[Paper Review] ChatGPT: Applications, Opportunities, and Threats

Aram Bahrini, Mohammadsadra Khamoshifar+5 more|arXiv (Cornell University)|Apr 14, 2023

Artificial Intelligence in Healthcare and Education41 citations

TL;DR

This paper surveys ChatGPT’s applications, opportunities, and threats across 10 domains and reports an educational experiment showing GPT-4 outperforms GPT-3.5. It discusses limitations, ethics, and governance for responsible use.

ABSTRACT

Developed by OpenAI, ChatGPT (Conditional Generative Pre-trained Transformer) is an artificial intelligence technology that is fine-tuned using supervised machine learning and reinforcement learning techniques, allowing a computer to generate natural language conversation fully autonomously. ChatGPT is built on the transformer architecture and trained on millions of conversations from various sources. The system combines the power of pre-trained deep learning models with a programmability layer to provide a strong base for generating natural language conversations. In this study, after reviewing the existing literature, we examine the applications, opportunities, and threats of ChatGPT in 10 main domains, providing detailed examples for the business and industry as well as education. We also conducted an experimental study, checking the effectiveness and comparing the performances of GPT-3.5 and GPT-4, and found that the latter performs significantly better. Despite its exceptional ability to generate natural-sounding responses, the authors believe that ChatGPT does not possess the same level of understanding, empathy, and creativity as a human and cannot fully replace them in most situations.

Motivation & Objective

• Review existing literature on ChatGPT and its foundational GPT architecture.
• Classify and analyze ChatGPT applications, opportunities, and threats across 10 domains.
• Illustrate domain-specific examples in business/industry and education.
• Conduct an experimental comparison of GPT-3.5 and GPT-4 in an educational setting.
• Discuss ethical, privacy, and governance considerations for responsible use.

Proposed method

• Literature review of ChatGPT developments and transformer-based models.
• Domain-based categorization of applications, opportunities, and threats across 10 areas.
• Detailed examples in business/industry and education to illustrate practical use.
• Experimental study comparing GPT-3.5 and GPT-4 using a university midterm exam (n=50 per model).
• Statistical analysis including normality tests and the Mann-Whitney U test to compare medians.
• Interpretation of results and discussion of limitations and implications for practice.

Experimental results

Research questions

• RQ1 What are the primary applications of ChatGPT across ten domains?
• RQ2 What opportunities and threats does ChatGPT present in each domain?
• RQ3 Does GPT-4 outperform GPT-3.5 in educational tasks, and by how much?
• RQ4 What ethical, privacy, and security concerns arise from deploying ChatGPT in education and other sectors?

Key findings

• GPT-4 significantly outperforms GPT-3.5 in the educational exam task, with a higher median score.
• ChatGPT has wide-ranging applications across ten domains, including business, education, science/tech, government, health, infrastructure, environment, communication, arts, and lifestyle.
• Main threats across domains include data bias, hallucinations, lack of transparency, privacy and security concerns, and ethical issues.
• Education-focused use of ChatGPT can support teaching, research, and writing, but requires careful mitigation of bias, reliance, and integrity risks.
• The experimental results include detailed statistics showing GPT-4’s superior performance and non-normal distribution of scores, necessitating nonparametric testing.