Carryover Effect in LLM Personality Measurement

Large language models are increasingly evaluated with human personality inventories, yet prior critiques mainly frame contextual variation as instability or explain directional shifts as user-targeted adaptation, such as sycophancy. We propose the personality carryover effect, a content-targeted contextual bias in which personality-related content in prior context shifts later questionnaire responses even when it does not reflect the user's own stance. Across MBTI and BFI-2 on two model families, third-person personality narratives systematically shift measured personality toward the described profile. Crucially, the shift persists when the user explicitly states dislike of that profile, distinguishing carryover from purely user-targeted adaptation. These findings identify a distinct validity threat to in-context personality measurement in LLMs.