TrueGreen AI: Intelligent Verification of Sustainability Claims for Smarter and Greener Consumer Decisions

Greenwashing, or the creation of environmentally misleading claims about products, is becoming a more prevalent phenomenon among products labelled as “eco-friendly,” “natural,” or “sustainable” without adequate evidence to prove such claims in the context of e-commerce. Prior research in greenwashing detection focuses on manual content analysis or single-modality text processing and is confined to food and beverages. Multimodal verification based on evidence in multiple modalities (text, image, speech), public benchmark datasets with annotated greenwashing instances, and explanation in machine learning or natural language processing approaches are relatively under-explored. We introduce True Green AI, a multimodal framework for auditing environmental sustainability claims based on textual, visual, and verbal evidence. The system considers the representation of a sustainability claim as an aggregation of evidence in text, image, and speech, independent analysis of each modality, and aggregation of outcomes into an interpretable risk profile. In addition to the greenwashing classification based on semantic entailment, we introduce a metric called the Cross-Modal Divergence Index (CMDI) that quantifies cross-modal divergence in a brand’s communication, presentation, emphasis, and evidence. The framework is developed as a decoupled architecture, composed of a React 19 client and Node.js/Express service layer with Firebase persistence. Browser-based technologies are used for capturing camera and microphone input, AI-assisted analysis of claims and evidence is performed by our system, and generation of justification based on evidence retrieval is constrained. Our system generates an interpretable A–E assessment of a sustainability claim, along with the evidence used for the evaluation and component scores, as opposed to just numbers or classifications. Such an approach is inspired by research on consumer perception that indicates that interpretative, graded labels can enhance consumer understanding as compared to numeric or unadjudicated environmental information. Our approach addresses four research gaps, including: multimodal verification, cross-modal consistency evaluation, retrieval-based generation, and interpretable output generation. However, quantitative evaluation is contingent upon the development of an annotated multimodal dataset of greenwashing instances, and the current study is presented as a framework for decision-making as opposed to an autonomous adjudication tool.