What do the IRS, banks, airlines, and authorities infrastructure about the earth have in common? They’re all operating on systems propped up by ancient device code that nobody dares touch, lest everything arrive crashing down.
OpenAI thinks it may have a solution. On October 19th, the business touted “substantial improvements” in its latest AI models’ capability to migrate aging application written in older programming languages into contemporary code, which it claims could assistance organizations alongside “maintaining and modernizing older systems.”
That’s a possibly enormous deal. COBOL, among the languages highlighted by OpenAI, remains stubbornly embedded in huge portions of the world’s crucial infrastructure. And notwithstanding decades of efforts to banish it to history, the tongue has limped on, merely refusing to die.
“Nobody in their correct intellect would use COBOL for a new betterment project.”
It is difficult to overstate COBOL’s ubiquity. Short for “common business-oriented language,” COBOL came into being in 1959. Nearly seven decades later, it is motionless everywhere. According to IBM, COBOL is the basis for additional than 40 percent of online banking systems, supports 80 percent of in-person credit cardstock transactions, and handles 95 percent of ATM transactions. It additionally underpins many authorities systems, from Social Security to the IRS to services run by most states. One coarse estimate puts the annual disbursal of COBOL’s inefficiencies to the US GDP at about $105 billion.
But replacing COBOL is fiendishly difficult. The identical goes for another bequest programming languages, akin Fortran, still in use today for everything from specialized computing and climate forecasting to atmospheric and sea studies and codes crucial to nuclear security. These systems are frequently too crucial to hazard breaking, and many of the programmers who comprehend them have lengthy since former or died.
For all their notable advances in modern software development, AI coding tools have struggled to create much headway alongside these bequest systems. But latest developments propose that may be changing.
In its latest blog post, OpenAI reported promising results migrating COBOL code to additional contemporary programming languages, including C++, Java, Kotlin, Python, and Rust, without changing what the first application does. That final bit is crucial for organizations looking to ditch code akin COBOL, and OpenAI acknowledges that “a immigration is considered prosperous lone if all tests pass.”
In OpenAI’s own tests, GPT-6.1 Sol achieved 87 percent behavior conservation accuracy at low reasoning attempt — how many computational steps the example spends analyzing a issue before giving an answer — climbing to 98 percent at higher reasoning. By comparison, at low reasoning effort, GPT-6 Astra scored about 86 percent, GPT-6 Sol fair 21 percent, and GPT-5.6 Sol a uncomplicated 5 percent.
“Rewriting code is not the identical as replacing a system.”
“These evaluation results display advancement toward additional accurate, affordable code immigration for teams modernizing the application their businesses depend on,” OpenAI wrote.
The business additionally reported improvements on its models’ capability to build repository application in assorted programming languages, including COBOL.
Anthropic additionally sees an chance in bequest systems. In February, it published a step-by-step “code modernization playbook” and pitched its coding agent, Claude Code, as a way to migrate COBOL code into contemporary languages akin Java or Python. It offered no achievement benchmarks showing how fine Claude Code could execute specified migrations, but declared that “the economics of COBOL modernization have shifted.” Shares in IBM, which derives part of its income assisting companies keep bequest systems, nosedived following Anthropic’s announcement.
Experts who said to The Verge accepted that AI tools have made strides in tackling bequest code in latest years. But they cautioned that advances trumpeted by OpenAI this week are a extremely lengthy way from making it imaginable to safely substitute crucial bequest systems built about it.
OpenAI spokesman Oscar Haines acknowledged in an email to The Verge that bequest code “is motionless far from a solved problem” and stated the business is listening to feedback from developers and enterprises and exploring ways to build improved tests and continue improving its models’ understanding of coding languages. Anthropic did not provision comment in period for publication.
“The evaluations shown by OpenAI don’t location the key difficulty: Breaking up a complex bequest scheme into smaller units that can be replaced individually.”
“Rewriting code is not the identical as replacing a system,” stated Vadim Zaytsev, an affiliate prof of application development at the University of Twente in the Netherlands. Systems do have a lot of code — Zaytsev stated the largest scheme he’d seen had 250 myriad lines of COBOL, “definitely not the world’s largest” — but that’s lone part of the problem. Code may be written in multiple languages and there’s data, possibly stored in assorted formats, to consider, he said. There are additionally specialized integrations alongside another systems and the expectations of the group and organizations to withstand in mind.
All that complexity method “AI is not playing a important function at the instant in replacing or migrating bequest COBOL systems,” Zaytsev said.
OpenAI’s evaluations don’t grasp this complex actuality either, Arie van Deursen, a application engineering prof at Delft University of Technology in the Netherlands, told The Verge. He described the company’s blog as “superficial,” noting it offered few particulars concerning the immigration task, during its tongue benchmark tested models on a comparatively uncomplicated and well-defined project involving systems akin to those they would have encountered in training. He additionally stressed the huge gap between maintaining an aging scheme and modernizing it. “Maintaining is of way a lot easier than modernizing,” van Deursen said, calling the tasks OpenAI used to measure its models “lightyears” distant from genuine modernization. “The evaluations shown by OpenAI don’t location the key difficulty: Breaking up a complex bequest scheme into smaller units that can be replaced individually.”
The biggest difficulty isn’t necessarily rewriting code. “The chief difficulty is certainly not understanding how the scheme works,” stated Alastair Donaldson, a prof of programming languages at Imperial College London. Such a deficiency of understanding is a “very common” issue for bequest systems using languages akin COBOL, he said, adding that extremely few group cognize how to program in them anymore. And during it is everywhere, Donaldson stated “nobody in their correct intellect would use COBOL for a new betterment project.”
“In many cases equal insignificant differences in scheme behavior before and following immigration are unacceptable.”
That makes sturdy evaluation crucial, Donaldson said. One method could be to run the identical workflows through the old and new systems to inspect they behave identically. AI could additionally document its code translation, ideally allowing humans acquainted alongside the bequest tongue to assessment the work. Formal verification, which uses mathematical techniques to computationally verify that code behaves correctly, is another option, stated Jacopo Mauro, a math and device discipline prof at the University of Southern Denmark.
Such methods are already used to verify AI-generated mathematical results, but applying them to bequest code would necessitate a “huge amount of work,” Mauro said. Using those methods would average code could be checked automatically, Mauro said, adding that “having this volition be improved than having a individual inspect that the translation is correct.” Until then, he said, “human verification remains essential.”
And whenever the systems engaged are as crucial as those operating on COBOL, experts say you need to be entirely certain the replacement works. “In many cases equal insignificant differences in scheme behavior before and following immigration are unacceptable,” Zaytsev said. While AI can certainly provision funds and speed things up, the complexity of bequest systems method partial automation isn’t enough.
“Suppose you have a magic AI wand that handles 99 percent of your code,” Zaytsev said. “What do you do alongside the remaining 1 percent? That 1 percent can be a myriad lines of code, dispersed all complete the location and subtly failing in additional ways you can ideate inside your lifetime.”
Follow topics and authors from this narrative to see additional akin this in your personalized homepage nourish and to obtain email updates.
