AI-Based Fake News Detection Using Natural Language Processing: A Review and Analysis
Year: 2026
Type: article
Abstract: Misleading news often reaches readers before its claims can be checked, which has encouraged research into artificial intelligence (AI) and natural language processing (NLP) as tools for flagging content that needs closer examination. This paper reviews 31 publications from 2011 to 2024 and organises their approaches into four groups: traditional machine learning, deep learning, transformer-based models, and multi-source or real-time systems. It then tests one concern that recurs in the literature, namely that scores measured on one dataset do not carry over to another. Four standard classifiers (Naive Bayes, Logistic Regression, Linear SVM, and Random Forest, all on TF-IDF features) were evaluated on two public datasets: Fake or Real News (2016 political articles) and CoAID (2020 COVID-19 articles). Within Fake or Real News, five-fold macro-F1 ranges from 0.87 to 0.94. When a model trained on one dataset is tested on the other, macro-F1 falls to between 0.06 and 0.42. A model trained on an early CoAID snapshot also detects few fake articles in later snapshots, although the later test set holds only 15 fake articles, so that result is inconclusive. The highest-weighted terms of the linear models include source and formatting words, which suggests dataset-specific shortcuts. The paper also specifies CAMMS, a multilingual and multimodal framework with an explanation layer. CAMMS is described and given an evaluation plan, but it has not been implemented or evaluated here. The review finds no basis for naming one model family as best in every setting, and it points to cross-dataset testing, multilingual evaluation, transparent reporting, and human oversight as priorities.
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7236600380
