The rapid dissemination of short claims on social media makes it challenging to detect misinformation, since it involves a classification problem where the model is required to distinguish various levels of truthfulness rather than just true and false. This paper presents an explainable transformer-based model for classifying social-media-style claims that decouples model explanation from factual verification. The public LIAR benchmark that this study revolves arounds contains a total of 12,836 manually fact-checked and labelled short statements from the source PolitiFact. It uses six different labels for truthfulness. A transformer encoder transforms any claim into a contextual representation, a multiclass classification head estimates the label probabilities, with a token-level explainability layer identifies words with the strongest influence on the prediction. Faithfulness is the reduction in predicted-class score after removing tokens that are highly attributed. The authors also offer a reproducible experimental protocol for BERT and RoBERTa while anchoring the task’s anticipated difficulty to released LIAR results, instead of inventing unexecuted measures. The literature demonstrates that six-class RIAR classification is considerably harder than binary fake-news detection, and that quality of the explanation cannot be inferred from accuracy alone. As a result, the proposed framework views the output of the transformer as a triage signal: it can indicate patterns associated with previously fact-checked claims and reveal the text that the classifier uses, but it does not by itself prove that a new claim is true or false. The effective use of automated misinformation detection on social networks and fact-checking workflows hinges on this distinction.