Detecting Greenwashing: A Natural Language Processing Literature Survey

Greenwashing refers to practices by corporations or governments that intentionally mislead the public about their environmental impact. This paper provides a comprehensive and methodologically grounded survey of natural language processing (NLP) approaches for detecting greenwashing in textual data, with a focus on corporate climate communication. Rather than treating greenwashing as a single, monolithic task, we examine the set of NLP problems, also known as climate NLP tasks, that researchers have used to approximate it, ranging from climate topic detection to the identification of deceptive communication patterns. Our focus is on the methodological foundations of these approaches: how tasks are formulated, how datasets are constructed, and how model evaluation influences reliability. Our review reveals a fragmented landscape: several subtasks now exhibit near-perfect performance under controlled settings, yet tasks involving ambiguity, subjectivity, or reasoning remain challenging. Crucially, no dataset of verified greenwashing cases currently exists. We argue that advancing automated greenwashing detection requires principled NLP methodologies that combine reliable data annotations with interpretable model design. Future work should leverage third-party judgments, such as verified media reports or regulatory records, to mitigate annotation subjectivity and legal risk, and adopt decomposed pipelines that support human oversight, traceable reasoning, and efficient model design.

💡 Research Summary

The paper presents the first comprehensive, methodologically grounded survey of natural language processing (NLP) approaches aimed at detecting greenwashing in corporate and governmental climate communications. Rather than treating greenwashing as a monolithic classification problem, the authors decompose it into a suite of “climate NLP” subtasks that collectively approximate the phenomenon. These subtasks include climate‑topic identification, risk or carbon‑transition classification (aligned with frameworks such as TCFD, ESG, ESRS), claim‑type categorization, sentiment, stance and argument‑quality assessment, and the detection of deceptive rhetorical patterns (e.g., overly positive language, vague commitments, selective disclosure).

The survey analyzes 61 peer‑reviewed works, detailing for each subtask the data collection strategies (expert‑annotated, crowd‑annotated, or rule‑based keyword labeling), model architectures (from classic TF‑IDF + SVM to modern transformer‑based models such as BERT, RoBERTa, DeBERTa, and large language models like GPT‑3/4), and evaluation protocols. The authors find that tasks with clear lexical cues—climate‑topic detection, risk classification, and claim‑type labeling—have reached near‑perfect performance (often >95 % F1) when fine‑tuned on balanced, noise‑free datasets. Simple keyword baselines sometimes achieve comparable scores, indicating that many of these tasks rely more on surface forms than deep semantic understanding.

Conversely, subtasks requiring nuance, subjectivity, or multi‑document reasoning—distinguishing specific versus vague commitments, assessing rhetorical deception, or evaluating consistency across a company’s disclosures—remain challenging. Reported inter‑annotator agreement for these tasks is low (Cohen’s κ often below 0.5), and datasets are small (hundreds to a few thousand instances), leading to over‑fitting concerns. Moreover, evaluation practices are frequently inadequate: most papers report only accuracy or F1, omit uncertainty estimates, ignore semantic proximity between labels, and lack strong baselines (random, majority, or keyword‑only). Experiments are typically conducted on curated, balanced corpora, which may not reflect the noisy, imbalanced nature of real‑world sustainability reports.

A critical gap identified is the absence of any publicly available dataset containing verified greenwashing cases—i.e., instances confirmed by regulators, courts, or investigative journalism. Existing work therefore relies on proxy indicators (excessive positivity, framing patterns, stance inconsistencies) that have not been validated against ground‑truth legal judgments. This limits the ecological validity of current models and hampers the ability to claim true greenwashing detection.

To address these shortcomings, the authors advocate for a principled data collection pipeline grounded in third‑party judgments, clearer and narrower definitions of misleading communication, and decomposed pipelines that combine automated analysis with human oversight. They suggest incorporating uncertainty quantification, semantic‑aware metrics, and robust baselines, as well as designing interpretable models that allow auditors to trace decisions.

In conclusion, while certain building blocks of greenwashing detection are technically mature, constructing a reliable, trustworthy end‑to‑end system remains an open challenge. Progress requires (1) creation of verified greenwashing datasets, (2) improved annotation protocols with higher inter‑rater reliability, (3) rigorous evaluation methodologies that reflect real‑world conditions, and (4) human‑in‑the‑loop designs that ensure transparency and accountability. The paper outlines these research directions as essential steps toward deploying NLP tools that can meaningfully support investors, journalists, regulators, and activists in combating greenwashing.