Gordon Burtch study finds AI coding tools standardize syntax, not semantics

A new study by Gordon Burtch, published on arXiv in July 2026, puts numbers behind a long-standing fear: are AI coding assistants making all developers write the same code? Using a massive dataset of Kaggle contest submissions from 2019 to mid-2026, Burtch measured homogenization at two levels — individual submission similarity and contest-level conceptual span.

The results are nuanced. First, there's a clear convergence toward using the number 42 as a random seed, a nod to Douglas Adams' 'Hitchhiker's Guide' that AI models have enthusiastically adopted. More importantly, syntactic homogenization is real: TF-IDF vector comparisons show submissions have become significantly more alike in literal code structure and surface syntax. The latent dimensionality of syntactic variation has also narrowed, meaning developers are literally writing more similar-looking code.

Yet the study's most striking finding is what hasn't changed. Semantic similarity — measured using Voyage 3 code embeddings, which capture code intent and problem-solving strategy — has remained essentially flat over the entire period. In fact, Burtch finds some evidence that the conceptual span of approaches has even modestly expanded. This suggests AI assistants are standardizing boilerplate and implementation details without narrowing the creative problem-solving space.

The implications are twofold: developers can use AI to speed up routine coding without fear of losing innovation, but teams should be aware that codebases may become syntactically monotonous. The study also raises important questions about how we measure code originality and whether current AI models might eventually converge on semantic solutions if they continue to train on their own outputs.