Rich Data Versus Quantity of Data in Code Generation AI: A Paradigm Shift for Healthcare

Abstract

This paper evaluates the critical trade-offs between "rich data" and "data quantity" approaches in Code Generation AI (Code Gen AI) and autonomous code agents, particularly in high-integrity sectors like healthcare. While Code Gen AI can enhance productivity by up to 55% in controlled environments, systems trained on unfiltered, large-scale datasets often increase code duplication, churn, and error rates. The paper demonstrates that in sectors where accuracy, auditability, and privacy are paramount, data richness consistently outperforms brute-force scaling strategies. Using Self-Evolving Software (SES) as a case study, the research contrasts outcomes from both paradigms and proposes a weighted matrix for data selection in Code Gen AI systems. The findings show that rich, curated, domain-specific datasets produce more reliable, compliant, and sustainable code with significantly reduced technical debt, particularly in regulated environments where quality and ethical considerations are essential. The paper concludes with best practice guidelines for implementing Code Gen AI in sensitive sectors, emphasizing that sustainable software engineering requires scaling ethically and intelligently rather than indiscriminately.