LLMs Are Not Yet Ready for Deepfake Image Detection

2025-06-12

AI Summary

This study investigates the zero-shot detection capability of four state-of-the-art vision-language models (VLMs)—ChatGPT, Claude, Gemini, and Grok—on three deepfake image categories: face swapping, facial expression reenactment, and synthetic generation. Method: We construct a multi-source deepfake benchmark and propose a dual-dimensional evaluation framework assessing both classification accuracy and reasoning depth. Contribution/Results: All VLMs fall significantly short of practical detection accuracy compared to specialized deepfake detectors, confirming their unsuitability as standalone forensic tools. We identify novel failure modes—including stylistic bias and vintage-image interference—for the first time. Conversely, VLMs demonstrate strong contextual understanding and natural language explanation capabilities. Accordingly, we advocate repositioning VLMs as interpretable, human-in-the-loop forensic components—not autonomous detectors. Our empirical findings and methodological framework provide foundational evidence and guidance for the paradigm shift toward explainable, collaborative digital forensics using VLMs.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

The growing sophistication of deepfakes presents substantial challenges to the integrity of media and the preservation of public trust. Concurrently, vision-language models (VLMs), large language models enhanced with visual reasoning capabilities, have emerged as promising tools across various domains, sparking interest in their applicability to deepfake detection. This study conducts a structured zero-shot evaluation of four prominent VLMs: ChatGPT, Claude, Gemini, and Grok, focusing on three primary deepfake types: faceswap, reenactment, and synthetic generation. Leveraging a meticulously assembled benchmark comprising authentic and manipulated images from diverse sources, we evaluate each model's classification accuracy and reasoning depth. Our analysis indicates that while VLMs can produce coherent explanations and detect surface-level anomalies, they are not yet dependable as standalone detection systems. We highlight critical failure modes, such as an overemphasis on stylistic elements and vulnerability to misleading visual patterns like vintage aesthetics. Nevertheless, VLMs exhibit strengths in interpretability and contextual analysis, suggesting their potential to augment human expertise in forensic workflows. These insights imply that although general-purpose models currently lack the reliability needed for autonomous deepfake detection, they hold promise as integral components in hybrid or human-in-the-loop detection frameworks.

Problem

• Evaluating VLMs' effectiveness in detecting deepfake images
• Identifying failure modes in VLM-based deepfake detection
• Assessing VLMs' potential as components in hybrid detection systems

Innovation

• Zero-shot evaluation of VLMs for deepfake detection
• Benchmark with diverse authentic and manipulated images
• Hybrid human-in-the-loop detection framework potential