In addressing various risks on social media, the HCI community has advocated for teen-centered risk detection technologies over platform-based, parent-centered features. However, their real-world viability remains underexplored by secondary stakeholders beyond the family unit. Therefore, we present an evaluation of a teen-centered social media risk detection dashboard through online interviews with 33 online safety experts. While experts praised our dashboard’s clear design for teen agency, their feedback revealed five primary tensions in implementing and sustaining such technology: objective vs. context-dependent risk definition, informing risks vs. meaningful intervention, teen empowerment vs. motivation, need for data vs. data privacy, and independence vs. sustainability. These findings motivate us to rethink “teen-centered” and a shift from a “fail fast” to a “mature safely” paradigm for youth safety technology innovation. We offer design implications for addressing these tensions before system deployment with teens and strategies for aligning secondary stakeholders’ interests to deploy and sustain such technologies in the broader ecosystem of youth online safety.

Ensuring the online safety of youth has motivated research towards the development of machine learning (ML) methods capable of accurately detecting social media risks after-the-fact. However, for these detection models to be effective, they must proactively identify high-risk scenarios (e.g., sexual solicitations, cyberbullying) to mitigate harm. This `real-time' responsiveness is a recognized challenge within the risk detection literature. Therefore, this paper presents a novel two-level framework that first uses reinforcement learning to identify conversation stop points to prioritize messages for evaluation. Then, we optimize state-of-the-art deep learning models to accurately categorize risk priority (low, high). We apply this framework to a time-based simulation using a rich dataset of 23K private conversations with over 7 million messages donated by 194 youth (ages 13-21). We conducted an experiment comparing our new approach to a traditional conversation-level baseline. We found that the timeliness of conversations significantly improved from over 2 hours to approximately 16 minutes with only a slight reduction in accuracy (0.88 to 0.84). This study advances real-time detection approaches for social media data and provides a benchmark for future training reinforcement learning that prioritizes the timeliness of classifying high-risk conversations.

While risk is highly subjective, especially when it comes to the private online interactions of youth, third-party annotations are often performed to identify risky content. Therefore, we conducted a mixed-methods study to examine if, how, and why risk perceptions might differ between youth and third-party annotators who were research assistants (RAs). We first asked 100 youth to share their Instagram private messages and flag media that made them feel unsafe. Then, we had RAs annotate the same media to identify what they thought was unsafe or risky. Compared to RAs, youth tended to flag images as risky when they perceived targeted harassment towards them or unwanted solicitations from strangers. In contrast, RAs were more likely to risk-flag sexual images with a humorous undertone shared among friends. Our findings highlight the differences between how online risks are perceived by youth compared to RAs. We provide recommendations for assessing online risks based on multiple perspectives to inform future youth-centered risk mitigation approaches.

Recent increases in self-harm and suicide rates among youth have coincided with prevalent social media use; therefore, making these sensitive topics of critical importance to the HCI research community. We analyzed 1,224 direct message conversations (DMs) from 151 young Instagram users (ages 13-21), who engaged in private conversations using self-harm and suicide-related language. We found that youth discussed their personal experiences, including imminent thoughts of suicide and/or self-harm, as well as their past attempts and recovery. They gossiped about others, including complaining about triggering content and coercive threats of self-harm and suicide but also tried to intervene when a friend was in danger. Most of the conversations involved suicide or self-harm language that did not indicate the intent to harm but instead used hyperbolical language or humor. Our results shed light on youth perceptions, norms, and experiences of self-harm and suicide to inform future efforts towards risk detection and prevention.

Accurate real-time risk identification is vital to protecting social media users from online harm, which has driven research towards advancements in machine learning (ML). While strides have been made regarding the computational facets of algorithms for “real-time” risk detection, such research has not yet evaluated these advancements through a human-centered lens. To this end, we conducted a systematic literature review of 53 peer-reviewed articles on real-time risk detection on social media. Real-time detection was mainly operationalized as “early” detection after-the-fact based on pre-defined chunks of data and evaluated based on standard performance metrics, such as timeliness. We identified several human-centered opportunities for advancing current algorithms, such as integrating human insight in feature selection, algorithms’ improvement considering human behavior, and utilizing human evaluations. This work serves as a critical call-to-action for the HCI and ML communities to work together to protect social media users before, during, and after exposure to risks.

Online harassment negatively impacts mental health, with victims expressing increased concerns such as depression, anxiety, and even increased risk of suicide, especially among youth and young adults. Yet, research has mainly focused on building automated systems to detect harassment incidents based on publicly available social media trace data, overlooking the impact of these negative events on the victims, especially in private channels of communication. Looking to close this gap, we examine a large dataset of private message conversations from Instagram shared and annotated by youth aged 13-21. We apply trained classifiers from online mental health to analyze the impact of online harassment on indicators pertinent to mental health expressions. Through a robust causal inference design involving a difference-in-differences analysis, we show that harassment results in greater expression of mental health concerns in victims up to 14 days following the incidents, while controlling for time, seasonality, and topic of conversation. Our study provides new benchmarks to quantify how victims perceive online harassment in the immediate aftermath of when it occurs. We make social justice-centered design recommendations to support harassment victims in private networked spaces. We caution that some of the paper's content could be triggering to readers.

