The IRS still runs on code from the Nixon administration, and now OpenAI wants to rewrite it with ChatGPT.
The Summary
- OpenAI announced October 19th that its latest models show "substantial improvements" in migrating legacy code like COBOL to modern languages, potentially solving a crisis that's kept banks, airlines, and government systems frozen in the 1970s.
- COBOL still powers critical infrastructure worldwide, but the programmers who understand it are retiring faster than replacements can be trained.
- If AI can actually do this reliably, it unlocks trillions in stuck modernization budget and eliminates one of the strangest technical-debt hostage situations in computing history.
The Signal
The COBOL problem isn't about nostalgia. It's about risk. Systems that handle Social Security payments, airline reservations, and ATM transactions run on code written when Watergate was still unfolding. Nobody wants to touch it because one wrong keystroke could brick a system that processes billions in transactions daily. The few developers who truly understand COBOL are in their 60s and 70s. When they retire, they take institutional knowledge that can't be Googled.
This creates a bizarre market dynamic. COBOL programmers command premium rates because scarcity, but organizations still won't migrate because the migration risk feels existential. So the technical debt compounds. Every year these systems run makes them harder to replace. Every patch adds another layer of fragility.
"The IRS, banks, airlines, and government infrastructure are all running on systems propped up by antique computer code that nobody dares touch."
OpenAI's pitch is that language models trained on billions of lines of code across dozens of languages can understand COBOL's logic and translate it into Python, Java, or JavaScript. The models don't need to have written COBOL professionally. They just need to recognize patterns. COBOL's verbosity actually helps here. It was designed to be readable by non-programmers, which means it's relatively explicit about what it does.
But there's a canyon between "substantial improvements" and "production ready." Code migration isn't just translation. It's archaeology. Legacy systems contain decades of undocumented business logic, workarounds for bugs that no longer exist, and edge cases that only surface once a decade. A human COBOL expert doesn't just read the code. They understand the context, the constraints of 1970s hardware, the tribal knowledge about why certain things were done certain ways.
Key unknowns:
- Can AI models preserve the edge-case handling that keeps these systems stable under weird conditions?
- Who takes liability when AI-migrated code fails at 2am on a Saturday during tax season?
- How do you test that translated code does exactly what 50-year-old code did, especially when the original requirements documentation is lost?
The economic implications are staggering. Governments and Fortune 500 companies have postponed modernization for decades because migration projects routinely fail. They go over budget, miss deadlines, or produce systems that work differently enough to break downstream dependencies. If AI can reduce that risk even 30%, it unlocks modernization budgets that have been frozen since the Obama administration.
The Implication
Watch how OpenAI prices this. If they're positioning it as a developer tool, it's incremental. If they're going after enterprise contracts to actually execute migrations, it's a services play that could reshape how infrastructure gets built. The real test isn't whether AI can translate COBOL. It's whether organizations trust it enough to bet their payment systems on it. That trust either gets built through small, boring, successful projects, or it gets destroyed by one spectacular failure that makes headlines.
The deeper shift is what this says about AI's expanding role in technical debt. Code migration is unglamorous work that smart engineers avoid because there's no glory in it. If AI can handle the tedious, high-risk parts of legacy modernization, it frees up human developers to build new things instead of maintaining old things. That's the real unlock, assuming the AI doesn't break anything important first.