Ensemble AI model flags and rewrites media bias in news text

Researchers have built an artificial intelligence system that detects bias in news content and then rewrites the flagged passages, reporting an accuracy of 77.5 percent and a Macro Average F1 score of 0.705 for the detection stage. The system, called FairJEn, is described in a paper published in the journal Neural Computing and Applications on 27 October 2026.

FairJEn combines three machine learning components in an ensemble: the transformer-based language model RoBERTa, a bidirectional long short-term memory network (Bi-LSTM), and XGBoost serving as a meta-classifier that merges their outputs. According to the authors, Rajalakshmi Sivanaiah, Pranav Vaithiya Subramani and Sanjjit Sounderrajan, this combination outperformed the individual approaches they compared it against. The work uses natural language processing and machine learning techniques to address a problem the authors frame as increasingly urgent: public opinion is largely shaped by media and digital social platforms, and growing bias in coverage can steer users away from objective information and foster skewed perceptions.

Detection alone does not change what readers see, so the second stage of the system focuses on mitigation. The researchers used Pointwise Mutual Information, a statistical measure of how strongly two terms co-occur, to identify biased wording within texts. Flagged content was then refined using BERT-based contextual replacements, in which the language model proposes substitute phrasing that fits the surrounding context. The authors report that this rewriting preserved meaning to a striking degree: the revised text maintained a semantic similarity score of 0.9997 with the original, with a minimal semantic shift of 0.0047, which they present as evidence that the intervention removes bias without distorting the substance of the reporting.

The third component tackles a subtler problem: even unbiased individual articles can produce a skewed picture if a reader’s recommendations keep circling the same narrow cluster of viewpoints. To address this, the team built a news recommendation system based on Siamese Networks, a neural architecture that learns to measure similarity between pairs of items. The system was designed to diversify the content a user is exposed to. The authors report that moving from cluster-based recommendations to a dataset-wide approach improved both bias coverage and diversity, with scores of 0.6 and 1.52 respectively.

The published abstract describes the work as an original research article appearing in Volume 38 of the journal. As with any machine learning system reported at these performance levels, the figures reflect the authors’ evaluation on their data, and the full paper, which is behind a subscription preview on the publisher’s site, would detail the datasets and comparison baselines. The authors position the combined pipeline — ensemble detection, statistically guided rewriting, and diversity-aware recommendation — as a step toward balanced content delivery and more transparent journalism, arguing that such tools can support fairer media consumption at a time when algorithmic feeds increasingly determine what audiences read.

Subject of Research: Machine learning detection and mitigation of media bias in news content

Article Title: FairJEn: media bias analysis using ensemble learning approach for fair journalism

Article References: Sivanaiah, R., Subramani, P. V., & Sounderrajan, S. (2026). FairJEn: media bias analysis using ensemble learning approach for fair journalism. Neural Computing and Applications, 38(20), 853. https://doi.org/10.1007/s00521-026-12540-y

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12540-y

Keywords: media bias, ensemble learning, RoBERTa, Bi-LSTM, XGBoost, natural language processing, BERT, Pointwise Mutual Information, Siamese Networks, news recommendation, fair journalism, FairJEn