Can AI help solve the COBOL code crisis?

AI coding assistants now demonstrate the ability to translate complex legacy code with high fidelity, reducing the risk of manual migration projects.

OpenAI has announced substantial improvements in its AI models' ability to migrate aging software code, specifically highlighting COBOL, into modern programming languages. The company claims its latest models, particularly GPT-6.1 Sol, show significant progress in preserving the original software's behavior, achieving up to 98 percent accuracy in its tests. This capability could address the ongoing crisis of critical infrastructure running on legacy code, which is difficult and risky to replace due to a shortage of skilled programmers and the complexity of the systems. IBM, which profits from maintaining legacy systems, saw its shares decline following similar announcements from competitors. While OpenAI and Anthropic are pushing AI solutions for code modernization, experts caution that current AI capabilities are still far from safely replacing entire critical legacy systems, noting the immense complexity beyond just code translation, including data, integrations, and system-wide behavior verification.

OpenAI’s claims suggest a significant leap in AI’s capacity to handle the nuanced and critical task of legacy code modernization, potentially unlocking billions in economic value lost to inefficient systems. The challenge of COBOL, a language embedded in 40% of banking and 95% of ATM transactions, has long been a major technical debt for global finance and government. While AI has historically struggled with the unstructured nature and sheer scale of these systems, OpenAI’s reported improvements in behavior preservation signal a potential pathway to more reliable and cost-effective modernization efforts.

However, the broader ecosystem of legacy systems involves more than just code translation; it encompasses data, intricate interdependencies, and operational context that AI currently struggles to fully grasp. Experts emphasize that even minor deviations in critical system behavior post-migration are unacceptable, meaning a 99% automated solution leaves an unmanageable 1% challenge. This suggests that while AI may accelerate parts of the modernization process, human oversight and specialized testing methods will remain essential for the foreseeable future, particularly for high-stakes infrastructure.