ORIGINAL ARTICLE
Figure from article: Identifying Arabic...
 
KEYWORDS
TOPICS
ABSTRACT
As large language models have become widely accessible, AI-generated Arabic text designed to mislead readers is increasingly prevalent on social media, making reliable detection across platforms a challenging task. This study investigates Arabic deepfake text detection across two contrasting environments, Twitter/X and YouTube, using benchmark datasets consisting of paired authentic and AI-generated posts. We examine the impact of text representation by applying three controlled preprocessing views built on a shared cleaning baseline. Within this framework, machine-learning classifiers, sequence-based models, and fine-tuned Arabic transformers are evaluated in both in-domain settings and cross-platform transfer scenarios. Transformer models consistently achieved the strongest in-domain performance, reaching Macro-F1 scores of 0.9912 on Twitter/X and 0.9409 on YouTube. However, the effect of preprocessing was highly variable: retaining the original text generally produced the best results, aggressive manual normalization often degraded performance, and deepfake-aware preprocessing yielded mixed outcomes depending on the model and platform. Cross-platform evaluation revealed a substantial performance gap compared to in-domain results, with the best transfer scores reaching 0.8364 from Twitter/X to YouTube and 0.9433 in the reverse direction. This study provides a controlled cross-platform evaluation of Arabic deepfake text detection and highlights the critical role of preprocessing and model pre-training in shaping robustness. Overall, the findings indicate that while high-accuracy detection within a single platform is achievable, developing models that generalize effectively across platforms remains an open challenge.
Journals System - logo
Scroll to top