The Vocabulary Fingerprint of AI Rewriting: Common Words AI Language Models Prioritize

Working paper and open lexicons. We compared 60,786 human-written academic texts with AI rewrites of the same texts (eight model configurations across six model families, 40.4 million tokens) and measured which words models inject and suppress when rewriting. The signature is dominated by formal connectives and Latinate substitutions such as "thereby", "consequently", and "utilized", while the famous chatbot words behave differently: "meticulously" is enriched 214-fold, but "delve" is not enriched in rewriting at all. Two lexicons are released, related as subset and whole. The studied lexicon (1,057 words) is a small subset covering the most prominent AI-preferred words of the common vocabulary; it uses a centered AI score, 100·σ(δ), where 50 means equal use on both sides and the score can lean both ways. The full-vocabulary lexicon scores the entire English vocabulary as represented in WordNet, 90,520 single-word lemma and part-of-speech entries, on a one-sided 0 to 100 scale where 0 means no observable AI preference; 3,371 entries pass the two-signal evidence check. Files: both lexicons (CSV and JSON), candidate and phrase score tables, document frequencies, discovered sequences, analysis summaries, all analysis and figure code, and the paper PDF. The lexicons can also be browsed interactively and downloaded at https://textpulse.ai/research/ai-vocabulary-explorer. Paper page: https://textpulse.ai/research/ai-vocabulary-fingerprint