Legacy system modernization remains one of the most persistent and costly challenges for enterprises worldwide, particularly within the financial, administrative, and government sectors where legacy COBOL codebases still form the backbone of critical operations. In recent years, generative artificial intelligence has been widely positioned as an automated, cost-effective solution for translating outdated mainframe languages into modern, object-oriented languages like Java. However, new research suggests that this approach has a critical structural flaw: AI-powered translators are replicating legacy software bugs alongside the functional code.

According to Hacker News Front Page, an academic paper published on the arXiv repository analyzed the performance of artificial intelligence systems in conducting these code translations. The findings indicate that while modern LLMs excel at syntax conversion and matching structural patterns, they struggle with deep semantic reasoning. Consequently, any hidden logical errors, undocumented edge cases, or functional vulnerabilities present in the original COBOL source code are faithfully translated directly into the newly generated Java code.

This "bugs-included" migration model poses significant operational risks for cloud migration initiatives. Organizations attempting to cut costs by bypassing manual code audits in favor of purely automated AI pipelines may find themselves running the exact same flawed business logic, now obscured within thousands of lines of newly generated Java. Industry experts argue that these findings emphasize the irreplaceable role of human-in-the-loop verification. While artificial intelligence can serve as an exceptionally fast drafting tool to accelerate the initial phases of software translation, extensive testing frameworks and manual oversight are still mandatory to ensure that legacy defects do not contaminate modern cloud environments.