A GitHub scraper just measured how much of human writing now reads like Claude

Abraham clustered eight distinct ways of writing found in GitHub pull request descriptions using KL-divergence k-means, then tracked which cluster each new PR belonged to over 85 complete weeks, January 6, 2025 through August 17, 2026. The dataset: 47,464 descriptions pulled from a randomized five-minute daily sampling window, after excluding bot and app accounts and capping any one author at three descriptions a week. One cluster, headlined by "load-bearing" and also carrying "seam," "quietly," "genuine," and "latent," made up 1.0 percent of the corpus at the start of 2025. It is 45 percent of it now, still climbing at roughly 0.8 percentage points a week. "Load-bearing" itself shows up 123 times more often inside that cluster than outside it.

Abraham published his own researcher degrees of freedom rather than burying them: two parameters, the cluster count and the minimum word-frequency floor, were, in his words, "chosen on the outcome," meaning tuned so the headline word would land where it does. Run the same method unconstrained across 32 independent fits of the data, and the cluster still shows up in 21 of them, with "load-bearing" ranking in the top five of its words in 29 of 32 and the top forty in all 32. The exact 45 percent figure comes from one fit, but the pattern appears across the unconstrained runs.

The Hacker News discussion had 110 points and 50 comments about seven hours after posting at this reading. Several threads argued about whether the shift reflects models training on their own RLHF-shaped output, and at least one commenter noted catching themselves adopting the same vocabulary in Slack conversations with colleagues.

Why it matters: This corpus has argued for months that AI-generated contributions are eroding the effort signal reviewers rely on to judge a pull request. This is the first attempt to measure that erosion instead of asserting it, and Abraham's transparency about which parameters he tuned to land his exact number is more convincing than the number itself: run the method blind and the same cluster keeps turning up anyway.