AI Now Writing Code That Humans Can’t Even Understand

As the use of generative artificial intelligence in the workplace grows, programmers are increasingly tasked with monitoring code produced by large language models. Some engineers are no longer able to comprehend AI-generated code, according to Jeremy Nixon, founder of chip optimization software company Infinity. He points out that researchers have found AI use can lead to cognitive decline and skill atrophy among computer scientists.

Developers are worried that reliance on AI for code generation is causing them to drift away from writing code and learning new skills needed to create larger projects. Even professionals at leading AI labs are alarmed by colleagues struggling to understand AI outputs. An interview with SemiAnalysis revealed that OpenAI engineers were baffled by code they reviewed for DeepSeek's graphics processing unit kernel.

Independent researcher Jordan Nanos noted that while AI engineers understand hardware concepts, they have no idea what the AI-generated code actually does. He believes humans may not need to deeply analyze the code because AI can manipulate data and processing elements effectively. However, this raises concerns about errors and hallucinations slipping through oversight cracks.

Amazon recently implemented a 90-day "code safety reset" after experiencing outages caused by AI-assisted code changes. Software company Undo's analysis found 35% of AI-generated code is not fully understood before production, and 29% of engineers witnessed productivity losses due to "unpicking" AI-generated code. Many engineers spend significant time deciphering AI code, prompting questions about the technology's worth.

Some researchers argue this is a paradigm shift, with AI handling architecture and correctness criteria, but will the optimism persist as software development becomes unrecognizable? There are concerns that younger engineers may grow up undervaluing code understanding in this environment.