Consequences for tic lists and detectors
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 Consequences for tic lists and detectors. 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.
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]
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]
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]
- 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]
- 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]
- 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]
Children
- No children recorded.