Maintaining a living tic list
Parent: Measuring LLM lexical tics: excess vocabulary, baselines, drift · Published reference · snapshot 2026-09-23
↓ Facts as markdownall context files
Source-anchored research for Maintaining a living tic list. 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.
Structure and components
- Wikipedia's guide groups AI vocabulary by era (2023-mid 2024 GPT-4; mid 2024-mid 2025 GPT-4o; mid 2025 onward GPT-5) and carries a banner saying parts about the most recent models need updating. (low confidence) (verified as of 2026-09-23) [source]
Measurements and reference values
- Antislop builds a model-specific slop profile automatically by comparing generated text to human baselines; some patterns appear over 1,000x more often than in human text. (medium confidence) [source]
- Antislop Sampler suppresses over 8,000 patterns, whereas plain token banning becomes unusable at about 2,000; FTPO reports ~90% slop reduction with benchmarks maintained. (low confidence) [source]
Comparisons and alternatives
- Instruction tuning with human feedback drives much overuse (e.g. 'nuanced' +8,342% vs base Llama), so lists should be regenerated after each tuning/RLHF change; the study covers only Llama and does not track drift across models or time. (low confidence) [source]
Facts and statements
- Overused vocabulary changes over time: 'delve' was heavily overused by ChatGPT in 2023-early 2024, declined later in 2024, and dropped sharply in 2025, so a static list decays. (medium confidence) (verified as of 2026-09-23) [source]
- Overuse of a word does not imply its synonyms are overused, so list entries should be exact terms, not thesaurus expansions. (low confidence) [source]
- Human writing is increasingly influenced by LLM patterns, so human-baseline checks for list entries also need refreshing. (low confidence) (verified as of 2026-09-23) [source]
Children
- No children recorded.