生成式AI之虛擬人格設計工程及其在市場研究之應用策略探析

Abstract

本研究旨在探討生成式人工智慧（Generative AI）於虛擬人格（AI Persona）建模之設計工程流程，並深入分析其在市場研究領域的應用策略與實務發展潛力。研究以Henry Murray（1938）提出之心理需求理論為核心理論架構，建構六類動機向度（成就、感情、資訊、佔有、權力、地位/保護），結合電通行銷傳播集團CCS（Consumer Connection System）調查去識別化資料，運用提示詞工程（Prompt Engineering）技術，生成具備特定人格特徵的AI Persona敘事內容，並進行語義一致性與動機再現性之系統化分析。研究結果證實，在結構化提示設計與精確資料配置條件下，AI Persona能夠有效模擬真實消費者之動機特徵表現，從而驗證本研究之核心假設。 然而，本研究發現語言模型在處理抽象或具爭議性的人格動機面向（如地位追求與權力展現）時，表現仍存在明顯侷限，特別是在負面語意表達上呈現系統性偏誤，例如傾向性誤判等現象。相較於McCrae & Costa（1987）所提出的大五人格模型之統計可區辨的五維特質架構，其五個連續性特質向度在語言模型中較易透過語言風格或表達習慣進行建模與區辨；Murray（1938）所建構之心理需求理論則涉及情境觸發與內在動機的交互關係，具有高度語境依賴性與需求重疊性。為提升語言模型對動機性人格的精確模擬能力，本研究提出三項改進策略：其一，設計具情境深度的提示詞場景以誘發目標動機；其二，建構動機導向的語言特徵詞庫以輔助模型識別；其三，採用混合式人格建模架構，以結合動機需求與特質性格的表徵優勢。綜合而言，生成式AI結合心理學理論與實證調查資料之虛擬人格設計工程，不僅為市場研究提供高效率且具創新性的洞察分析工具，更為AI驅動之合成資料應用與消費行為模擬研究領域建立嶄新的研究典範。

This study aims to investigate the design engineering processes of AI Persona modeling through Generative AI and to conduct an in-depth analysis of its application strategies and practical development potential in the field of market research. Grounded in Henry Murray's (1938) psychological needs theory as the core theoretical fundamental, this research constructs six motivational dimensions (achievement, affection, information, acquisition, power, and status/protection), integrating the survey data from Dentsu Group Inc. Through proper prompt engineering techniques, AI Persona narratives with specific personality characteristics were generated, followed by systematic analysis of semantic consistency and motivational reproducibility. The findings confirm that under structured prompt design and data configuration, AI Personas can effectively simulate the motivational characteristics of real consumers, thereby validating the core hypothesis of this study. Furthermore, the study reveals that language models exhibit performance limitations when processing abstract or controversial motivational dimensions (such as status pursuit and power manifestation), while simultaneously demonstrating systematic bias in expressing negative semantic content. Compared to the bipolar dimensional structure of the Big Five personality model, Murray's psychological needs theory faces challenges including high contextual dependency and technical difficulties in feature separation during the language mapping. Based on these findings, this research proposes three optimization strategies: contextualized prompt design, systematic construction of linguistic feature lexicons, and hybrid personality modeling, to further enhance the precise representational capabilities of language models for motivational personalities. In conclusion, the virtual persona design engineering that combines generative AI with psychological theories and empirical survey data not only provides market research with highly efficient and innovative analytical tools for insights, but also establishes a novel research paradigm for AI-driven synthetic data applications and consumer behavior simulation research fields.