# Human adoption of LLM vocabulary and its effect on tic baselines > Source-anchored research for Human adoption of LLM vocabulary and its effect on tic baselines. Run status: COMPLETED. Parent: [Writing and Documentation](https://llms-explorer.com/tree/writing-and-documentation/) · 5 facets · 28 facts · page: https://llms-explorer.com/tree/ai-isms-human-convergence/ ## How it works - A preregistered experiment (496 participants) found brief chatbot interaction makes participants adopt specific words, persisting through distractor tasks, supporting exposure-to-active-vocabulary as the mechanism. (low confidence) — [source](https://arxiv.org/abs/2409.01754) - 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](https://arxiv.org/html/2412.11385v1) - 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](https://arxiv.org/abs/2409.01754) ## Measurements and reference values - Enriching TOEFL essay vocabulary with ChatGPT cut the false positive rate from 61.22% to 11.77%, showing detectors key on vocabulary and predictability, not authorship. (medium confidence) — [source](https://arxiv.org/html/2304.02819) - LLM-marker usage varies widely: about 20% in computational fields and in China/South Korea/Taiwan, about 5% in English-speaking countries; MDPI (21%) and Frontiers (20%) exceed Nature/Science (~7-10%). Measured rates are likely lower bounds. (low confidence) (verified as of 2025-07-01) — [source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12219543/) - Sourati et al. (USC) report declining writing-complexity variance on Reddit, news and academic preprints, and that LLM polishing/rewriting reduces that variance (reported 21-50% across datasets/models); the review cites this as evidence of stylistic homogenization in scientific writing. (medium confidence) — [source](https://arxiv.org/html/2508.01491v1) - Even when prompted for grammar-only edits, LLMs substantially alter meaning; in a user study heavy LLM use raised neutral-stance essays by nearly 70%, and AI-generated peer reviews scored about one point higher than expected and devalued clarity and significance. (low confidence) — [source](https://arxiv.org/abs/2603.18161) - 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](https://arxiv.org/abs/2409.01754) ## Problems, failure modes and limitations - Model collapse under a replace-data regime (each generation's synthetic data replaces the prior data) degrades models, but if synthetic data accumulate alongside the original real data, test error has a finite upper bound independent of iteration count; this held across language models, diffusion models and VAEs. (medium confidence) — [source](https://arxiv.org/abs/2404.01413) - The excess-word method cannot flag individual abstracts, cannot separate direct LLM use from stylistic borrowing, and cannot fully separate LLM effects from other emerging linguistic trends. (low confidence) — [source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12219543/) ## Comparisons and alternatives - Seven GPT detectors misclassified non-native English TOEFL essays as AI with 61.22% average false positive rate, versus 5.19% on US 8th-grade essays; the mechanism is perplexity, since lower lexical variability reads as machine-like. (high confidence) — [source](https://arxiv.org/html/2304.02819) - Excess-vocabulary analysis found 454 excess style words in 2024 biomedical abstracts, versus 190 at the 2021 COVID peak; the authors estimate at least 13.5% of 2024 abstracts were LLM-processed. (low confidence) (verified as of 2025-07-01) — [source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12219543/) - Cited evidence that people adopt LLM framing (Jakesch et al. 2023) and that AI-assisted ideation reduces semantic diversity (Anderson et al. 2024) supports the mechanism, but the review notes longitudinal effects and correlation-versus-causation gaps remain open. (low confidence) — [source](https://arxiv.org/html/2508.01491v1) - Rewriting experiments apply one LLM uniformly to all abstracts, so they give an upper bound on homogenization rather than real-world practice; effects are stronger for Reddit and academic writing than for news. (low confidence) — [source](https://arxiv.org/html/2508.01491v1) - Florida State University (Juzek et al.) analysed 22.1 million words of science/tech conversational podcasts; nearly three-quarters of target AI buzzwords rose after ChatGPT, some more than doubling. 'Underscore' rose while its synonym 'accentuate' did not, which points to AI-specific rather than generic drift. (medium confidence) (verified as of 2025-08-26) — [source](https://news.fsu.edu/news/education-society/2025/08/26/on-screen-and-now-irl-fsu-researchers-find-evidence-suggesting-chatgpt-influences-how-we-speak/) - 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](https://arxiv.org/abs/2406.07016) - 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](https://arxiv.org/abs/2409.01754) ## Facts and statements - Wikipedia's Signs of AI writing guide says its signs indicate potential problems, not proof, that it is not a ban on words, and that treating signs as problems to fix could make detection harder. (low confidence) (verified as of 2026-09-23) — [source](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) - Tic vocabulary shifts by model generation (earlier delve/intricate, newer enhance/showcase), so a fixed word list decays and needs dating. (low confidence) (verified as of 2026-09-23) — [source](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) - A 2026 arXiv paper reports that writing style (vocabulary complexity, formality) is a confound causing detector false positives on academic writing, so formal human prose can be flagged. (low confidence) (verified as of 2026-08-28) — [source](https://arxiv.org/pdf/2608.26710) - The Nigerian-English explanation for 'delve' is contested: comparing base Llama 2 to its RLHF-tuned version suggests RLHF preference itself promotes such words. (low confidence) — [source](https://arxiv.org/pdf/2412.11385) - A Max Planck study of 737,083 hours of podcast speech (824,634 episodes) found abrupt post-ChatGPT increases in words such as delve, showcase, boast, intricacies and meticulous; synthetic-control analysis supports a causal link, and the authors frame it as machines trained on human data feeding their traits back into human language. (low confidence) (verified as of 2026-07-16) — [source](https://arxiv.org/abs/2409.01754) - A preregistered experiment (N=496) found brief chatbot interaction led participants to adopt the chatbot's words, persisting through a distractor task and in forced lexical choice. (low confidence) — [source](https://arxiv.org/abs/2409.01754) - The term 'model collapse' is used inconsistently in the literature and is conflated with generic performance loss or any synthetic-data training; critics argue claims of inevitable collapse need scoping to the specific definition and data regime. (low confidence) — [source](https://arxiv.org/pdf/2503.03150) - Yakura et al. (Max Planck) analysed about 737,083 hours of unscripted conversation from 824,634 podcast episodes and report that ChatGPT-favoured words (delve, showcase, boast, meticulous) rose abruptly in spontaneous speech after ChatGPT's release; they use synthetic-control analysis to argue for a causal link. (high confidence) (verified as of 2026-07-16) — [source](https://arxiv.org/abs/2409.01754) - The 'GPT words' list was derived by having ChatGPT edit millions of emails, essays and articles and extracting words it repeatedly added; usage was then tracked in about 360,000 YouTube academic talks and 771,000 podcast episodes, with a surge within roughly 18 months of release. (medium confidence) — [source](https://www.scientificamerican.com/article/chatgpt-is-changing-the-words-we-use-in-conversation/) - The measured effect is lexical only; the Max Planck paper does not test semantic or grammatical influence and frames vocabulary as an early indicator of wider cultural influence. (medium confidence) — [source](https://arxiv.org/abs/2409.01754) - 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](https://arxiv.org/html/2412.11385v1)