AI Advances COBOL Migration But Experts Warn Legacy Replacement Remains Tough

AI is being positioned as a possible remedy for the global reliance on obsolete enterprise software, but experts caution that current systems remain far from being safe enough to replace critical infrastructure wholesale. OpenAI said on October 19 that its latest models had achieved “substantial improvements” in translating legacy code, including COBOL, into modern languages such as C++, Java, Kotlin, Python and Rust, while preserving the original software’s behavior. The claim is significant because COBOL, created in 1959, still powers large parts of the global economy. IBM estimates it underpins more than 40 percent of online banking systems, supports 80 percent of in-person credit card transactions, and handles 95 percent of ATM transactions. It also supports government services including the IRS, Social Security and state systems. Inefficiencies linked to COBOL have been estimated at roughly $105 billion a year in the US GDP. Legacy languages such as Fortran are similarly entrenched in scientific computing, weather forecasting, atmospheric and ocean modeling, and nuclear security. Replacing them is difficult because the systems are mission-critical, poorly documented and often maintained by a shrinking pool of programmers with expertise in obsolete tools. Many organizations have tried to modernize for decades, but the risk of breaking something essential has usually favored leaving the systems in place. OpenAI reported that GPT-6.1 Sol reached 87 percent behavior-preservation accuracy at low reasoning effort and 98 percent at higher effort, compared with lower scores for earlier models. The company also said its models improved at building database software in languages including COBOL. Anthropic has also promoted legacy modernization, publishing a code modernization playbook in February and marketing Claude Code as a tool for moving COBOL workloads to Java or Python, although without publishing performance benchmarks. Shares in IBM, which earns revenue from maintaining legacy systems, fell after that announcement. However, several computer science researchers argued that these results do not show the technology can yet modernize large, real-world systems. OpenAI’s own spokesperson acknowledged that legacy code “is still far from a solved problem.” Vadim Zaytsev of the University of Twente said the problem is not merely translating millions of lines of code, but understanding system architecture, data flows, integrations and human expectations. A single major system can contain hundreds of millions of lines of COBOL, and even a small failure rate can affect many components. Arie van Deursen of Delft University of Technology criticized OpenAI’s evaluation as too narrow, saying the tasks were far simpler than the hard work of decomposing complex legacy systems into replaceable parts. Alastair Donaldson of Imperial College London emphasized that understanding undocumented behavior is the central challenge, especially as few programmers remain fluent in COBOL. He said rigorous testing, including parallel workflow comparisons, is essential. Jacopo Mauro of the University of Southern Denmark suggested formal verification could eventually validate translations automatically, but he said that would require enormous effort and human review remains indispensable today. The broader message from the debate is that AI may reduce costs and accelerate certain migration tasks, but it does not eliminate the need for systematic documentation, testing and human oversight. For institutions whose services depend on decades-old COBOL, the technology may help ease maintenance and incremental modernization, yet it is not yet a trustworthy replacement for wholesale system overhauls.