August 2, 2026
Old code, new panic
AI migrated legacy COBOL programs to Java, bugs included
AI moved ancient banking code to Java and dragged the gremlins along too
TLDR: Researchers say AI can move old COBOL business software to Java and then test it hard enough to catch whether the new version acts like the old one. Commenters weren't sold: they warned the examples were tiny compared with real banking and government systems, and joked AI may preserve old bugs while inventing brand-new disasters.
The paper itself is pretty serious: researchers say they found a way to use AI to move old COBOL business software into Java, then hammer both versions with generated tests until the new one behaves like the old one. In their small case studies, the Java version matched the original on accepted tests, and coverage got impressively high. Sounds neat! But the comments immediately turned this from a lab win into a full-on "okay, but in the real world?" showdown.
That skepticism was the main character. One commenter pointed to a flashy conference talk about AI-powered COBOL migration and basically said, sure, everyone sounds confident in the middle of the demo—until reality shows up with edge cases and broken workflows. Another went straight for the jugular with the line that summed up the vibe: maybe fix the bugs before replacing the "average COBOL dev" with the "average Java dev." Ouch. The biggest side-eye came from people noting the scale problem: the paper's largest example was about 4,000 lines, while real government and banking systems can be hundreds of thousands of lines each. Translation: the internet is not ready to hand over the mainframe keys just because three smaller tests looked good.
And then came the practical panic. How do you migrate messaging? What about hidden weirdness? One commenter argued the real danger isn't preserving old bugs—it's inventing shiny new ones, because AI tools can be unpredictable. The crowd's verdict: promising research, but nobody's canceling the veteran COBOL folks just yet.
Key Points
- •The paper focuses on validating AI-assisted migration of legacy COBOL programs to Java, a process complicated by limited test data and hard-to-check corner cases.
- •It proposes an agentic test-synthesis method called the Locksmith Loop that runs both COBOL and generated Java in instrumented runtime environments off-mainframe.
- •The method uses Witness Search and parity-preserving mutations to explore additional branches and identify blocked execution conditions called Locked Paragraphs.
- •The evaluation covered three COBOL-Java case studies, including two open-source programs and one internal production-like program ranging from 430 to 4,114 source lines.
- •Reported results showed nearly complete coverage on the two open-source programs, 91.90% branch coverage on the internal program, and deterministic parity between Java and COBOL in all accepted test cases.