allowed to remain online.
The participants were randomly assigned to one of two conditions.
The Individual Condition
In this condition, participants worked entirely alone. They reviewed images and decided whether each one violated content rules or should remain on the platform. This setup closely mirrors how moderation decisions are often handled in practice.
The Network Condition
In the second condition, participants worked in structured networks of 50 people. During each round, participants were paired with a partner and assigned one of two roles: speaker or hearer.
The speaker was shown three randomly selected images, with one image highlighted. The speaker had 30 seconds to assign a violation tag to the highlighted image or select the Do Not Remove option. The hearer’s task was to identify which image the speaker was describing based only on the tag provided.
Correct matches resulted in cash rewards, while incorrect matches caused both participants to lose money. After each round, participants were randomly paired with a new partner. Importantly, participants had no visibility into decisions made by anyone outside their immediate partner.
The Results Were Striking
The differences between the two conditions were dramatic.
In the individual condition, agreement was low. Participants agreed on whether images should remain on Facebook only 38% of the time by the end of the experiment. Political polarization was especially pronounced: Democrats and Republicans agreed on only about 30% of classifications.
In contrast, participants in the network condition achieved near-perfect agreement across all eight experimental networks. Even more notable was the reduction in political disagreement. Working in networks reduced partisan gaps by 23 percentage points, meaning Democrats and Republicans within the same networks ended up agreeing on most moderation decisions.
This level of agreement emerged without participants ever seeing the full network’s decisions, suggesting that the structured interaction itself — not persuasion or authority — drove the convergence.
Emotional Impact on Moderators
The study didn’t just measure accuracy and agreement. The researchers also examined emotional responses during the task.
Participants in the network condition reported significantly more positive feelings and lower emotional stress compared with those working alone. This finding is especially important because content moderation is known to be psychologically demanding, particularly when moderators are exposed to disturbing or violent material for extended periods.
The results suggest that collaboration doesn’t just improve decisions, it may also protect moderators’ mental well-being.
Why These Findings Matter
At a time when technology companies and policymakers are struggling to design fair, transparent, and consistent moderation systems, this study offers a promising path forward. Rather than relying solely on individual judgment or automated systems, platforms could benefit from structured collaborative moderation models.
The findings also challenge the assumption that political polarization makes consensus impossible. Under the right conditions, even deeply divided groups can align on difficult moral decisions.
The study was funded through Facebook’s Content Moderation Research Award, but Facebook had no role in designing the experiment, collecting data, or analyzing results.
Broader Implications for Online Platforms
While the study focused on image moderation, the concept of structural synchronization could apply more broadly. Similar network-based approaches might help with:
- Identifying misinformation and fake news
- Moderating harassment and hate speech
- Reviewing borderline or ambiguous content that automated systems struggle to classify
By reducing disagreement and emotional strain at the same time, group-based moderation could lead to more consistent and humane systems.
Final Thoughts
Content moderation is unlikely to ever be simple, but this research shows that how decisions are made matters just as much as who makes them. Structured collaboration appears to unlock a powerful mechanism that brings people together, even across political divides.
For an internet that feels increasingly fragmented, that’s a hopeful finding.
Research Paper:
https://doi.org/10.23919/jsc.2025.0024
