The spread of fake news in Arab media outlets, including social media, poses a major challenge to the integrity of information and public trust. In this regard, Dr. Ahmed Shehata and other researchers joined forces to develop a comprehensive deep learning framework called ArabFake, which is an algorithm for detecting fake news, classifying misleading information and predicting potential risks.

Built on MARBERTv2, an advanced model for multi-dialectal Arabic tweets, ArabFake efficiently handles the complexities of the Arabic language while performing the three main tasks of detecting fake news, classifying content and assessing possible risks.

The researchers trained the algorithm by using a dataset of 2,495 manually classified news items, with labels carefully verified by experts in classifying news according to fake news categories and risk levels. The model was applied to the ANS Corpus and AraNews datasets, where the data size amounted to nearly 199,331 news items in Arabic, including both authentic and fake news, thus providing a solid basis for evaluating its efficiency. 

ArabFake proved remarkably effective, achieving 94.12 percent accuracy in detecting fake news, 84.92 percent in classifying content and 88.91 percent in assessing risks. These results reflect the reliability of the algorithm and its ability to efficiently perform multiple tasks. The researchers believe that this outstanding performance may have significant practical implications, whether for news organisations or for fact-checking projects, thereby contributing to enhancing the credibility of news and reducing the spread of disinformation. 

An experimental analysis of the gathered information further revealed important patterns in the distribution of Arabic-language fake news. About 60.4 percent of this was fabricated stories, while 22.4 percent was economic disinformation. These are the most prevalent forms of fake news in the Arab media.

The researchers revealed that nearly two-thirds of fake news content poses a high risk to society, which confirms the urgent need for developing effective detection systems. The implementation of equivalence scoring in the new algorithm has proven particularly effective in identifying linguistic patterns associated with fake news, providing new insights into the characteristics of misleading content. 

The algorithm provides an instant assessment of published news, determines whether the content is true or not, estimates the level of potential risk, prioritises intervention and quickly debunks high-risk disinformation. In other words, it enables practical assessment of fake information in Arabic, while providing an opportunity for media, news agencies, content moderation systems, media literacy initiatives and analysts to investigate patterns of misinformation.