Univrsity Of Tehran Press Journal of Information Technology Management 2980-7972 18 3 2026 07 01 Generative AI Health Assistants: From Theory to Practice 181 201 108010 10.22059/jitm.2026.108010 EN Ghada Alhudhud Prof., Department of Data Science and Artificial Intelligence, Faculty of Information Technology, Al-Ahilya Amman University, Amman, Jordan. 0000-0002-2810-6582 Sanaa Alhalabi Assistant Prof., Department of Computer Science, Faculty of Information Technology, Al-Ahilya Amman University, Amman, Jordan. 0009-0002-2839-0349 Marwan Al-Akaidi Professor, Vice Chancellor and CEO Wigwe University, Isiokpo, Nigeria. Journal Article 2026 07 20 Generative Artificial Intelligence (GenAI) has evolved significantly from early models like variational autoencoders (VAEs) to sophisticated diffusion and hybrid models, enabling the creation of entirely new data from input prompts. In the domain of natural language, powerful large language models (LLMs) such as GPT-4 now generate remarkably coherent responses, while diffusion models have opened up new applications in text-to-image generation and image editing across unimodal and multimodal spaces. These advancements are built upon foundational theories of latent variable modeling, denoising processes, and attention/transformer architectures. This paper traces the development of GenAI, with a focus on its application in mobile AI assistants integrated into smartphones and edge computing platforms. While transformative, these models are typically large and resource-intensive. Their deployment in mobile or edge environments necessitates architectural compression techniques like pruning and quantization, model partitioning between the device and server, or the development of lightweight models designed for constrained settings. This study discusses the practical transition from model architecture to a functional, AI-powered mobile assistant that utilizes dynamic inference across device and edge layers. A mixed-methods approach is employed, combining quantitative benchmarking of latency, accuracy, energy consumption, and resource utilization with qualitative assessments of user experience, trust, and privacy perceptions. The system's design incorporates principles of data protection and user authentication through voice biometric-based watermarking, aligning with our previous work on secure multimedia exchange. Using a heterogeneous dataset of benchmarked prompts, mobile hardware telemetry, and semi-structured user interviews, this research reports on performance metrics, usability findings, and privacy trade-offs. The paper concludes by offering recommended practices and reference architectures, providing a holistic framework for transitioning GenAI from a theoretical innovation to a practical, secure, and ethically aligned tool in mobile and health-oriented applications. https://jitm.ut.ac.ir/article_108010_fe78e3a154c6b0b2934e16b54a315bdc.pdf