Measuring convergence against pre-LLM baselines
Parent: Human adoption of LLM vocabulary and its effect on tic baselines · Published reference · snapshot 2026-09-23
↓ Facts as markdownall context files
Source-anchored research for Measuring convergence against pre-LLM baselines. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
How it works
- Corpus check on the International Corpus of English found no evidence that focal words like delve are especially prevalent in any English variety, contradicting the Nigerian-English/RLHF-rater-origin explanation; RLHF itself remains a candidate cause. (medium confidence) [source]
- A preregistered experiment (496 participants) showed brief chatbot interaction made people adopt its vocabulary, giving a causal mechanism beyond correlational baseline comparisons. (medium confidence) [source]
Measurements and reference values
- Marker words are not stable baselines: per search summaries of the Yakura paper (v3), delve later declined below its pre-ChatGPT baseline in some categories (Science & Tech -15%, Education -11%, Business -30%), so a static pre-LLM frequency list can mislead. (low confidence) (verified as of 2026-09-23) [source]
Comparisons and alternatives
- Kobak et al. measure LLM vocabulary influence as excess word frequency in 15M+ PubMed abstracts (2010-2024) versus pre-LLM years, an approach modelled on excess-mortality studies; they report a lower bound of at least 13.5% of 2024 abstracts LLM-processed, up to 40% in some subcorpora. (high confidence) [source]
- Yakura et al. use a synthetic-control method on 737,083 hours of podcast speech (824,634 episodes) plus YouTube academic talks: treated words like delve are compared with a weighted mix of donor words sharing pre-ChatGPT trends, and show an abrupt post-release rise. (high confidence) [source]
Facts and statements
- Juzek and Ward take PubMed 2020 as the pre-LLM reference (LLMs not widespread before 2021), compare it with 2024, and define 21 focal words by three criteria: significant 2020-2024 rise, unexplained spike, and ChatGPT overuse relative to human abstracts. (medium confidence) [source]
Children
- No children recorded.