← Back to Blog

AI Unleashed: Revolutionizing Cyberbullying Detection for the PAN SafeKids App

Published by Manuel Jenni on Mon, May 15, 2023


In today's interconnected world, the rise of cyberbullying has become a significant concern. Defined as the intentional harassment of a specific person in cyberspace over an extended period of time, cyberbullying can have detrimental effects on individuals, particularly children and adolescents. As technology advances, it becomes imperative to develop effective tools to detect and prevent cyberbullying, ensuring a safer online environment for everyone.

Understanding Cyberbullying

Cyberbullying can take on different forms, such as offensive comments, voice messages, photos, and offensive jokes. However, it is crucial to distinguish between the presence of cyberbullying indicators and actual instances of cyberbullying. While a single negative comment may serve as an indicator, it only qualifies as cyberbullying if it is repetitive and specifically directed at the victim.

Challenges in Cyberbullying Detection

Detecting cyberbullying is a complex task that goes beyond traditional sentiment analysis. It requires considering not only textual features but also temporal features, interactions, relationships between users, emotions, and more. To address this challenge, advanced AI technologies are employed in SafeKids to analyze historical data from user sessions and compute various features, such as repetitiveness scores, to determine an overall cyberbullying score.

The Role of AI in Cyberbullying Detection

AI-powered systems, like the SafeKids app, utilize advanced algorithms and machine learning models to detect cyberbullying incidents. By analyzing multiple modalities, such as text, images, and user interactions, these systems can identify patterns and correlations that indicate instances of cyberbullying. Textual analysis plays a crucial role in understanding the content of messages, while temporal analysis helps identify persistent and repetitive behaviors over time.

AI in SafeKids App: A Closer Look

In the quest to combat cyberbullying and protect children in the digital realm, SafeKids has implemented advanced Natural Language Processing (NLP) models as the backbone of its AI-powered detection system. These sophisticated models employ a range of features to identify and analyze various aspects of cyberbullying incidents. Let's delve deeper into the specific implementation of these features in SafeKids:

  1. Repetitiveness Score: Research in the field of cyberbullying highlights the importance of repetitiveness in distinguishing cyberbullying traces from messages with negative sentiment that may not qualify as cyberbullying. SafeKids analyzes historical data to compute the repetitiveness score, allowing for the differentiation of cyberbullying patterns that involve repetitive and offensive activity directed at the victim.

  2. Anger Score: This feature captures the correlation between cyberbullying incidents and the expression of anger in messages specifically targeting an individual (victim) within a cyberbullying session.

  3. Flirtation Score: While flirtation may not be directly associated with cyberbullying, SafeKids takes into account its low correlation with cyberbullying activity when assessing potential risks and harmful behaviors.

  4. Identity Attack Score: This score exhibits a high correlation with cyberbullying activity, although it is not always considered a direct cyberbullying trace. Its relevance stems from its association with the identity of the victim.

  5. Insult Score: Almost always directed against somebody, the insult score carries a high correlation with cyberbullying activity. However, it may not always target the victim within the discussion.

  6. Joy Score: Joyful comments, often positive in nature, display a low correlation with cyberbullying activity, as they are less likely to be indicative of harmful behaviors.

  7. Negative Sentiment Score: While negative comments alone may not constitute cyberbullying traces, they exhibit a high correlation with cyberbullying activity, as cyberbullying traces are predominantly negative in nature.

  8. Optimism Score: Optimistic activity typically bears no relation to cyberbullying, resulting in a low correlation with cyberbullying activity.

  9. Positive Score: Positive comments generally do not align with cyberbullying traces and exhibit a low correlation with cyberbullying activity.

  10. Profanity Score: This feature demonstrates a remarkably high correlation with cyberbullying activity, as profane language often accompanies harmful behaviors.

  11. Sadness Score: Emphasizing the strong connection between sadness and cyberbullying, this score exhibits a very high correlation with cyberbullying activity.

  12. Severe Toxicity Score: The severity of toxicity is a critical indicator of cyberbullying activity, thus showcasing a very high correlation with such incidents.

  13. Suicidal Score: Considered dangerous and alarming, the suicidal score demonstrates a very high correlation with cyberbullying activity, as it highlights potential harm directed towards the victim.

  14. Threat Score: Toxic, offensive messages carrying the threat score exhibit a very high correlation with cyberbullying activity, serving as a red flag for harmful behaviors.

  15. Toxicity Score: While some comments may be classified as toxic without being explicitly offensive, this score demonstrates a high correlation with cyberbullying activity, contributing to the overall assessment.

Privacy-Preserving Approach

Respecting the privacy of children is a fundamental aspect of SafeKids. Parents are not granted direct access to their child's social media accounts or the ability to read their messages or view their images. Instead, the app's AI technology ensures that the child's privacy is preserved while providing timely and context-specific guidance to address potential cyberbullying situations.


The rise of cyberbullying demands effective solutions that harness the power of AI technology. Through the analysis of textual and temporal features, as well as other modalities, AI-powered systems like SafeKids can help detect and prevent cyberbullying incidents, providing a safer online environment for children. By combining advanced algorithms, multi-modal analysis, and privacy-preserving approaches, we can empower parents and children to navigate the digital landscape with confidence, ensuring their well-being and fostering positive online experiences.

← Back to Blog
PAN SafeKids Logo

Technoparkstrasse 1 8005 Zürich - Switzerland