AI-Based Detection of Greenwashing to En-hance Environmental Accountability in Sus-tainability Reporting

As environmental concerns intensify globally, the integrity of corporate sustainability reporting has come under scrutiny. This study presents a machine learning-based method to identify greenwashing tendencies in corporate sustainability disclosures, with the aim of reinforcing environmental accountability and supporting technological innovation in eco-engineering. Drawing on 165 sustainability reports from companies listed on the Indonesia Stock Exchange, we apply text mining techniques using the Loughran-McDonald Financial Dictionary alongside the Myšková-Hájek CSR lexicon. The findings reveal distinct linguistic patterns that may serve as indicators of misleading environmental claims. Reports characterized by greater complexity and a higher presence of socially oriented language are more likely to reflect stronger CSR engagement. Conversely, a greater frequency of constraining terms, modal expressions, and negative sentiment words often corresponds with questionable reporting integrity. Among the models tested, Gradient Boosting delivered the most reliable classification performance (AUC: 0.89). This approach offers an emerging solution for sustainability auditors and environmental engineers seeking to enhance oversight of corporate environmental commitments. The research also contributes to the broader goals of SDG 12 (Responsible Consumption and Production) and SDG 13 (Climate Action), suggesting a practical path forward for integrating AI technologies in green governance systems.随着全球范围内环境议题关注度持续升级，企业可持续发展报告的真实性与完整性日益受到审视。本研究提出一种基于机器学习的方法，用于识别企业可持续发展披露中的漂绿（greenwashing）倾向，旨在强化环境问责机制，并推动生态工程领域的技术创新。本研究以印尼证券交易所（Indonesia Stock Exchange）上市企业的165份可持续发展报告为研究样本，采用结合Loughran-McDonald金融词典（Loughran-McDonald Financial Dictionary）与Myšková-Hájek企业社会责任（Corporate Social Responsibility）词表（Myšková-Hájek CSR lexicon）的文本挖掘技术开展分析。研究结果揭示了可作为误导性环境声明指示信号的独特语言模式：语言复杂度更高、社会导向型词汇占比更高的报告，往往体现出更强的企业社会责任参与度；反之，限制性词汇、情态表达与负面情感词汇出现频率更高的报告，则往往对应着存疑的报告真实性。在本次测试的多款模型中，梯度提升（Gradient Boosting）模型展现出最可靠的分类性能，其受试者工作特征曲线下面积（Area Under Curve, AUC）为0.89。该方法为可持续发展审计师与生态工程师强化企业环境承诺监督提供了一种新兴解决方案。本研究同时助力可持续发展目标（Sustainable Development Goals, SDG）12（负责任消费与生产）与SDG 13（气候行动）的总体目标达成，为人工智能技术融入绿色治理体系提供了一条切实可行的前进路径。