The observation
Louis Abraham collected public GitHub pull requests that name Claude as the author and grouped the descriptions by the exact words they contain. The result is a set of eight fixed sentence patterns that account for every example in the collection. No description falls outside these patterns, and the patterns themselves show no dependence on the code changes they accompany.
How the data was assembled
The underlying site pulls every qualifying pull request description from public repositories. Abraham then clustered the text by lexical form rather than by topic or intent. Early samples from the start of 2025 showed the eight patterns present in roughly equal shares. By the middle of 2026 one pattern had grown from 1.0 percent of the total to 45 percent while the remaining seven stayed flat or declined. The site makes the raw clusters available for download so others can verify the same distribution in their own data.
Technical detail
The eight templates differ only in their opening and closing clauses; the middle sentences are interchangeable across examples. None of the templates reference specific functions, files, or diff content. Abraham notes that the repetition occurs even when users give no explicit instruction to use stock language. The same convergence appears across unrelated repositories and programming languages.
The project readme connects the finding to watermarking research. It cites Google’s SynthID-Text paper, which states that its detection method produces “some reduction to inter-response diversity.” Anthropic has said it intends to adopt the same watermarking approach for Claude. Abraham observes that the existing loss of variety predates any watermarking layer and will likely be amplified once the scheme is deployed.
Reactions
The Hacker News thread and Daring Fireball post both treat the result as evidence of a broader trend rather than an isolated quirk. Commenters point out that the eight patterns function as load-bearing vocabulary: they carry the surface appearance of a description while conveying almost no information about the actual change. No counter-analysis has yet appeared that disputes the clustering method or the reported percentages.
Why it matters
Teams that accept Claude-generated pull request text now receive descriptions that are increasingly interchangeable. Reviewers who once skimmed the text for intent must now treat it as decorative and look elsewhere for substance. The shift also changes the baseline for any future watermarking system. Because the model already favors a narrow set of phrases, adding a technique that further reduces diversity will start from an already constrained state rather than from varied output.
The data set itself supplies a concrete measurement. Anyone can download the clusters and test whether the same eight patterns appear in their own Claude-assisted work. The finding therefore moves the discussion from anecdotal complaints about repetitive AI text to a verifiable, growing concentration on a handful of fixed expressions.
---
Sources:
{"word_count": 612, "sources_used": ["https://louisabraham.github.io/load-bearing/"]}
No comments yet