Fake news detection is often treated as a content-centered task, yet social media posts are embedded in user reactions that express skepticism, endorsement, correction, ridicule, and affective responses. This paper asks what makes user reactions useful for fake news detection. We study performance according to different dimensions: text only, emotions in comments, and number of available comments. We study this question across Fakeddit and Weibo20 under both standard and temporal splits (in a temporal split, most recent posts are included in the testing split). Robustness of results is measured across two encoder families (BERT-base, RoBERTa-base) and two English emotion taxonomies (dair-ai/emotion, GoEmotions). Across these configurations, comment-only representations outperform post-only baselines in most evaluation regimes, though this gain depends on dataset, split, and number of available comments. Emotion-fine-tuned comment encoders produce a temporal-shift-specific significant benefit compared to non-temporal splits.

We compare post-only representations, generic comment representations, emotion-fine-tuned comment representations, and post–reaction fusion models, while varying comment availability, showing that a minimum number of comments is necessary to improve detection of fakes. In addition, we show that fine-tuning comment or reaction-encoders on relevant auxiliary tasks such as emotions has a significant impact on the distribution of embeddings of fake and real posts in their representation space.

In light of these findings, we believe that user responses—and their emotional structure for example—should be taken into account when detecting fake news or compiling datasets for this purpose.