(Image by StockSmartStart on Shutterstock)
Restorative Replies Keep Social Media Healthier, Cornell Study Finds
In A Nutshell
- Restorative replies, which appeal to values and invite apologies, are judged fairer than insults or threats.
- People support restorative responses with more upvotes and fewer flags compared to retributive ones.
- Offenders seen as unwilling to change weaken the appeal of restorative approaches, but kindness still holds ground.
- When restorative voices appear first, they set a constructive tone and reduce support for harsh replies.
ITHACA, N.Y. — Someone posts a racist comment on a viral video. Within minutes, replies pile up. Some lash out with insults and threats. Others urge the poster to reflect and consider the harm. A Cornell University study tested which type of response actually helps online communities thrive, and the results challenge the instinct to punish.
Researchers found that restorative responses, which appeal to shared values and invite apologies, not only win more support but also create digital spaces where people feel more satisfied and more likely to stay engaged. Retributive replies, built on insults or threats, may feel natural, but they often fuel the very hostility they are meant to stop.
How Researchers Tested Online Conflict
To measure real reactions, the Cornell team built a simulated platform called VidShare. It looked and felt like a video-sharing site, and participants thought they were beta-testing a new app. In reality, every video, comment, and reply was scripted. More than 3,500 participants encountered an offensive comment aimed at a video creator. Depending on their assigned condition, they saw a retributive reply, a restorative reply, or sometimes both in sequence. In later tests, some participants also saw a neutral reply that acknowledged the comment without addressing the harm.
Researchers measured how participants actually behaved. Did they upvote a reply? Did they downvote or flag it as inappropriate? These simple actions matter because they reflect how users shape what rises and what sinks on real social media feeds.
Why Restorative Replies Outperform Punishment
Across both experiments, restorative responses consistently stood out. Participants judged them to better serve justice than retributive ones. They were more likely to support restorative replies with upvotes and less likely to punish them with flags or downvotes. Retributive replies, while not meaningless, often drew pushback.
Exposure to restorative replies also shaped how people felt about the community itself. When participants believed justice had been served in a fair and constructive way, they reported higher satisfaction and a greater willingness to engage further. Retribution still had some legitimacy, especially compared to ignoring harm altogether, but it carried a risk of prolonging hostility rather than resolving it.
When Revenge Feels Justified In Trolling Situations
The researchers tested whether perceptions of the offender influenced the effectiveness of each approach. They introduced the idea of moral corrigibility, which refers to whether someone seems capable of learning from their mistakes. When offenders were perceived as corrigible, restorative replies were strongly favored. When offenders were seen as beyond change, the preference for restorative replies weakened and retribution seemed just as legitimate.
This finding helps explain why social media often skews toward harsh replies. Many users assume strangers online are unlikely to change, and so they default to punishment. Interestingly, the study also found that when no restorative voice was present, retribution gained more approval. In other words, the absence of constructive alternatives can make punitive responses appear justified by comparison.
The second experiment tested what happens when restorative and retributive replies appear together. Sequence shaped perception. If a retributive reply came first, people judged it less favorably once a restorative reply appeared after it. But when a restorative reply appeared first, it maintained its strength even if a retributive one followed. This suggests that setting a constructive tone early can prevent conflict from escalating. Even one restorative reply, if introduced quickly, can make later retributive replies look less appealing and unnecessary.
Building Healthier Online Communities
The research highlights a core problem in digital spaces. Online vigilante justice often spirals into hostility. Harsh responses may seem like justice, but they frequently trigger more attacks and create an environment where people feel unwelcome. By contrast, restorative replies provide a path that acknowledges harm while aiming to repair it.
The findings can also help developers reconsider platform design. Social media companies could build in features that encourage users to pause before posting harsh replies. Algorithms could be tuned to surface constructive responses that appeal to shared values and invite accountability. Even small shifts, such as making restorative options visible earlier in a thread, could help change the culture of online conversations.
While kindness and restorative justice will not always be the right fit, especially when offenders appear unwilling to change, the study shows that most participants recognize its value. Encouraging restorative replies can help transform online exchanges from toxic trolling cycles into constructive dialogues.
Paper Summary
Methodology
Researchers conducted two controlled experiments using VidShare, a simulated social media platform designed to look like TikTok or YouTube Shorts. A total of 3,516 participants were recruited through CloudResearch and believed they were beta-testing a new video platform. In reality, all content and user interactions were pre-programmed by researchers. Participants were exposed to offensive comments followed by either retributive responses (threatening or shaming) or restorative responses (appealing to values and requesting apologies). The researchers tracked real behavioral responses like upvotes, downvotes, and flags, then measured participants’ perceptions of justice, community satisfaction, and future engagement intentions through surveys.
Results
Restorative responses consistently outperformed retributive ones across multiple measures. Participants rated restorative responses as more effective at serving justice and were 2.35 to 2.43 times more likely to upvote them compared to retributive responses. They were also significantly less likely to downvote or flag restorative responses. Participants exposed to restorative responses reported higher community satisfaction and greater intentions for future platform engagement. However, when offenders were perceived as morally incorrigible (unable to change), the preference for restorative over retributive responses diminished. When multiple responses appeared together, restorative responses maintained their effectiveness regardless of order, while retributive responses were viewed less favorably when they appeared after restorative ones.
Limitations
The study used a simulated environment that doesn’t fully replicate the networked interactions of real social media platforms. Each justice framework was represented by carefully crafted exemplar messages that cannot capture the full range of these approaches. The research focused primarily on immediate behavioral responses rather than long-term community effects. Future research should examine how these interventions perform in naturalistic settings and investigate their sustained impact on reducing online toxicity over time.
Funding and Disclosures
This research was funded by the National Science Foundation under grant number 2106476. The authors declared no competing interests. The study was approved by the Institutional Review Board at Cornell University, and all participants provided informed consent.
Publication Information
Zhao, P., Bazarova, N. N., Bae, I., Hui, W., Kizilcec, R. F., & Margolin, D. (2025). Restorative justice appeals trump retributive vigilance on social media. PNAS Nexus, 4, pgaf255. https://doi.org/10.1093/pnasnexus/pgaf255
