Green or Greenwashing? A novel deep learning method integrating multi-source data for ESG greenwashing detection

The Environmental, Social, and Governance (ESG) reports of listed companies offer stakeholders valuable insights into a company's sustainability practices. However, some companies exaggerate their green, low-carbon, and energy-saving initiatives in their ESG reports, even though they face environmental penalties in reality, engaging in “greenwashing” behavior. This severely undermines public confidence in the reliability of ESG reports and threatens long-term sustainability. Existing studies on company greenwashing have mainly focused on explaining greenwashing behavior, rather than detecting it, which fails to provide early warnings for relevant stakeholders. Additionally, while multi-source data from financial reports, ESG reports, and other sources contribute to fully capturing greenwashing clues, existing studies have rarely focused on exploring their joint role in greenwashing detection. Under these circumstances, we propose a novel deep learning method integrating multi-source data for ESG greenwashing detection, named DLM-ESG. Specifically, the proposed DLM-ESG employs large language models (LLMs) to extract ESG-related content from financial reports and uses attention mechanisms to capture the varying importance of different data sources. By fully leveraging the complementary strengths of multi-source data, DLM-ESG enables more accurate and robust detection of greenwashing. Experiments on real-world datasets demonstrate that DLM-ESG significantly outperforms existing baseline methods in identifying ESG greenwashing behaviors.