The Load-Bearing Vocabulary of Claude
Louis Abraham conducted a data analysis project revealing Claude's repetitive word choices in GitHub pull requests. The analysis grouped descriptions by their word usage, identifying eight distinct writing styles. One of these styles dramatically increased its prevalence from 1.0% to 45% in just over a year. This suggests a growing uniformity in how Claude generates these descriptions. Further research into AI watermarking, specifically Google's SynthID-Text, highlights a related issue. This watermarking technique causes a reduction in inter-response diversity. Essentially, different answers to the same prompt become more alike. This means there is less variety in watermarked text output. This problem is particularly concerning given Claude's existing struggles with response diversity. The implications are that watermarking might exacerbate Claude's tendency towards repetitive language.