Stereotypes—simplified, often inaccurate mental shortcuts—shape attention, memory, and judgment, influencing outcomes from education to public policy. This project investigates political party stereotypes in the United States, addressing three core questions: (i) can stereotypes be measured more precisely; (ii) what beliefs people hold about their own and opposing parties; and (iii) which interventions can durably alter these beliefs to reduce stereotype-driven affective polarization. Using an online experiment, we ask participants to allocate 100 hypothetical Democrats and Republicans across seven positions on key issues. This yields subjective belief distributions whose moments identify “stereotypical” members without imposing a fixed profile and capture within‑ and between‑party polarization. Validation occurs via an incentivized social‑categorization task where participants purchase diagnostic cues to classify a mystery individual, revealing alignment between cue selection and measured stereotypes and exposing stereotype‑driven classification errors. We then compare factual corrections with AI‑driven, personalized dialogues to test lasting stereotype change. Immediate and delayed effects are assessed through belief elicitations, affective‑polarization scales, and cooperation games. Findings show that belief‑distribution elicitation provides a more accurate misperception metric, that individual stereotypes predict information demand and classification behavior, and that data‑driven, tailored interventions can reduce stereotype‑driven affective polarization.