ETD

• ai autophagy
• human-ai coevolution
• large language model

The widespread adoption of artificial intelligence, and especially Large Language Models (LLMs), is reshaping how information is produced, presented, and evaluated across many domains. Recruitment is a prominent example: employers increasingly rely on automated screening tools to process large volumes of applications, extract relevant signals, and support faster shortlisting decisions. In practice, however, resumes may be filtered out by algorithmic systems before reaching a human recruiter, often based on rigid matching criteria or opaque scoring rules. To adapt, candidates increasingly use LLMs and AI-powered resume builders to rewrite and tailor their applications to specific job descriptions. This raises a central question: if many candidates rely on the same generative tools, do resumes converge toward similar language and structure, reducing their ability to signal individual differences?
This thesis investigates linguistic homogenization in LLM-assisted resume rewriting, whether it intensifies when synthetic resumes are generated iteratively, and what implications this may have for automated screening. We conduct a pretest--posttest simulation comparing original resumes with LLM-augmented versions produced under four prompting strategies. We evaluate changes in semantic similarity, lexical diversity, and information-theoretic measures in both resume--resume and resume--job description settings, and we examine how these patterns evolve across successive rewriting steps. Across strategies, augmented resumes become more similar to each other and exhibit reduced vocabulary variety and lower uncertainty, indicating more standardized and predictable text. Repeated iterations further amplify these effects, consistent with progressive information loss under iterative generation and with reduced differentiation among candidates in downstream matching and ranking.