When Understanding Becomes Risk: Authenticity and Safety Risks of the Emerging Image Generation Paradigm MLLM

The paper systematically compares the safety risks of multimodal large language models (MLLMs) and diffusion models in image generation, covering two dimensions: unsafe content generation and fake image synthesis. On multiple unsafe generation benchmark datasets, MLLMs tend to generate more unsafe images than diffusion models, partly because diffusion models struggle to understand abstract prompts and output corrupted images, whereas MLLMs can understand these prompts and generate unsafe content. Existing state-of-the-art fake image detectors also find it harder to identify images generated by MLLMs, and even if the detector is retrained with MLLM-specific data, it can still be bypassed by giving the MLLM longer, more descriptive inputs. The paper argues that the safety risks of this emerging generation paradigm have not yet been fully recognized, posing new challenges to real-world safety.