The proliferation of AI-generated voice calls, often synthesized using Text-to-Speech (TTS) systems, poses significant challenges in security and authenticity verification. Current TTS methods employed to generate fake AI calls often suffer from long latency and degraded voice quality, leading to lower Mean Opinion Scores (MOS). However, real calls can also experience poor call quality due to network conditions, including high latency, which can cause audio distortions. In such cases, the Real-time Transport Control Protocol (RTCP) packets can be used to calculate network-based MOS, which should align with the AI-predicted MOS. This paper proposes a novel approach to detect faked AI calls by leveraging AI-based MOS prediction models like the Neural Integrated Speech Quality Assessment (NISQA) and comparing them with RTCP-derived MOS. We include mathematical formulations for calculating MOS using RTCP statistics and analyze how latency affects MOS. Additionally, we discuss related work on machine learning-based audio analysis for fraud detection in wireless networks and latency compensation technologies. Experimental results show that inconsistencies between these two MOS measurements are indicative of faked AI calls, providing a viable solution for enhancing communication security.