| Abstract: |
Generative artificial intelligence (GenAI) and large language models (LLMs) have moved from research laboratories into everyday computer science practice, yet empirical evidence on how they are used, what benefits they deliver and which risks they introduce remains fragmented. The study is motivated by the need to connect adoption, efficiency, quality and perceived risk within a single analytical frame. This paper presents a data-driven analysis of the applications and challenges of GenAI and LLMs in computer science. A structured questionnaire was administered to 240 respondents comprising undergraduate students, postgraduate students, software developers and researchers. It was complemented by a timed comparison of five programming tasks and a quality audit of 300 LLM-generated code snippets. Descriptive statistics, paired comparisons and correlation analysis were applied to five datasets. Code generation (87.5%), debugging (72.1%) and documentation (68.3%) emerged as the most adopted applications, and assisted task completion times fell by 27.3% to 52.4%. Weekly usage hours correlated positively with perceived productivity gain (r = 0.64, p < 0.001). However, only 72.3% of generated snippets were functionally correct, 16.0% carried flagged vulnerabilities, and 78.3% of respondents identified hallucination as a major concern. Correctness declined steadily with complexity and security sensitivity, from 91.3% for simple functions to 47.5% for security-sensitive code. These findings indicate that LLMs are strong accelerators of routine work but unreliable autonomous agents for complex or security-critical development. Human verification, secure prompting, retrieval grounding and curriculum reform are therefore essential for responsible adoption. Implications for software practice, education and research governance are outlined. |
