Artificial Intelligence and Misinformation Detection in Digital Media: An Explainable Machine Learning Framework
Year: 2026
Type: preprint
Abstract: The spread of misinformation on social media is a pressing problem in digital information landscapes, especially in times of sensitive political discourse, like national elections. Although machine learning and deep learning models have proven effective in predicting misinformation, they make decisions in a black box fashion, which hinders their application in a real-world platform in which accountability and human supervision are important. In this study, a machine learning model that is explainable is presented for detecting misinformation from social media text, using a set of 8495 labeled tweets gathered throughout the 2016 United States presidential election cycle. Five classifiers were considered as baseline classifiers: logistic regression, support vector machine, random forest, and two types of feed-forward neural networks (FFNN-A and FFNN-B) trained over a combined lexical and account-level feature space were compared along with a proposed stacked ensemble framework that combines linear, tree-based, and neural classifiers through a meta-learning layer that validates and calibrates the thresholds of each base-level classifiers. Using SHAP (SHapley Additive exPlanations), the lexical and metadata features that informed the classification decision were highlighted and a structured error analysis was conducted to analyse the successful and unsuccessful cases. The proposed framework performed best on the test set in terms of F1-score (0.497) and ROC-AUC (0.761) of all six approaches considered, and this was confirmed using five-fold cross-validation, as well as by finding in a bootstrap confidence interval on the difference between the F1-scores, a value that is just above zero and a McNemar's test that reveals a statistically significant difference (p < .001) in the patterns of classification errors between the proposed approach and the strongest individual baseline approach. While the results reported are headline findings, two additional robustness checks should be reported as central results and not as incidental results: Chance-corrected inter-annotator agreement on the ground-truth labels of the dataset was found to be low, at Cohen's κ = 0.084; A temporal generalization test (training on pre-election tweets and testing on post-election-day tweets) showed that the performance of every model deteriorated to near chance levels (ROC-AUC ≈ 0.56–0.57), indicating that improvements in in-domain performance – such as those reported in the proposed framework – do not necessarily imply robustness to genuine temporal distribution shift. The analysis by SHAP showed that the most important predictors of misinformation, in addition to lexical ones like second-person pronouns and mentions of the candidates' names, were structural features like the number of media in an account, the number of links embedded in the text, and temporal markers related to the election, plus verification status. In aggregate, this suggests that explainability mechanisms may be incorporated into pipelines for detecting misinformation without a complete loss of predictive performance, and that reliability of the labels, and their generalizability over time, not architectural sophistication will be the ultimate limits on trustworthy deployments of systems like these. Discussions and implications of platform governance, algorithmic accountability, and future extensions to multimodal are offered.
Source: Research Square
Authors: Muhammad Usman Aslam, Hafiz Aqib Sattar
Institutions: University of West Florida, Norfolk State University
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225821197