Using AI to Expose Corporate Greenwashing Lies: Practical Analysis of an ESG Report Intelligent Detection System

This article introduces a corporate greenwashing detection project that combines large language models (LLMs) and machine learning technologies. The project builds a multi-dimensional detection framework integrating three types of features: semantic, lexical, and financial. It aims to address the problems of low efficiency and strong subjectivity in traditional manual auditing, providing technical support for ESG investment decisions, regulatory screening, and corporate self-assessment.

"Greenwashing" refers to false advertising or exaggerated marketing by enterprises to shape an eco-friendly image. Common tactics include using vague environmental terms, framing routine compliance as achievements, and selective disclosure. Traditional detection relies on manual auditing, which is inefficient and highly subjective when handling large volumes of reports, thus spurring the need for AI automated detection.

The core innovation of the project is a detection framework integrating three types of features: 1. Semantic features (using LLMs to capture implicit greenwashing signals in text); 2. Lexical features (analyzing keyword frequency, sentence patterns, emotional polarity, etc.); 3. Financial features (verifying whether environmental investment matches claims). The experiment designed 6 feature combination schemes (M1-M6) and compared multiple architectures including gradient boosting trees (XGBoost, etc.), deep learning (MLP, etc.), and sequence models (LSTM, etc.).

Ablation experiments reveal: Lexical features improve recall rate (identifying more potential greenwashing cases); financial features improve precision rate (filtering false positives); semantic features provide deep understanding. Comparison between ChatGPT and Llama: ChatGPT has low prompt sensitivity, stable semantics, and reliable results; Llama has high variability and prompt sensitivity, but stronger discriminative ability in some dimensions.

Application value: Assists investors in ESG decision-making, improves screening efficiency for regulatory authorities, and helps enterprises self-assess the quality of their ESG reports. Limitations: Relies on the quality and coverage of training data; model needs continuous updates as greenwashing behaviors evolve; AI detection should assist rather than replace manual auditing.

This open-source project demonstrates the possibility of building a multi-dimensional greenwashing detection system by combining LLM semantic understanding with traditional machine learning, providing a noteworthy direction for sustainable finance, corporate social responsibility, and technological innovation fields.