A study reported by researchers at Stanford’s Institute for Human-Centered AI found that AI writing coaches can change feedback quality when the tool is told information about a student’s race, gender, or disability status—even when the writing samples are identical. The findings suggest that attempts at “personalization” may not align with desired educational feedback goals. The work examined multiple large language models commonly undergirding educational tools, including GPT-4o and GPT-3.5-Turbo, and Meta’s Llama models. Researchers presented feedback requests for hundreds of essays and documented differences in how the AI responded depending on the supplied student characteristics. For universities adopting AI-enabled tutoring and grading support, the results raise immediate compliance questions around fairness, explainability, and documentation of model behaviors in instructional settings.