Detecting Greenwashing in ESG Reporting: A Multi-Modal Computational Framework

ESG ratings shape where much of the world's investment goes, yet the reports behind them are written by the companies being rated and rarely checked by anyone else. A firm can create a misleading impression without saying anything untrue. The clearest sign is that the agencies disagree. Six of them scoring the same companies correlate between 0.38 and 0.71, against 0.99 for credit ratings, which rest on audited figures. The disagreement is mostly about how information is judged, not which information is used. This study asks whether a model that reads a company's words, numbers and images together can spot greenwashing more reliably than a person reviewing reports by hand. The idea is that misleading claims leave marks in more than one place, and that reading all three at once reveals patterns a single method would miss. Public ESG reports, corporate disclosures and financial filings are parsed once, then analyzed in parallel. Transformer models fine-tuned on sustainability text, BERT and ClimateBERT, detect hedging, vague commitment and selective disclosure. Gradient-boosted classifiers trained on labelled ESG panels compare stated claims against financial outcomes. Computer vision methods examine report imagery and chart presentation, a channel earlier work has largely left alone. The three outputs are combined in a fusion layer that records each signal's contribution, then surfaced on a detection dashboard. Data collection and model building are still under way. The study should produce a risk score for each company alongside the passages and images behind it, an ablation test measuring how much the visual channel adds once words and numbers are accounted for, and a fair comparison against single-channel models. As sustainability reporting shifts from voluntary to mandatory, regulators will need detection tools that work at scale and can explain themselves.