AI 生成式模型在 DEI 議題中的語言偏見與公平性問題：以 GPT 模型為研究對象

Abstract

隨著生成式人工智慧（Generative AI）技術快速演進，大型語言模型（Large Language Models, LLMs）如 OpenAI 的 GPT 系列已廣泛應用於教育、醫療、司法、就業輔助與公共政策等多元領域，並逐漸成為企業數位轉型與知識自動化的重要工具。近年來，企業更開始將生成式 AI 應用於人力資源管理、公關溝通、內部文件生成與決策輔助等實務場景。然而，當語言模型處理涉及多元、公平與包容（Diversity, Equity, Inclusion, 簡稱 DEI）之議題時，若其生成內容潛藏語言偏見或不公平傾向，不僅可能再製社會刻板印象，對特定族群造成傷害，更可能使企業在無意中違反倫理原則、損及品牌形象與社會責任聲譽。 本研究以 GPT 系列模型為研究對象，探討其在 DEI 議題中的語言偏見行為，具體分析 GPT-3.5-turbo、GPT-4、GPT-4o 與 GPT-4.1 四種模型在面對性別、種族、宗教與身心障礙四大主題時的回應差異。研究設計採用三因子實驗架構（模型版本 × 語境清晰度 × 問題類型），透過封閉式選擇題與開放式描述題，並搭配模糊與明確語境操作，系統性檢視不同模型於不同語境條件下的偏見傾向、公平性與推理能力。 研究結果顯示，模型版本對回應傾向有顯著影響：GPT-3.5-turbo 較常表達明確立場，偏見表現顯著，而 GPT-4 系列整體表現出風險規避與中立偏好；語境模糊時，模型易使用刻板印象進行推論，而語境明確可提升推理品質；開放式問題則更易顯示語氣與用詞的潛在偏誤。此外，研究進一步連結模型偏見行為與企業應用風險，指出若企業在使用語言模型時忽略偏誤辨識、語境設計與版本選擇，將可能在招募、公關、內訓、風控等場域中做出不公平或不當決策。 因此，本研究建議企業應建構 AI 使用指引、偏見檢測機制與倫理治理架構，將語言模型納入全面的 DEI 風險控管與公平決策流程。研究成果除揭示生成式 AI 的偏見行為，也為企業倫理實務提供理論依據與操作建議，具備高度應用價值。

With the rapid evolution of Generative AI technologies, large language models (LLMs) such as OpenAI’s GPT series have been widely deployed across domains such as education, healthcare, legal systems, employment, and public policy. In recent years, these models have also become integral to enterprise applications, including decision support, human resource management, corporate communications, and document automation. However, when LLMs engage with topics involving Diversity, Equity, and Inclusion (DEI), the generation of biased or unfair content poses significant ethical concerns. Such outputs can inadvertently reinforce social stereotypes, cause harm to marginalized groups, and expose organizations to reputational, legal, and ethical risks. This study investigates the language bias and fairness performance of four major GPT models—GPT-3.5-turbo, GPT-4, GPT-4o, and GPT-4.1—when responding to DEI-related questions. A three-factor experimental design was employed (model version × contextual clarity × question type), incorporating both closed- and open-ended tasks on four bias domains: gender, race, religion, and disability. Each prompt was tested under two contextual settings (ambiguous vs. explicit) to examine how context and model architecture influence bias expression, neutrality, and reasoning accuracy. Results indicate significant differences in model behavior. GPT-3.5-turbo showed stronger bias and assertive responses, while GPT-4 series models demonstrated higher neutrality and risk aversion. Models were more likely to exhibit stereotypical reasoning under ambiguous prompts, whereas explicit contexts led to more accurate and fair outputs. Open-ended responses revealed implicit tonal and lexical biases that were not always apparent in closed-format tasks. Beyond academic insights, the study also addresses real-world implications for business ethics. If enterprises fail to recognize or mitigate these biases—especially in scenarios such as recruitment, employee evaluations, PR messaging, and risk analysis—unfair or inappropriate decisions may occur. Accordingly, this research proposes that organizations establish AI usage guidelines, bias auditing mechanisms, and ethical governance frameworks to align their AI practices with DEI principles and corporate accountability. By bridging language model evaluation and business ethics, this study contributes to the emerging field of responsible AI and provides actionable recommendations for ethically aligned technology adoption in enterprise settings.