As LLMs are increasingly used to grade or evaluate written academic work, an open question is whether the way something is written — its apparent personality traits or emotional tone — systematically shifts the score a model assigns, independent of the underlying content quality.

This project tests that directly: holding academic content constant while varying personality and emotional framing in the writing, then measuring whether an LLM evaluator’s scores drift as a result. The findings speak to whether LLM-as-grader setups need explicit debiasing safeguards before being trusted for high-stakes academic evaluation.