FakeXplain: AI-Generated Images Detection via Human-Aligned Grounded Reasoning

Core task: AI-generated image detection with interpretable explanations. The field has evolved from purely classification-driven approaches toward systems that not only identify synthetic content but also provide human-understandable rationales for their decisions. The taxonomy reflects this dual emphasis through eight major branches. Deep Learning Classification Approaches and Forensic and Signal-Based Detection Methods focus on discriminative accuracy using CNNs, vision transformers, and frequency-domain analysis. Explainable AI Techniques for Detection Transparency leverage saliency maps, attention mechanisms, and post-hoc interpretation tools to reveal which image regions or features drive predictions. Multimodal Large Language Model-Based Detection and Explanation represents a newer direction that integrates vision-language models to generate natural-language justifications and localize artifacts. Meanwhile, Datasets, Benchmarks, and Evaluation Frameworks establish standardized testbeds, User-Centric and Interactive Detection Systems explore human-in-the-loop workflows, Domain-Specific and Specialized Applications address niche contexts like medical imaging or sign language, and Generalization, Robustness, and Cross-Generator Detection tackle the challenge of maintaining performance across diverse generative models. Recent work has increasingly emphasized grounded reasoning that pinpoints suspicious artifacts rather than offering only global verdicts. FakeXplain[0] exemplifies this trend by combining multimodal large language models with explicit artifact localization, situating itself within the Grounded Reasoning with Artifact Localization cluster. This approach contrasts with earlier explainability efforts such as Deepfake Detection Explainable[1] and CNN Explainable Detection[5], which primarily relied on gradient-based saliency or attention overlays without structured linguistic explanations. Closely related works like ForenX[3] and AIGI Holmes[10] similarly pursue fine-grained localization and interpretable outputs, yet FakeXplain[0] distinguishes itself by leveraging the reasoning capabilities of large language models to articulate why specific regions appear synthetic. A key open question across these branches is how to balance detection accuracy with explanation fidelity, especially when models must generalize to unseen generators or adversarially perturbed images.