Vericoding: The End of "Trust Me Bro, The AI Wrote It" (blog.icme.io)

GitHub reports that 46% of new code is now AI-generated, with startups in Y Combinator's Winter 2025 cohort showcasing codebases where 21% of the code is 91% AI-created. Despite the efficiency of AI in code generation, a concerning trend has emerged: AI-generated code exhibits 2.74 times more security vulnerabilities than human-written code, and nearly half of it fails security tests. This raises significant alarms within the AI/ML community as scalable automated code generation outpaces the capability to verify that code, leading to increased technical debt and security risks, as seen in recent breaches. To address this challenge, researchers from the Beneficial AI Foundation introduced "vericoding," a technique for producing formally verified code based on mathematical proofs rather than simply relying on testing. This method contrasts sharply with "vibe coding," allowing developers to generate code that is directly proven against specifications, ensuring correctness. Early benchmarks indicate an 82% success rate using existing large language models for vericoding. The potential of this approach is demonstrated through its application in companies like AWS, which has successfully implemented verification in their systems. As the tech landscape shifts towards AI-generated code, the push for vericoding could provide the compliance and safety assurances necessary to bridge the gap between automation and security, enabling regulated industries to participate in the AI coding revolution.