Teens on social media need both protection and privacy – AI could help get the balance right

Much of adolescent online safety research is from self-reported data such as surveys. There’s a need for more investigation of young people’s real-world private interactions and their perspectives on online risks. To address this need, my colleagues and I collected a large dataset of young people’s Instagram activity, including more than 7 million direct messages. We asked young people to annotate their own conversations and identify the messages that made them feel uncomfortable or unsafe.

Using this dataset, we found that direct interactions can be crucial for young people seeking support on issues ranging from daily life to mental health concerns. Our finding suggests that these channels were used by young people to discuss their public interactions in more depth. Based on mutual trust in the settings, teens felt safe asking for help.

Research suggests that privacy of online discourse plays an important role in the online safety of young people, and at the same time a considerable amount of harmful interactions on these platforms comes in the form of private messages. Unsafe messages flagged by users in our dataset included harassment, sexual messages, sexual solicitation, nudity, pornography, hate speech and sale or promotion of illegal activities.

However, it has become more difficult for platforms to use automated technology to detect and prevent online risks for teens because the platforms have been pressured to protect user privacy. For example, Meta has implemented end-to-end encryption for all messages on its platforms to ensure message content is secure and only accessible by participants in conversations.

Also, the steps Meta has taken to block suicide and eating disorder content keep that content from public posts and search even if a teen’s friend has posted it. This means that the teen who shared that content would be left alone without their friends’ and peers’ support. In addition, Meta’s content strategy doesn’t address the unsafe interactions in private conversations teens have online.

Striking a balance

The challenge, then, is to protect younger users without invading their privacy. To that end, we conducted a study to find out how we can use the minimum data to detect unsafe messages. We wanted to understand how various features or metadata of risky conversations such as length of the conversation, average response time and the relationships of the participants in the conversation can contribute to machine learning programs detecting these risks. For example, previous research has shown that risky conversations tend to be short and one-sided, as when strangers make unwanted advances.

We found that our machine learning program was able to identify unsafe conversations 87% of the time using only metadata for the conversations. However, analyzing the text, images and videos of the conversations is the most effective approach to identify the type and severity of the risk.

These results highlight the significance of metadata for distinguishing unsafe conversations and could be used as a guideline for platforms to design artificial intelligence risk identification. The platforms could use high-level features such as metadata to block harmful content without scanning that content and thereby violating users’ privacy. For example, a persistent harasser who a young person wants to avoid would produce metadata – repeated, short, one-sided communications between unconnected users – that an AI system could use to block the harasser.

Ideally, young people and their care givers would be given the option by design to be able to turn on encryption, risk detection or both so they can decide on trade-offs between privacy and safety for themselves.

Afsaneh Razi receives funding from NSF.