Findings

Measurements too small for a paper and too specific for an essay. Each one states what it found, on what data, and what it does not support. The caveats live here with the finding rather than in a reply somewhere, because a chart travels and its footnotes do not.

Dialect drift in AI-signed code review

Where AI-signed prose sits on the grammatical axis between US-located and India-located professional English, and which way each model is moving. Drawn from 153,448 GitHub comments, 2022 to 2026, parsed for construction rather than vocabulary and matched on repository and year.

Figure 1 Position on the US → India axis, 2025 → Opus 5 era In every panel the US human rate sits at 0 and the India-located human rate at 100, with the corridor between them shaded. Up to the India line all panels share one scale. Above the line each panel compresses its own scale to fit; read the labeled values, and do not compare distances across panels except inside the shaded band. An arrow runs from a cohort's 2025 position to its Opus 5-era position (Aug 12–27, 2026). The small dots are the windows in between (Jan 1–Jun 8, and the Fable era Jul 1–Aug 11). Jun 9–30, the release-and-suspension window, is excluded.
Figure 2 Raw trajectories against the human baseline bands The four habits plus the two supporting measures, in native units per panel. Shaded bands are the pre-2025 human baselines ±2 standard errors; dashed where no interval was computed; open markers flag cells under 100 texts. Scales and directions follow Figure 1: India side up ("axis reversed" where that requires flipping), compression past the India band at the break mark.

Reading and caveats

The four Figure 1 metrics are the constructions that most separate India-located from US-located professional GitHub prose in 2022–24 (all p < 10−30, repo- and year-matched samples of ~4,800 comments per side): verbless status sentences, low clause subordination, triple-noun compounds and please/kindly politeness marking.

Claude-signed text enters 2025 past the India baseline on all four at once and drifts back toward the human range through 2026. Copilot shares the noun compression but keeps US-style subordination and has converged to US politeness. A third cohort of 370 AI-signed comments whose signature names no model (chiefly Claude Code PR footers with the link text stripped at harvest) is excluded from the charts: its dialect position does not track the Claude cohort, and whether that reflects different authorship or human editing of generated text cannot be determined from the corpus.

Sample sizes are honest but thin at the edges: Claude cells run 51–304 texts, Copilot cells 12–47, so read Copilot lines as sketch, not measurement. AI-signed text is largely PR-body prose while the baselines are conversation comments; that genre gap inflates verblessness and noun stacking for every AI cohort, though it cannot manufacture the politeness or subordination signatures. Parsing is spaCy en_core_web_sm; the parser is trained on US-style English, which may flatten Indian-text parses slightly.

Data table — every value plotted

Unpublished note, August 2026. Not pre-registered. It measures how text is built, not who wrote it, and it makes no claim about any group of people.