When sexual assault survivors choose to share their stories online, they often turn to anonymous platforms like Reddit.
The posts they write can be long, emotionally complex and deeply personal, and reach up to 40,000 characters per post. While the parameters of the platform offer freedom and anonymity, they also pose a challenge. The very length and intensity of the stories can make it difficult for “helpers,” fellow Reddit users who offer support and advice, to assist the posters effectively.
Vaibhav Garg, an affiliate faculty member at Virginia Tech’s Institute for Advanced Computing in Alexandria, looked to bridge that gap. He assembled a multidisciplinary team of collaborators: Hannah Javidi, a psychologist at Indiana University’s Department of Applied Health Science; Jiaqing Yuan, applied scientist at Amazon Web Services; Ruijie Xi, machine learning scientist at Meta; and Munindar Singh, alumni distinguished graduate professor of computer science at North Carolina State University.
Together, they developed an artificial intelligence (AI) model that highlights key parts of long-form Reddit posts about sexual violence, such as sexual harassment and assault, to make it easier for helpers to read, understand and respond to the victims with care.
“This project is very close to my heart because we can see its direct impact,” Garg said. “If we can use AI to make it easier for survivors to get the support they seek, that’s what social good really means.”
The project, recently accepted to the 2025 International AAAI Conference on Web and Social Media, is entitled “Analyzing Reddit Stories of Sexual Violence: Incidents, Effects, and Requests for Advice.”
Unlike traditional summarization models that condense stories into just a few lines or focus on one component, Garg’s AI tool preserves the original narrative while highlighting three specific sentence categories:
- Incident: The act of sexual violence, denoted in yellow in the project
- Effects: The emotional or psychological effects on the survivor, denoted in blue
- Advice-seeking: The specific advice the survivor is seeking, such as whether they should tell their parents, seek legal help, see a counselor, etc., denoted in green
Garg said the approach avoids reducing or erasing parts of a survivor’s story.
“Summarization can invalidate someone’s experience. By highlighting these three key sentence categories, we keep the story intact and make each part visible – no matter how short or emotionally difficult.”
Team members made careful, deliberate choices throughout the research process keeping in mind the sensitivity of the subject matter.
“We didn’t want to hire anyone else to read these stories. Some are traumatic. What if the person labeling the data has had a similar experience? That’s an ethical line we didn’t want to cross,” Garg said.
They avoided outsourcing any of the data annotation work, a common practice in AI research, because of such ethical concerns. Instead, Garg and his co-authors, who were each fully aware of the emotional demands of the work, manually labeled nearly 9,000 sentences. They consulted with Javidi to define and identify the three sentence types their AI model would be trained to detect.
“We were very conscious of our choices,” Garg said. “Even in our survey design, we made sure to use sensitive, trauma-informed language.”
To test the effectiveness of their AI model, team members conducted a survey in which Reddit helpers were asked to read both unmarked stories and stories enhanced with AI-generated highlights. The results were striking: 23 of 24 participants reported the highlighted stories were easier to understand and respond to. The tool reduced the perceived reading difficulty by nearly a full point on a four-point scale.
“This means the model actually helps. It grabs attention, helps the reader process the emotional and narrative content more easily, and ultimately enables better support,” Garg said.
While the model isn’t yet deployed live on Reddit, the team sees potential for integration into real-time systems. Garg envisions a future in which as soon as a survivor posts a story online, the AI model could automatically highlight the most important parts and make them more accessible to helpers.
Garg is already applying the same philosophy, called AI for social good, to a new project. He and two Virginia Tech master’s degree students are developing a model that analyzes posts in depression and anxiety forums. The aim is to determine whether the advice given by helpers actually resonates with individuals who are seeking support.
“This research is just the beginning. It can inspire others across the tech sector to build tools that ensure stories from vulnerable communities are heard, and responded to, with empathy,” Garg said
Share
Rebecca Barnabi
