Graphite Finds Claude Opus 5.5 Writes More Like Humans but Still Has 2,548 AI Tells

Summary

Graphite’s updated AI Tells study adds Claude Opus 5.5 to a comparison of model-generated and human articles across 9,974 aligned topics. Opus 5.5’s unigram word-distribution divergence from human writing fell from 0.064 for Opus 5 to 0.052, while GPT-6 Astra’s divergence rose to 0.109 from 0.101 for GPT-5.6 Sol. The researchers say Claude’s move toward human-like word use is not explained only by common function words: nouns, verbs, adjectives, and adverbs account for about 79% of the divergence gap between Astra and Opus 5.5. Opus 5.5 also scored 10.57 for mannered prose, 37% below Opus 5’s 16.75, but remained about 1.6 times the human score; Astra was about 1.2 times the human score. Its em-dash use fell 99%, from 2.92 to 0.015 per 1,000 words, and both Opus 5.5 and Astra used em dashes less often than the pre-ChatGPT human articles in the comparison. However, the model still produced 2,548 disproportionately common words, phrases, and frames, only 4% fewer than Opus 5’s 2,666. Opus 5.5 used more helpfulness language, while Opus 5 used emphatic phrases more often, and Opus 5.5 used more superlatives than Astra. The study says well-known AI tells have weakened across model versions, but different tells have replaced or persisted alongside them. Results use shared article topics and defined frequency thresholds; earlier-model figures can shift slightly because adding Opus 5.5 changed the shared sample. The researchers provide the tells and raw article data for download.