The debate on whether social media has a net positive or negative effect on youth is ongoing. Therefore, we conducted a thematic analysis on 2,061 posts made by 1,038 adolescents aged 15-17 on an online peer-support platform to investigate the ways in which these teens discussed popular social media platforms in their posts and to identify differences in their experiences across platforms. Our findings revealed four main emergent themes for the ways in which social media was discussed: 1) Sharing negative experiences or outcomes of social media use (58%, n = 1,095), 2) Attempts to connect with others (45%, n = 922), 3) Highlighting the positive side of social media use (20%, n = 409), and 4) Seeking information (20%, n = 491). Overall, while sharing about negative experiences was more prominent, teens also discussed balanced perspectives of connection-seeking, positive experiences, and information support on social media that should not be discounted. Moreover, we found statistical significance for how these experiences differed across social media platforms. For instance, teens were most likely to seek romantic relationships on Snapchat and self-promote on YouTube. Meanwhile, Instagram was mentioned most frequently for body shaming, and Facebook was the most commonly discussed platform for privacy violations (mostly from parents). The key takeaway from our study is that the benefits and drawbacks of teens' social media usage can co-exist and net effects (positive or negative) can vary across different teens across various contexts. As such, we advocate for mitigating the negative experiences and outcomes of social media use as voiced by teens, to improve, rather than limit or restrict, their overall social media experience. We do this by taking an affordance perspective that aims to promote the digital well-being and online safety of youth "by design."

We conducted a study with 173 adolescents (ages 13-21), who self-reported their offline and online risk experiences and uploaded their Instagram data to our study website to flag private conversations as unsafe. Risk profiles were first created based on the survey data and then compared with the risk-flagged social media data. Five risk profiles emerged: Low Risks (51% of the participants), Medium Risks (29%), Increased Sexting (8%), Increased Self-Harm (8%), and High Risk Perpetration (4%). Overall, the profiles correlated well with the social media data with the highest level of risk occurring in the three smallest profiles. Youth who experienced increased sexting and self-harm frequently reported engaging in unsafe sexual conversations. Meanwhile, high risk perpetration was characterized by increased violence, threats, and sales/promotion of illegal activities. A key insight from our study was that offline risk behavior sometimes manifested differently in online contexts (i.e., offline self-harm as risky online sexual interactions). Our findings highlight the need for targeted risk prevention strategies for youth online safety.

Research involving sensitive data often leads to valuable human-centered insights. Yet, the effects of participating in and conducting research about sensitive data with youth are poorly understood. We conducted meta-level research to improve our understanding of these effects. We did the following: (i) asked youth (aged 13-21) to share their private Instagram Direct Messages (DMs) and flag their unsafe DMs; (ii) interviewed 30 participants about the experience of reflecting on this sensitive data; (iii) interviewed research assistants (RAs, n=12) about their experience analyzing youth's data. We found that reflecting about DMs brought discomfort for participants and RAs, although both benefited from increasing their awareness about online risks, their behavior, and privacy and social media practices. Participants had high expectations for safeguarding their private data while their concerns were mitigated by the potential to improve online safety. We provide implications for ethical research practices and the development of reflective practices among participants and RAs through applying trauma-informed principles to HCI research.

Instagram, one of the most popular social media platforms among youth, has recently come under scrutiny for potentially being harmful to the safety and well-being of our younger generations. Automated approaches for risk detection may be one way to help mitigate some of these risks if such algorithms are both accurate and contextual to the types of online harms youth face on social media platforms. However, the imminent switch by Instagram to end-to-end encryption for private conversations will limit the type of data that will be available to the platform to detect and mitigate such risks. In this paper, we investigate which indicators are most helpful in automatically detecting risk in Instagram private conversations, with an eye on high-level metadata, which will still be available in the scenario of end-to-end encryption. Toward this end, we collected Instagram data from 172 youth (ages 13-21) and asked them to identify private message conversations that made them feel uncomfortable or unsafe. Our participants risk-flagged 28,725 conversations that contained 4,181,970 direct messages, including textual posts and images. Based on this rich and multimodal dataset, we tested multiple feature sets (metadata, linguistic cues, and image features) and trained classifiers to detect risky conversations. Overall, we found that the metadata features (e.g., conversation length, a proxy for participant engagement) were the best predictors of risky conversations. However, for distinguishing between risk types, the different linguistic and media cues were the best predictors. Based on our findings, we provide design implications for AI risk detection systems in the presence of end-to-end encryption. More broadly, our work contributes to the literature on adolescent online safety by moving toward more robust solutions for risk detection that directly takes into account the lived risk experiences of youth.

