A SYSTEMATIC REVIEW OF GENERATIVE ARTIFICIAL INTELLIGENCE: ETHICS, AGENTIC AI, AND FUTURE DIRECTIONS

Abstract

The rapid proliferation of Large Language Models (LLMs), multimodal generative architectures, and autonomous agentic AI systems since 2017 has fundamentally transformed computational intelligence, yet the field lacks a unified systematic review encompassing ethics, agentic capabilities, and emerging future trajectories. This paper presents a PRISMA-guided systematic review of 312 peer-reviewed studies published between 2017 and 2024, drawn from Scopus, Web of Science, IEEE Xplore, and ACM Digital Library. We synthesize empirical evidence across three dimensions: (i) the technical evolution of generative AI from Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) through transformer-based LLMs to multimodal foundation models; (ii) the ethical landscape including bias, hallucination, privacy, accountability, and regulatory governance; and (iii) the emergence of agentic AI — autonomous, tool-using, self-correcting agents capable of multi-step planning and environmental interaction. Results reveal that 68.4% of reviewed studies identify hallucination and factual inconsistency as the primary technical barrier, while 54.2% flag the absence of enforceable accountability mechanisms as the dominant governance gap. Agentic systems demonstrate 3.2x improvement in task completion benchmarks over static LLMs when equipped with planning modules and external tool integration, yet introduce novel safety risks including misalignment cascades and adversarial prompt exploitation. We propose a Unified Governance and Capability Framework (UGCF) for responsible agentic AI deployment and identify six priority research directions for 2025–2030.