Constructing an Arabic Deepfake Detection Dataset from Government Sources and LLM-Generated Text

Najla A. Aljaloud; Amal Sunba; Tarek Helmy

doi:10.1016/j.procs.2026.01.047

Constructing an Arabic Deepfake Detection Dataset from Government Sources and LLM-Generated Text

Najla A. Aljaloud^*
, Amal Sunba
, Tarek Helmy

^*Corresponding author for this work

Research output: Contribution to journal › Conference article › peer-review

Abstract

The increasing use of large language models (LLMs) such as ChatGPT and DeepSeek poses new challenges for detecting AI-generated (deepfake) Arabic news. This paper constructs a high-quality dataset of authentic governmental Arabic news and corresponding deepfake articles generated using advanced LLMs. A BERT model fine-tuned on this dataset achieved an accuracy of 84.3% on an unseen test set. ross-domain assessment indicated some generalization, achieving 77.3% accuracy on X (Twitter) but exhibiting reduced performance in other domains. These results emphasize the capabilities and constraints of transformer-based methods for identifying Arabic deepfake news.

Original language	English
Pages (from-to)	385-392
Number of pages	8
Journal	Procedia Computer Science
Volume	275
DOIs	https://doi.org/10.1016/j.procs.2026.01.047
State	Published - 2026
Event	7th International Conference on AI in Computational Linguistics, ACLing 2025 - Hybrid, Dubai, United Arab Emirates Duration: 6 Dec 2025 → 7 Dec 2025

Bibliographical note

Publisher Copyright:
© 2025 The Authors. Published by Elsevier B.V.

Keywords

Arabic Deepfake Detection
BERT
Cross-Domain Validation
Dataset Construction
Large Language Models (LLMs)

ASJC Scopus subject areas

General Computer Science

Access to Document

10.1016/j.procs.2026.01.047

Cite this

@article{9f0fd234ba3e419d968c82f4faaf28bc,

title = "Constructing an Arabic Deepfake Detection Dataset from Government Sources and LLM-Generated Text",

abstract = "The increasing use of large language models (LLMs) such as ChatGPT and DeepSeek poses new challenges for detecting AI-generated (deepfake) Arabic news. This paper constructs a high-quality dataset of authentic governmental Arabic news and corresponding deepfake articles generated using advanced LLMs. A BERT model fine-tuned on this dataset achieved an accuracy of 84.3\% on an unseen test set. ross-domain assessment indicated some generalization, achieving 77.3\% accuracy on X (Twitter) but exhibiting reduced performance in other domains. These results emphasize the capabilities and constraints of transformer-based methods for identifying Arabic deepfake news.",

keywords = "Arabic Deepfake Detection, BERT, Cross-Domain Validation, Dataset Construction, Large Language Models (LLMs)",

author = "Aljaloud, \{Najla A.\} and Amal Sunba and Tarek Helmy",

note = "Publisher Copyright: {\textcopyright} 2025 The Authors. Published by Elsevier B.V.; 7th International Conference on AI in Computational Linguistics, ACLing 2025 ; Conference date: 06-12-2025 Through 07-12-2025",

year = "2026",

doi = "10.1016/j.procs.2026.01.047",

language = "English",

volume = "275",

pages = "385--392",

journal = "Procedia Computer Science",

issn = "1877-0509",

publisher = "Elsevier B.V.",

}

TY - JOUR

T1 - Constructing an Arabic Deepfake Detection Dataset from Government Sources and LLM-Generated Text

AU - Aljaloud, Najla A.

AU - Sunba, Amal

AU - Helmy, Tarek

PY - 2026

Y1 - 2026

N2 - The increasing use of large language models (LLMs) such as ChatGPT and DeepSeek poses new challenges for detecting AI-generated (deepfake) Arabic news. This paper constructs a high-quality dataset of authentic governmental Arabic news and corresponding deepfake articles generated using advanced LLMs. A BERT model fine-tuned on this dataset achieved an accuracy of 84.3% on an unseen test set. ross-domain assessment indicated some generalization, achieving 77.3% accuracy on X (Twitter) but exhibiting reduced performance in other domains. These results emphasize the capabilities and constraints of transformer-based methods for identifying Arabic deepfake news.

AB - The increasing use of large language models (LLMs) such as ChatGPT and DeepSeek poses new challenges for detecting AI-generated (deepfake) Arabic news. This paper constructs a high-quality dataset of authentic governmental Arabic news and corresponding deepfake articles generated using advanced LLMs. A BERT model fine-tuned on this dataset achieved an accuracy of 84.3% on an unseen test set. ross-domain assessment indicated some generalization, achieving 77.3% accuracy on X (Twitter) but exhibiting reduced performance in other domains. These results emphasize the capabilities and constraints of transformer-based methods for identifying Arabic deepfake news.

KW - Arabic Deepfake Detection

KW - BERT

KW - Cross-Domain Validation

KW - Dataset Construction

KW - Large Language Models (LLMs)

UR - https://www.scopus.com/pages/publications/105041089394

U2 - 10.1016/j.procs.2026.01.047

DO - 10.1016/j.procs.2026.01.047

M3 - Conference article

AN - SCOPUS:105041089394

SN - 1877-0509

VL - 275

SP - 385

EP - 392

JO - Procedia Computer Science

JF - Procedia Computer Science

T2 - 7th International Conference on AI in Computational Linguistics, ACLing 2025

Y2 - 6 December 2025 through 7 December 2025

ER -

Constructing an Arabic Deepfake Detection Dataset from Government Sources and LLM-Generated Text

Abstract

Bibliographical note

Keywords

ASJC Scopus subject areas

Access to Document

Fingerprint

Cite this