Information disorder research overwhelmingly focuses on fabricated or manipulated content (fake news, deepfakes, propaganda) while comparatively neglecting the most pervasive form of distorted information: media bias. Unlike outright falsehoods, media bias operates within the boundaries of factual reporting, distorting public understanding through framing, omission, and word choice rather than fabrication. This makes it harder to detect, harder to regulate, and paradoxically more influential, since it originates from trusted mainstream sources rather than marginal actors. In this position paper, we argue that media bias should be recognized as a first-class category within information disorder frameworks. Drawing on the Wardle and Derakhshan (2017) taxonomy, communication theory, and a systematic review of over 100 studies on automated media bias detection, we demonstrate that current frameworks inadequately account for the systematic distortion of true content. We present a consolidated taxonomy of media bias types organized by linguistic level, compare detection paradigms across the information disorder and media bias communities, and identify four properties that make media bias uniquely dangerous: its scale, its source credibility, the invisibility of omission, and its cumulative normative effect. We conclude with an integrated research agenda grounded in specific gaps identified through the review.

Current multilingual benchmarks cover only a fraction of the languages represented in modern pretraining corpora and often conflate language proficiency with the ability to answer questions grounded in specific regions and communities. We introduce INCLUDE-2.0, a large-scale benchmark for region-aware multilingual factual evaluation, which covers 91 languages from 86 countries (127 total language–country pairs), comprising over 220K curated samples drawn from native regional sources. Unlike existing benchmarks that often treat language as a proxy for region, IINCLUDE-2.0 distinguishes between questions whose answers are independent of context, vary by country, concern widely documented facts about a place, or depend on knowledge primarily expressed through local linguistic and cultural sources. Each item is annotated with its regional knowledge type and domain, and paired with an English translation, enabling controlled analysis of the relationship between language proficiency and possession of regional knowledge.

Evaluating frontier and open-weight models on IINCLUDE-2.0, we find that aggregate multilingual accuracy masks substantial variation across regional knowledge types, languages, and regions. We find that frontier models perform worse on locally embedded knowledge than on widely documented facts about regions, with an average drop of ~14.5 points. We also show that for open-weight models, scale improves accuracy but does not reliably make performance invariant to whether a question is asked in English or its original language. INCLUDE-2.0 is available at https://huggingface.co/datasets/epfl-nlp/include-2.0

How similar is the detection of media bias to the detection of persuasive techniques? We have explored how transferring knowledge from one task to the other may help to improve the performance. This paper presents the systems developed for participating in the SemEval-2023 Task 3: Detecting the Genre, the Framing, and the Persuasion Techniques in Online News in a Multi-lingual Setup. We have participated in both the subtask 1: News Genre Categorisation, and the subtask 3: Persuasion Techniques Detection. Our solutions are based on two-stage fine-tuned multilingual models. We evaluated our approach on the 9 languages provided in the task. Our results show that the use of transfer learning from media bias detection to persuasion techniques detection is beneficial for the subtask of detecting the genre (macro F1-score of 0.523 in the English test set) as it improves previous results, but not for the detection of persuasive techniques (micro F1-score of 0.24 in the English test set).