AI-Written Software Beat Human Code in a Reliability Test

A team of computer scientists wanted an honest answer to a big marketing claim: is software written by AI actually more reliable than software written by people? So they ran a fair test. They took ten well-known Linux utility programs — the small, everyday tools that keep computers running — and had AI agents (AI that can take actions on its own) recreate them from scratch using a standard professional workflow.

Then they attacked both versions with fuzz testing, which means feeding a program millions of random, messy inputs to see if it crashes. This is how security researchers find hidden bugs. The AI-written programs held up better than expected: they were typically as reliable as the human versions, and often more reliable. They also had fewer memory errors — the kind of flaw, like a buffer overflow, that hackers can exploit to break into systems.

But there was a trade-off. The AI code was more likely to hang, meaning it froze in an endless loop instead of finishing its job. On your laptop, that's an app that spins forever and needs force-quitting. In a hospital or bank system, that's a service that quietly stops responding.

The study's other big finding is about people, not machines. Getting good code out of AI required careful human supervision — the prompts (instructions) people wrote and how they reacted to mistakes made a huge difference. The researchers also noticed the written instructions could double as documentation, helping future developers maintain the software affordably. In short: AI can code well, but it still needs a skilled human in charge.