Popular AI Code Isn't Better, Study of 265,000 Notebooks Finds

If you've ever copied a popular recipe online only to find it tastes mediocre, you already understand this study. Researchers looked at 265,363 public AI projects posted on Kaggle, a site where data scientists compete to build the best prediction models. They wanted to test a common assumption: that clean, well-written code produces better AI results. It turns out that assumption is only half true.

They measured two kinds of "code quality." The first was general tidiness — things like clear naming and neat formatting, the coding equivalent of good grammar. The second was AI-specific care — things like handling data properly and using models correctly, more like following food-safety rules in a kitchen. The tidiness had essentially zero connection to how well the AI performed. But the AI-specific mistakes mattered: projects that cut those corners consistently did a little worse.

The team also checked whether the crowd's favorite projects were actually better. They weren't. A project's popularity gave no clue about its quality or its results. Neither did the author's general coding reputation. The one thing that did predict better results was experience in competitions specifically — people who had competed a lot scored higher and made fewer AI-specific mistakes, even though their code was slightly messier in other ways.

The takeaway is practical. If you're hiring someone, buying an AI tool, or copying code from the internet, don't trust stars, likes, or job titles as proof of quality. Look instead at whether the work handles data and models carefully — that's the part that actually shows up in results. General polish is nice, but it's mostly decoration.