UAlbany, Rutgers Researchers Develop Early-Warning Model to Predict Toxic Social Media Storms

By Bethany Bump

ALBANY, N.Y. (March 10, 2026) — Researchers at the University at Albany and Rutgers University have developed an early-warning framework that can predict harmful social media interactions before they erupt, paving the way for interventions that can minimize harm and make platforms safer for users.

Using publicly available datasets from Reddit and Instagram, two social media platforms with distinct conversation dynamics, researchers trained models to predict from just the first 10 comments whether a thread would escalate into “concentrated waves of toxic interactions” — or what they have dubbed a “negative storm” or “neg storm.”

These “neg storms” are distinct from isolated abusive remarks, said Pradeep Atrey, associate professor in the Department of Computer Science at UAlbany’s College of Nanotechnology, Science, and Engineering. Most research focuses on detecting whether an individual comment is toxic, but misses the situational dynamics that can cause a social media thread to escalate, he said.

“We like to give the analogy, would you rather detect a burning tree or a burning forest? In isolation, it could be a burning tree,” he said. “But if you look at the entire situation, it could be a burning forest.”

The early-warning framework could prove useful to social media platforms that are looking to prevent toxic social media situations from escalating, researchers said. Online moderators could be equipped with tools to slow or stop harmful interactions before they occur, they said.

Detecting a Burning Forest

Atrey collaborated with Irien Akter, a PhD student in Computer Science at UAlbany, and Vivek Singh, an associate professor of Library and Information Science at Rutgers University, on the model, which they unveiled in a paper presented at the IEEE International Symposium on Multimedia Conference in Italy this past December.

Their paper introduces a new metric, Comment Storm Severity (CSS), which quantifies how intensely toxicity concentrates on a social media thread within a short span of time, normalized against early baseline behavior. A “neg storm” develops once the CSS exceeds a predefined threshold.

Notably, their analysis found early signals within just the first 10 comments can indicate whether a social media thread is going to escalate into a toxic wave of interactions. Signals include comment timing and text pattern, said Akter, a second-year PhD student working in Atrey’s lab.

“When comments arrive quickly and start showing small toxic cues, that timing pattern is often more predictive than the actual words,” she said.

Most automated moderation in online communities treats toxicity as a problem to be solved one comment at a time, with per-comment models often trained on text devoid of any broader context. This can encode biases and over-flag language from marginalized dialect communities, researchers state.

Taking both timing of comments and content cues into account proved most accurate for detecting whether escalation would occur, they found. This is because harm often arises from the evolution of a thread over time, where sequences of replies can accumulate into a burst.

“To move from looking at trees to seeing forests, we argue for modeling situations, not isolated events,” researchers wrote.

Early Warning, Early Intervention

The findings have practical relevance for social media platforms.

By forecasting several trajectories from the earliest comments, platforms could shift to more proactive moderation strategies that anticipate and react to harm before it unfolds.

“The key is that if we can predict a toxic event before it occurs, we can implement safeguards to prevent it,” said Singh, who has been collaborating with Atrey and Akter.

Social media companies could integrate early-detection tools into their platforms that flag when a thread is predicted to escalate. At that point, moderators could apply friction such as rate limits, subtle background changes, warning nudges, or temporary slow-mode before storms fully form. Moderators could also route uncertain or high-risk cases for further review using calibrated probabilities.

Future research could delve into the network status of commenters — things like recent activity, post history and followers — to determine whether a thread is likely to escalate, Akter said.

“Social media is an integral part of our daily life nowadays,” she said. “Everyone wants to check what's going on around the world, to share their thoughts and pictures, to connect with families and with society. Our model aims to create a safe social media environment for everyone.”