Recent forgery forensics methods based on deep neural networks and large models have achieved promising results in localizing image forgeries. However, localizing tampered regions in AI-generated images remains challenging due to their pixel-level consistency and the lack of interpretable evidence. To address this, we propose an explainable image forgery localization framework via Difference Amplification and Cross-Clue Querying (DA-CCQ). Specifically, we propose a Difference Amplification Module (DAM) to capture semantic inconsistencies within latent features, which are then decoded into pixel-level reconstruction errors that are sensitive to tampering traces. To amplify forgery artifacts, DAM incorporates a hybrid generator composed of a reconstruction generator and a frequency-aware forgery predictor. The predictor guides the generator by providing both pixel-wise supervision and semantically-aware latent features, helping it focus on tampered regions. Furthermore, we design a Cross-Clue Querying (CCQ) mechanism to integrate three complementary clues: pixel-level consistency, physics-based illumination, and amplified reconstruction errors. Through mutual querying between image features and forensic clues, CCQ achieves enhanced cross-clue reasoning, improving the generalization of localization. Extensive experiments on two AI-generated and three traditional tampering datasets demonstrate that our method outperforms existing approaches in localization, particularly in challenging AI-generated scenarios.