The biggest problem is not that bugs are migrated with COBOL, but that lots of new bugs are going to be introduced. AI is not deterministic, it will be making tons of mistakes. The only realistic low-error approach is incremental step-by-step migration using Cursor or similar tools. However, it requires much more time as each step must be prompted, tested and committed manually. Any hope that one-shot migraton of a large code base will not introduce enormous number of bugs is very naive. LLM is very bad on handling long context - it is their nature unfortunately. There is no answer to this problem yet.
> The only realistic low-error approach is incremental step-by-step migration using Cursor or similar tools.
Correction: AI is not deterministic, the only realistic low-error solution is not a more complex use of non-deterministic AI, but deterministic transpilation.
The problem is that this results in COBOL-in-Java which runs correctly but it is a nightmare to maintain.
> AI is not deterministic, it will be making tons of mistakes.
From the paper:
> The COBOL source is passed through an internal deterministic Migrator to produce a generated Java target.
Also, humans are not deterministic either. Give the same COBOL -> Java translation to multiple developers and each will come up with a different solution. Heck, even the same developer will produce a different output for the same task, depending on the day of the week.
> AI is not deterministic, it will be making tons of mistakes.
Just set the sampling temperature to zero and remove any unintended non-determinism during the parallel computation of the token probability distribution. The problem is solved? Of course, not. Non-determinism has little to do with LLMs' mistakes.
You don’t need the entire codebase in context in every moment to migrate it.
Also AI being non deterministic does not prevent it from one-shotting perfect solutions 100% of the time for simple enough problems.
And every model generation brings this bar higher. So that’s really not a fundamental problem.
And we can also implement llm inference deterministically if we want, it’s just that it’s not worth the loss in performance to do it.
How does ai implement all the not Cobol parts a Cobol program rely on? Job Contol, CICS, sort processors? Cobol and mainframe technologies are non existant in java on any modern machine
I remember a similar story shared this year at JAX2026 from the Sparkasse Group, they said they were using AI to migrate from COBOL, but they still were in the middle of the migration. Maybe they faced the same issues / problems? It seemed pretty zealous to me, that everything was working smoothly, but this article highlights the limitations
300 to 4000 lines of "production like" (whatever that is) cobol code which is easily ported to a non-mainframe env. Our's sometimes uses assembler in its innards, so good luck with real legacy code spanning a dozen files and 50k loc...
I recently threw in (want to check those intelligence metrics!) some real production code into a non-agentic system (just to get a feel how things perform without a custom harness) and results where ... interesting. The particular program uses some preprocessor no LLM we have access to (newest was GPT 5.5) has any clue about - so they confabulate what it could do (Gemini 2.5 didn't even notice there was a preprocessor...). This is expected of course but it somehow seems the problem of this technology that unless you feed it masses of data or mechanically break up the tasks in rote subunits, it just doesn't do anything sensible still...
Anything running on COBOL to day is a large enterprise system. You'll have reports running in other systems in subsidiary companies that rely on bugs in the upstream cobol code.
Yeah. Bug-for-bug migration is a real thing in large code-bases in the industry. You want to replicate all behavior of the code regardless whether the behavior is a feature or a bug.
COBOL is not about the language, it's about the whole environment in the mainframe. LLM fanboys won't understand this. You need something like a mainframe with a 99.99% uptime no matter what happens in hardware, with live CPU swapping and such.
Yes and that mainframe runs several other technologies besides COBOL, and more importantly, that don't depend on it (COBOL fanboys won't understand this ;) )
Correction: AI is not deterministic, the only realistic low-error solution is not a more complex use of non-deterministic AI, but deterministic transpilation.
The problem is that this results in COBOL-in-Java which runs correctly but it is a nightmare to maintain.
From the paper:
> The COBOL source is passed through an internal deterministic Migrator to produce a generated Java target.
Also, humans are not deterministic either. Give the same COBOL -> Java translation to multiple developers and each will come up with a different solution. Heck, even the same developer will produce a different output for the same task, depending on the day of the week.
Just set the sampling temperature to zero and remove any unintended non-determinism during the parallel computation of the token probability distribution. The problem is solved? Of course, not. Non-determinism has little to do with LLMs' mistakes.
At least if tescoverage is good, but well... That's something llms can also be used for
https://www.pega.com/insights/resources/break-free-legacy-ma...
https://www.ibm.com/products/watsonx-code-assistant-z
https://global.fujitsu/en-global/pr/news/2026/03/30-01
https://www.rocketsoftware.com/en-us/insights/ai-powered-cob...
There are hundreds of billions of lines of cobol in production.
The IRS alone has approx 160 cobol programs, averaging 230kloc each.
As specified, please find 99997 correct parts and the 3 defects (do not use)
I recently threw in (want to check those intelligence metrics!) some real production code into a non-agentic system (just to get a feel how things perform without a custom harness) and results where ... interesting. The particular program uses some preprocessor no LLM we have access to (newest was GPT 5.5) has any clue about - so they confabulate what it could do (Gemini 2.5 didn't even notice there was a preprocessor...). This is expected of course but it somehow seems the problem of this technology that unless you feed it masses of data or mechanically break up the tasks in rote subunits, it just doesn't do anything sensible still...
COBOL isn't hard, the tooling around it on old systems are a pain though.
See Hyrum's Law: https://www.hyrumslaw.com/