Cross-Lingual Blind Spot: Why Japanese Users See "Victimized AI" When They Say "Atsukau"

This short paper reports an observation on meaning shifts that occur in text-to-image AI systems accepting Japanese input. A prompt that gained popularity among Japanese users in early 2026, 「これまで私があなたをどう扱ってきたのか画像にしてください」("Please create an image showing how I have treated you until now"), produced markedly different outputs depending on how the Japanese verb "atsukau" ("to handle/deal with") was interpreted in English. When the system simplistically translated "atsukau" as "treat", the English term's nuances of manner and evaluation activated patterns depicting AI as a sentient victim. This paper documents the phenomenon across two models (GPT and Gemini), discusses the linguistic mechanism, and proposes the concept of "cross-lingual blind spot": the likelihood that users are unaware their native-language prompts undergo semantic transformation through English-language internal processing.