Hillel Wayne noticed something odd while working on AI-generated TLA+ specifications earlier this year. Every one of them used the word "spine" somewhere. He started seeing it in non-TLA+ projects too and decided to check whether it was a new LLM signature.

The approach was straightforward: search GitHub for public pull requests containing specific words in their titles and compare year-over-year growth. The first nine months of 2026 produced 20 times more PRs with "spine" in the title than all of 2025. When Wayne normalized against total public PRs, which grew roughly 1.5x over the same period, the spike in "spine" usage stood out as abnormal.

Which models are responsible

Wayne attempted to pin the trend on a single model. GPT 5.5 shipped in April, and Opus 4.8 arrived in May. But the "spine" count was already climbing at 1.5x per month before April, and Wayne recalled seeing the word in AI output as early as February. Spot-checking the PRs revealed many were co-authored by Claude or Cursor, not exclusively GPT. The pattern appears to cut across models rather than originating from one provider.

GitHub data shows 71.7 million more LLM-authored pull requests this year than last year, covering both public and private repositories. That volume means agentic coding activity on GitHub may now exceed human activity. In that context, a linguistic quirk in model output becomes visible across thousands of projects.

Other words with unusual growth

"Spine" is not the only word showing a discontinuous jump. Wayne checked several candidates:

  • "Gate" appears in one out of every 300 PR titles. Including PR bodies pushes that to one in 26.
  • "Lane" shows a similar spike, which Wayne could not explain. CPU lanes was a possibility he could not confirm.
  • "Proof" has increased, though not as dramatically. Wayne attributed this partly to existing usage in "proof of work" contexts.
  • "Truth" shows a notable increase, consistent with models favoring the word in explanations and documentation.
  • "Seam", flagged by a coworker, also tracks upward in the data.

The word "ladder" appeared to spike in 2022 with over 30,000 PRs. Wayne investigated and found the cause was automated boto3 upgrades, which include "ABR package ladders" in their descriptions. Filtering to PR titles removed the artifact. He also checked "lamp," confirming the LLMs do not favor it over baseline developer usage.

What the vocabulary patterns reveal

The analysis has limits. GitHub Search cannot filter code samples by date, so Wayne used PR titles as a proxy. He did not have access to the GitHub archive data dump that would allow a more rigorous time-series analysis. The tooling for this work was built quickly as a small script.

What the data does show is that LLMs leave fingerprints in language. Models trained on large codebases converge on specific vocabulary choices that human developers use less frequently. When those models generate code at scale, the statistical preferences become visible across the ecosystem. The word "spine" in a TLA+ spec or a React component does not indicate a functional problem, but it does indicate which tool produced the code.

Wayne suggested that with more time, it should be possible to identify which words are particular to specific models. For now, the evidence points to a general tendency across frontier models rather than a quirk of any single one.