# Interpretable and Reliable Detection of AI-Generated Images via Grounded Reasoning in MLLMs

**Authors:** Yikun Ji, Hong Yan, Jun Lan, Huijia Zhu, Weiqiang Wang, Qi Fan, Liqing Zhang, Jianfu Zhang  
**arXiv:** [2506.07045](https://arxiv.org/abs/2506.07045)  
**Date:** 2025-06-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.07045v1/

## Key points

- Growing demand for interpretable and robust detection methods due to rapid advancement of image generation technologies
- MLLMs (Multimodal Language Models) have strong analytical and reasoning capabilities for forgery detection when fine-tuned
- Challenges faced by existing MLLMs include hallucination and difficulty aligning visual interpretations with actual image content and human reasoning
- Dataset of AI-generated images annotated with bounding boxes and descriptive captions created to highlight synthesis artifacts, enabling human-aligned visual-textual grounded reasoning
- Multi-stage optimization strategy improves performance of MLLMs in detecting AI-generated images and localizing visual flaws
- Introduction of dataset containing 8,772 AI-generated images annotated with bounding boxes highlighting visual anomalies represents a significant contribution to the field
- Qwen-2.5-VL (32B) fine-tuned on dataset enables end-to-end system capable of detecting and explaining AI-generated images through binary authenticity decisions, predicted bounding boxes, and natural language justifications
- Fine-tuning on FakeXplained allows MLLMs to perform fine-grained visual reasoning and articulate observations clearly, marking a new paradigm in human-interpretable AI-generated image detection

## Summary

In response to the rapid advancement of image generation technologies, there is a growing demand for interpretable and robust detection methods. Existing approaches often achieve high accuracy but operate as black boxes without providing human-understandable justifications. have emerged as powerful tools with strong analytical and reasoning capabilities, despite not originally being intended for forgery detection. When properly fine-tuned, MLLMs can effectively identify AI-generated images and offer meaningful explanations. However, existing MLLMs still face challenges such as hallucination and difficulty aligning their visual interpretations with actual image content and human reasoning. To address this gap, a dataset of AI-generated images annotated with bounding boxes and descriptive captions has been constructed to highlight synthesis artifacts. This dataset serves as a foundation for human-aligned visual-textual grounded reasoning. Through a multi-stage optimization strategy that balances the objectives of accurate detection, visual localization, and coherent textual explanation, MLLMs have been fine-tuned to improve performance in detecting AI-generated images and localizing visual flaws. The resulting model significantly outperforms baseline methods by achieving superior performance in both tasks. The introduction of the  dataset containing 8,772 AI-generated images annotated with bounding boxes highlighting visual anomalies and illogical details along with concise captions explaining each flaw represents a significant contribution to the field. By fine-tuning Qwen-2.5-VL (32B) on this dataset using a two-stage training approach, an end-to-end system capable of detecting and explaining AI-generated images has been developed. This system not only produces binary authenticity decisions but also provides predicted bounding boxes paired with natural language justifications for identified regions in fake images. Through grounded reasoning enabled by fine-tuning on FakeXplained, MLLMs can now perform fine-grained visual reasoning and articulate their observations clearly. This advancement marks a new paradigm in human-interpretable AI-generated image detection where models are able to provide comprehensive rationales comparable to human annotators.