The growing diversity of generative models has made distinguishing AI-generated images increasingly challenging, especially under unknown generator scenarios. While diffusion reconstruction-based methods have shown promise in detecting generated content, a key limitation persists: the entanglement between artifact and image content distributions during reconstruction. In this paper, we propose DDRE (Decoupled Diffusion Reconstruction Error), a novel, training-free detection framework that explicitly disentangles these two distributions via a two-stage reconstruction process. By comparing reconstruction error maps from successive denoising stages and applying a simple norm-based ratio metric, DDRE effectively isolates generative artifacts. Furthermore, we introduce a self-conditioning strategy that leverages CLIP embeddings to guide reconstruction in the absence of prompts, enhancing the fidelity of inversion. To provide a more comprehensive evaluation, we propose GenNEXT, a benchmark that extends GenImage by reducing JPEG bias and incorporating more recent generative models. Extensive experiments on GenNEXT demonstrate that DDRE achieves state-of-the-art performance and exhibits strong generalization to previously unseen generators, all without requiring additional training.