We collected Instagram data from 150 adolescents (ages 13-21) that included 15,547 private message conversations of which 326 conversations were flagged as sexually risky by participants. Based on this data, we leveraged a human-centered machine learning approach to create sexual risk detection classifiers for youth social media conversations. Our Convolutional Neural Network (CNN) and Random Forest models outperformed in identifying sexual risks at the conversation-level (AUC=0.88), and CNN outperformed at the message-level (AUC=0.85). We also trained classifiers to detect the severity risk level (i.e., safe, low, medium-high) of a given message with CNN outperforming other models (AUC=0.88). A feature analysis yielded deeper insights into patterns found within sexually safe versus unsafe conversations. We found that contextual features (e.g., age, gender, and relationship type) and Linguistic Inquiry and Word Count (LIWC) contributed the most for accurately detecting sexual conversations that made youth feel uncomfortable or unsafe. Our analysis provides insights into the important factors and contextual features that enhance automated detection of sexual risks within youths' private conversations. As such, we make valuable contributions to the computational risk detection and adolescent online safety literature through our human-centered approach of collecting and ground truth coding private social media conversations of youth for the purpose of risk classification.

We collected Instagram Direct Messages (DMs) from 100 adolescents and young adults (ages 13-21) who then flagged their own conversations as safe or unsafe. We performed a mixed-method analysis of the media files shared privately in these conversations to gain human-centered insights into the risky interactions experienced by youth. Unsafe conversations ranged from unwanted sexual solicitations to mental health related concerns, and images shared in unsafe conversations tended to be of people and convey negative emotions, while those shared in regular conversations more often conveyed positive emotions and contained objects. Further, unsafe conversations were significantly shorter, suggesting that youth disengaged when they felt unsafe. Our work uncovers salient characteristics of safe and unsafe media shared in private conversations and provides the foundation to develop automated systems for online risk detection and mitigation.

We conducted a user study with 380 Android users, profiling them according to two key privacy behaviors: the number of apps installed and the Dangerous permissions granted to those apps. We identified four unique privacy profiles: 1) Privacy Balancers (49.74% of participants), 2) Permission Limiters (28.68%), 3) App Limiters (14.74%), and 4) the Privacy Unconcerned (6.84%). App and Permission Limiters were significantly more concerned about perceived surveillance than Privacy Balancers and the Privacy Unconcerned. App Limiters had the lowest number of apps installed on their devices with the lowest intention of using apps and sharing information with them, compared to Permission Limiters who had the highest number of apps installed and reported higher intention to share information with apps. The four profiles reflect the differing privacy management strategies, perceptions, and intentions of Android users that go beyond the binary decision to share or withhold information via mobile apps.

Sexual exploration is a natural part of adolescent development; yet, unmediated internet access has enabled teens to engage in a wider variety of potentially riskier sexual interactions than previous generations, from normatively appropriate sexual interactions to sexually abusive situations. Teens have turned to online peer support platforms to disclose and seek support about these experiences. Therefore, we analyzed posts (N=45,955) made by adolescents (ages 13--17) on an online peer support platform to deeply examine their online sexual risk experiences. By applying a mixed methods approach, we 1) accurately (average of AUC = 0.90) identified posts that contained teen disclosures about online sexual risk experiences and classified the posts based on level of consent (i.e., consensual, non-consensual, sexual abuse) and relationship type (i.e., stranger, dating/friend, family) between the teen and the person in which they shared the sexual experience, 2) detected statistically significant differences in the proportions of posts based on these dimensions, and 3) further unpacked the nuance in how these online sexual risk experiences were typically characterized in the posts. Teens were significantly more likely to engage in consensual sexting with friends/dating partners; unwanted solicitations were more likely from strangers and sexual abuse was more likely when a family member was involved. We contribute to the HCI and CSCW literature around youth online sexual risk experiences by moving beyond the false dichotomy of "safe" versus "risky". Our work provides a deeper understanding of technology-mediated adolescent sexual behaviors from the perspectives of sexual well-being, risk detection, and the prevention of online sexual violence toward youth.

In the era of big data and artificial intelligence, online risk detection has become a popular research topic. From detecting online harassment to the sexual predation of youth, the state-of-the-art in computational risk detection has the potential to protect particularly vulnerable populations from online victimization. Yet, this is a high-risk, high-reward endeavor that requires a systematic and human-centered approach to synthesize disparate bodies of research across different application domains, so that we can identify best practices, potential gaps, and set a strategic research agenda for leveraging these approaches in a way that betters society. Therefore, we conducted a comprehensive literature review to analyze 73 peer-reviewed articles on computational approaches utilizing text or meta-data/multimedia for online sexual risk detection. We identified sexual grooming (75%), sex trafficking (12%), and sexual harassment and/or abuse (12%) as the three types of sexual risk detection present in the extant literature. Furthermore, we found that the majority (93%) of this work has focused on identifying sexual predators after-the-fact, rather than taking more nuanced approaches to identify potential victims and problematic patterns that could be used to prevent victimization before it occurs. Many studies rely on public datasets (82%) and third-party annotators (33%) to establish ground truth and train their algorithms. Finally, the majority of this work (78%) mostly focused on algorithmic performance evaluation of their model and rarely (4%) evaluate these systems with real users. Thus, we urge computational risk detection researchers to integrate more human-centered approaches to both developing and evaluating sexual risk detection algorithms to ensure the broader societal impacts of this important work.