AI Vocabulary Explorer

Every AI score is measured from one paired corpus. We compared 60,786 human-written academic texts with AI rewrites of the same texts, so the topic and content are constant and any vocabulary difference is caused by the AI model. For each word we measure two independent signals: (1) how often it is used, its token rate; and (2) how widely it is used, the share of documents that contain it at least once. Each signal is expressed as a regularized log-odds ratio of AI use vs. human use, the standard method for comparing word use between corpora (Monroe et al., 2008). The two signals are averaged and mapped onto a 0 to 100 scale through the logistic (sigmoid) function σ:

A word is scored above 0 only when it occurs over a predefined threshold in the corpus: at least 25 combined occurrences and a z-score of at least 3.29 (p < 0.001) on each signal independently, in the AI direction. Requiring both signals prevents a word from scoring from heavy repetition in a small handful of documents. From the 90,520 WordNet lemma entries we attempted, 3,371 qualified, namely, 1,502 nouns, 838 adjectives, 779 verbs, and 252 adverbs. Every other word is scored 0 by default, and is considered to be used roughly equally in AI and human writing. Words that human authors use significantly more than models are considered 'human-leaning'. This full lexicon represents the entire AI-scored English vocabulary.

The 1,057-word studied lexicon from the underlying paper uses a centered version of the same statistic, AI score = 100 · σ(δ), where 50 means equal use on both (human + AI) sides and the score can lean both ways. This lexicon represents the most prominent AI-preferred words of the common vocabulary, a small subset of the full lexicon. Both lexicons, the per-word statistics, and the analysis code are released openly under CC BY 4.0.