Discussions around AI and financial infrastructure have drawn attention in recent weeks, particularly as models are now smart enough to go far beyond chatboxes.
Sam Altman, who leads OpenAI, has long described a future in which advanced systems handle complex tasks on behalf of people, from research and software development to everyday interactions.
He has also co-founded Tools for Humanity, the organization behind World, arguing that as AI grows more capable it will become harder to tell whether an action online comes from a person or an automated agent.
The two efforts remain separate companies, yet they address related concerns about how institutions and platforms will confirm human involvement when agents can browse, shop, approve, or transact.
On October 19, OpenAI published details of stronger performance from its latest models on tasks involving languages such as COBOL, Fortran, and Visual Basic.
These languages receive far less public code and documentation for training than more widely used options like Python or JavaScript, yet they continue to support significant volumes of transactions and operational logic at banks, governments, and other organizations.
COBOL in particular still underlies large shares of online banking systems, in-person card transactions, and ATM activity.
The company introduced two evaluations.
MultiLangBench asks a model to implement a SQLite-compatible database engine from scratch across twenty languages, measuring how well it organizes code, manages state, and preserves functionality under tests derived from SQLite's own suite.
Results reported for GPT-6 Astra showed an average test pass rate of 87.4%, compared with 26.8% for an earlier model, with gains across every language tested and scores above 83% in nineteen of the twenty. A lower-cost model reached 40.3% at substantially reduced estimated API cost. A separate evaluation focused on migrating COBOL programs to languages including C++, Java, Kotlin, Python, or Rust while requiring byte-for-byte output parity on hidden tests. GPT-6.1 Sol reached 86.6% accuracy at lower reasoning effort versus 5.4% for the prior model, and 97.9% at higher effort. OpenAI released the MultiLangBench materials on GitHub and noted that production use of such migrations still requires expert review against actual business rules.
OpenAI has also published guidance on using its coding agent Codex for modernization work, including an example built around a COBOL investment portfolio system.
The approach emphasizes structured planning documents, inventory of existing programs and data flows, design of target architectures, and parity testing so that observed behavior remains consistent. These materials position the tools as aids for engineers who retain responsibility for architecture and validation.
Around the same period World announced a suite called World ID for banking and finance at the Money20/20 conference.
The offering combines cryptographic proofs that a person is a unique human, that they are present at the moment of verification, and that specific attributes such as age or nationality hold, without necessarily sharing underlying documents.
Design partners listed include firms in investment and financial services.
The tools can be applied to requirements such as approving wire transfers or recovering accounts, and a beta component allows a verified person to delegate certain proofs to an AI agent so the agent can present them to a third party.
World also made a Zoom integration generally available for confirming that video call participants match prior Orb-based verification.
The project cites risks from deepfakes and agent-driven fraud as motivation, including a 2024 case in which an employee wired roughly 25.6 million dollars after a video call in which every other participant was a deepfake, along with a Deloitte forecast that AI-enabled fraud losses in the United States could reach 40 billion dollars by 2027.
No formal joint product has been released that combines OpenAI's models directly with World ID at scale.
Community-developed connectors exist that allow assistants to interact with World ID flows, and reports have mentioned internal discussions about potential uses, but public offerings continue independently.
OpenAI's finance-related features include read-only account analysis through third-party connections and agentic commerce partnerships that keep authorization controls with users, while World focuses on the verification layer institutions can adopt.
More than 18 million people hold a Proof of Human, the foundation of the real human network World describes.
These developments occur against the fact that systems still rely on decades-old code for core operations and rising interest in confirming human involvement in high-stakes digital actions.
