Identifying disturbing online content being targeted at children is an important content moderation problem. However, previous approaches to this problem have focused on features of the content itself, and neglected potentially helpful insights from the reactions expressed by its online audience. To help remedy this, we present the Elsagate Corpus, a collection of over 22 million comments on more than 18,000 videos that have been associated with disturbing content. We describe the how we collected this corpus and present some insights from our initial explorations, including the surprisingly positive reactions from audiences to this content, some unusual non-linguistic commenting behavior of uncertain purpose and references to some concerning themes.

Large Language Models (LLMs) have rapidly been adopted by the general public, and as usage of these models becomes commonplace, they naturally will be used for increasingly human-centric tasks, including security advice and risk identification for personal situations. It is imperative that systems used in such a manner are well-calibrated. In this paper, 6 popular LLMs were evaluated for their propensity towards false or over-cautious risk finding in online interactions between real people, with a focus on the risk of online grooming, the advice generated for such contexts, and the impact of prompt specificity. Through an analysis of 3840 generated answers, it was found that models could find online grooming in even the most harmless of interactions, and that the generated advice could be harmful, judgemental, and controlling. We describe these shortcomings, and identify areas for improvement, including suggestions for future research directions.

In this work we use consumed text to infer Big-5 personality inventories using data we have collected from the social media platform Reddit. We test our model on two datasets, sampled from participants who consumed either fiction content (N = 913) or news content (N = 213). We show that state-of-the-art models from a similar task using authored text do not translate well to this task, with average correlations of r=.06 between the model’s predictions and ground-truth personality inventory dimensions. We propose an alternate method of generating average personality labels for each piece of text consumed, under which our model achieves correlations as high as r=.34 when predicting personality from the text being read.