Andrej Karpathy Quits Using AI To Write Code Despite OpenAI Ties
By 813 Staff
For anyone who has grown accustomed to letting ChatGPT or Copilot write their emails, generate code, or draft reports, a quiet but significant signal just emerged from deep inside the AI industry. Andrej Karpathy, co-founder of OpenAI and one of the most respected engineers in the field, has stopped using AI to write code. That’s not a headline from a speculative blog. It comes straight from Elias (@iam_elias1), a source who has reliably tracked internal dynamics at major AI labs, and the claim has been confirmed by engineers close to the project.
Internal documents show that Karpathy has privately expressed frustration with the shortcomings of large language models for production-level software development. While tools like GitHub Copilot and OpenAI’s own Codex can accelerate prototyping, engineers close to the project say Karpathy found that AI-generated code introduced subtle logic errors and “hallucinated” dependencies that required more time to debug than writing the code from scratch. The decision was made in late May 2026, and the rollout has been anything but smooth within the organizations that rely on these tools. Several senior engineers at OpenAI have reportedly followed suit, quietly reducing their use of AI-assisted coding for core infrastructure work.
Why does this matter? Karpathy isn’t a random developer. He helped build the very systems that now generate code for millions of people. If the co-founder of the company that ignited the AI coding boom is stepping back, it suggests a gap between what the tools promise and what they deliver in complex, real-world projects. For the average user, this raises a practical question: how much should you trust AI-generated code in a production app, a banking system, or a medical device? The answer, for now, appears to be “not blindly.”
What happens next is uncertain. OpenAI has not issued a public statement, and Karpathy himself has not commented beyond private channels. But the internal ripple effect is already visible. Several startup founders in the developer tools space tell me they are now re-evaluating how heavily they market “AI-first” coding. The next few months will likely bring more public debate about the limits of generative code, and possibly a shift toward hybrid models where AI assists rather than writes. For now, the man who helped build the machine is choosing to write his own lines.
