Artificial intelligence (AI) has become mainstream nowadays, with more and more systems relying on AI for judgment and decision-making. For instance, during the pandemic, AI was used to determine whether customers were wearing masks, deciding if they could enter stores. Similarly, factories utilize AI for automated defect detection, categorizing non-compliant products for reprocessing or recycling. However, as AI continues to evolve, ensuring the information security of images has become increasingly critical. Our study investigates whether fragile watermarking methods can effectively safeguard the integrity of AI-based automated image judgments. To explore this issue, we implemented previously proposed fragile watermarking techniques and simulated tampering scenarios. Subsequently, we employed YOLOv7 to assess whether restored images significantly differed from their original counterparts. Our experimental results revealed that under a 30% tampering scenario, accuracy dropped to approximately 63%. We conclude that the recovery mechanism proposed by previous scholars minimally impacts AI’s feature-based judgments, validating our research findings. This study sheds light on the delicate balance between image security and AI performance, providing valuable insights for future applications.