The influence of a deceptive AI agent in a multi-agent system

Abstract

This thesis explores the domain of artificial intelligence, multi-agent systems, and large language models. The term artificial intelligence was first invented in the 1950s, and it is indeed an ever-growing area of research. With the advancement of technology, large language models have been introduced and they have become one of the most prominent inventions in the history of artificial intelligence. However, due to the black-box nature of such language models, it is challenging to truly understand the working principles inside them. Despite the marvellous effort of the developers to construct honest LLM agents, such models have shown misleading and emergent behaviours. Thus, the nature of deception developed in LLMs is still ambiguous and researchers have been trying to achieve clarity in defining, detecting, and mitigating deception. It has also been proven in previous studies that current state-of-the-art LLMs are capable of deceiving humans without any direct instructions from the developers. This thesis aims to explore the deceptive nature of LLMs or AI agents in general as well as deception detection and mitigation methods. The experiment presented in this thesis also shows that deception appears in competitive scenarios and that appropriate prompting is crucial to define the characteristics of an LLM-based agent whether it is cooperative or manipulative.