New research suggests AI writing tools can change the quality and scope of feedback in ways that may not match educators’ goals for personalization. Researchers at Stanford University’s Institute for Human-Centered AI found that when AI writing coaches received information about a student’s background—such as race, gender, language, disability status, or achievement level—the feedback shifted significantly. The study, presented at the International Learning Analytics and Knowledge conference in Bergen, Norway, tested models including GPT-4o and Meta’s Llama variants on 600 eighth-grade persuasive essays. Findings showed that supplying student attributes can lead the system to provide more advanced suggestions for higher-performing profiles while reducing struggling students’ feedback toward spelling corrections and narrower rewrite suggestions. For higher education faculty using AI-assisted assessment or supporting ed-tech pilots, the result highlights a compliance and pedagogy issue: personalization features may reproduce bias through differential coaching—even when the underlying writing sample is identical.