Chatgpt And Beyond: A Study on Conversational Generative AI Systems

Kamunuri Ganapathi Babu1, Thota Keerthini2
Assistant Professor, Department of Computer Science and Engineering, St. Martin’s Engineering College,
Hyderabad, India ganapathicse2@gmail.com
Student, Department of Computer Science and Engineering, St. Martin’s Engineering College, Hyderabad, India
keerthinithota@gmail.com

1. ABSTRACT:This paper presents a comprehensive study of conversational generative artificial intelligence (AI) systems, with a primary focus on ChatGPT (GPT-4) and a systematic comparison with Google Gemini Ultra, Anthropic's Claude 3 Opus, and Meta's LLaMA 2, examining their transformer-based architectures, training methodologies including Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI (CAI), and performance across five industry-standardbenchmarks — MMLU, HumanEval, HellaSwag, GSM8K, and TruthfulQA — while also investigating the ethical challenges of hallucination, demographic bias, data privacy, adversarial misuse, and environmental cost of large-scale model training; through a systematic review of existing literature and comparative analysis, we evaluate the capabilities and limitations of these systems across real-world application domains including education, healthcare, legal practice, finance, and software engineering, ultimately finding that while these systems exhibit remarkable and often human-competitive generative and reasoning abilities, the gap between benchmark performance and reliable real-world deployment remains substantial, and that the value of conversational AI is not intrinsic but relational — shaped by how these systems are designed, contextualized, and governed — contributing to the growing body of knowledge on responsible AI development and deployment.Keywords: Generative AI, Large Language Models, ChatGPT, Claude, Gemini, Conversational AI, Natural Language Processing, RLHF, Transformer Architecture, Ethics in AI