應用大語言模型於城市永續揭露之評分與潛在漂綠偵測：以 ChatGPT 為例

Abstract

「漂綠」（greenwashing）一詞通常指涉「選擇只揭露好消息、隱匿壞消息」，或「表面象徵性承諾、實際行動敷衍」的誤導手法。近年 ESG 規範日趨嚴謹，針對「企業漂綠」已有多項國際準則規範，惟「城市漂綠」之辨識與治理仍屬研究及實務缺口。過往文獻將漂綠進一步細分為「企業/組織層級」與「產品/服務層級」。本研究回顧並整合各漂綠架構，從「資訊面」與「行動面」歸納八項漂綠種類後，再將其拓展到城市治理情境。 本研究使用 2023 年 CDP Cities Questionnaire 共 537 份城市問卷資料，擷取其中四個核心開放式題目（Q1.1a 風險評估、Q4.1a 調適目標、Q5.1a 減量目標、Q7.1a 行動計畫），結合 OpenAI ChatGPT o4-mini 作為評分工具，產出「各題分數」、「整體分數」、「漂綠傾向」、「可能漂綠種類」等指標。後續分析以線性迴歸模型及 SHAP 探討大語言模型評分邏輯的一致性，輔以 T 檢定、Precision@K 檢驗其是否能有效區分 A-List 與 Non-A-List 城市。最後，以卡方檢定判斷 A-List 與 Non-A-List 城市間是否展現出不同的潛在漂綠行為模式。 實證結果顯示，大語言模型可以區分 A-List 與 Non-A-List 城市，各指標 T 檢定皆達顯著；Precision@25 精準度達 0.8 以上。而大語言模型所偵測出的潛在漂綠類型顯示，A-List 城市較傾向於出現「目標過於理想化」；Non-A-List 城市則更常見「行動計畫與監督不足」等行動面漂綠風險。 傳統的永續揭露仰賴受認證的第三方機構及專家逐字審查，既耗時又耗力。本研究提出組織層級的漂綠評估框架，並期望透過大語言模型、在低人力成本下快速完成初步的城市永續評分與潛在漂綠偵測，為地方政府、自願性揭露平台及投融資機構提供即時且可追溯的決策輔助。

“Greenwashing” usually refers to two forms of misrepresentation: selectively disclosing only positive information while concealing negative aspects, and making symbolic pledges that are not matched by substantive action. Although recent ESG regulations have established several international standards to address corporate greenwashing, tools for identifying and governing city-level greenwashing remain scarce in both research and practice. Previous studies categorize greenwashing into corporate/organization level and product/service level. Drawing on these studies, this research synthesizes existing frameworks, and specifies eight greenwashing categories, which are then extended to the context of municipal governance. Using 537 responses from the 2023 CDP Cities Questionnaire, we focus on four open-ended items—Q1.1a (assessment), Q4.1a (adaptation goals), Q5.1a (mitigation targets), and Q7.1a (action plans). Leveraging the OpenAI ChatGPT o4-mini model, we automatically generate item-level scores, an overall score, a greenwashing tendency index, and potential greenwashing categories. We then examine the model’s internal scoring logic via linear regression and SHAP analyses, and test its validity with Welch’s t-tests, and Precision@K to determine whether the scores can effectively differentiate CDP A-List cities from Non-A-List peers. Finally, chi-square tests ascertain whether these two groups exhibit distinct patterns of potential greenwashing behavior. Empirical results show that the large language model clearly distinguishes A-List from non-A-List cities (t-tests, p < 0.001), achieving a Precision@25 above 0.80. The model further reveals that A-List cities are more prone to “overly idealistic targets,” whereas non-A-List cities more frequently exhibit behavioural risks such as “insufficient action planning and monitoring.” Conventional sustainability disclosure relies on certified third-party reviewers who conduct painstaking, line-by-line evaluations, which is time-consuming and labor-intensive. By proposing an organization-level greenwashing assessment framework and demonstrating how large language models can deliver rapid, low-cost preliminary city-level sustainability ratings and potential greenwashing detection, this study offers an immediately traceable decision-support tool for municipalities, disclosure platforms, and investment institutions.