GPT-3.5 Turbo Evaluation Tool
No Thumbnail Available
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Authors
Date
2024-03-11
Department
Major/Subject
Cloud and Network Infrastructures
Mcode
ELEC3059
Degree programme
Master's Programme in ICT Innovation
Language
en
Pages
59
Series
Abstract
This study investigates the challenges and implications associated with differentiating between computer-generated and human-generated text, with a specific focus on the utilization of the GPT-3.5 Turbo model. It explores the impact of artificial intelligence (AI) across various domains and endeavours to ascertain whether ChatGPT can discern the origin of a given text. The research is motivated by the escalating concern surrounding the misrepresentation of AI-generated text as human-authored, a practice that undermines transparency and integrity, necessitating the development of effective detection and mitigation strategies. Research Question: Can AI-generated text be distinguished from human-generated text, and if so, with what accuracy? The study examines the diverse applications of GPT-3, devises a mechanism to mask the origin of content generated by the ChatGPT-3.5 Turbo model to simulate human authorship, evaluates the efficacy of this mechanism, and proposes enhancements to facilitate the identification of AI-generated content when human-authored content is required. The study reveals that substituting synonyms proves to be a more effective method for concealing AI-generated text than rephrasing. Presently, AI models such as GPT-3.5 Turbo are unable to emulate human writing entirely. Consequently, there is a pressing need to bolster security measures to prevent instances of AI-generated text being misrepresented as human-generated. It is imperative to underscore that while such occurrences are not ubiquitous, unethical practices such as leveraging AI for academic dishonesty are unequivocally condemned. It is important to note that this study has inherent limitations, and future research endeavours will delve into the structural aspects of language, contextual considerations, alternative detection tools, diverse domains such as business and science, and multilingual contexts.Description
Supervisor
Manner, JukkaThesis advisor
Manner, JukkaKeywords
Natural Language Processing (NLP), Large Language Models (LLM), GPT-3.5 Turbo, human-written text