Impersonal Statements: LLM-Era College Admissions Essays Exhibit Deep Homogenization Despite Lexical Diversity

Growing evidence indicates that AI homogenizes content, and outputs of human-AI collaborations in laboratory contexts are also more similar than content generated by humans alone. But important questions remain about whether AI homogenization is changing the way humans generate ideas in the real world, including whether its effects extend into contexts where writing is expressly personal and to high-stakes writing where AI use is transgressive. Parallel questions concern how human writing may be changing – whether AI homogenizes all aspects of writing, or only some. In a preregistered natural experiment in 83,160 college applicants, we analyzed semantic diversity at the word, sentence, and document levels in high-stakes personal statement essays submitted before and after ChatGPT’s public release. A counterintuitive pattern of effects emerged whereby post-GPT essays exhibited increased superficial diversity at the level of word choice but simultaneous homogenization of the underlying ideas and themes those words expressed. These findings replicated across multiple demographic subgroups but the homogenizing pressure was more pronounced among applicants from backgrounds outside the cultural mainstream, indicating that AI may have disproportionately homogenizing effects on non-mainstream voices.