AI Reads Between the Lines of Corporate Green Reports in China

Corporate sustainability reports have become the public face of the green economy, promising investors and regulators a window into how companies are managing their environmental impact. But a growing problem shadows this flood of disclosure: how can anyone tell whether a company’s confident, optimistic language is actually backed by concrete strategies and measurable commitments? A new study published in Environmental and Sustainability Indicators offers a fresh, computationally grounded answer. Researchers led by Peng Zhou, Donatella Valente, Vincenzo Giannico, Mario Elia and Raffaele Lafortezza have developed a text-based tool called the Sentiment–Substance Indicator, which uses natural language processing to measure the alignment between how positively a corporation talks about the environment and how densely its reports describe specific environmental strategies. Applied to twenty of China’s largest and most visible corporations, the indicator reveals striking differences between sectors — and offers a transparent new lens on the perennial problem of greenwashing, without ever claiming to detect it directly.

The study arrives at a pivotal moment for Chinese corporate disclosure. As the world’s second-largest economy and one of the largest contributors to global greenhouse gas emissions, China plays a decisive role in global climate mitigation. Since the government announced its ‘Dual Carbon’ targets — peaking carbon emissions before 2030 and achieving carbon neutrality by 2060 — Chinese corporations have shifted from largely voluntary corporate social responsibility practices toward increasingly structured and mandatory ESG reporting. This transition has dramatically increased the volume and visibility of environmental disclosures, but it has also created new challenges in evaluating their credibility and comparability. Reporting practices remain highly variable across sectors, with companies differing in how they present strategies, how much detail they provide, and how they balance narrative description against quantitative data.

The conceptual heart of the research lies in distinguishing between two dimensions of disclosure that are often conflated: sentiment and substance. Sentiment captures the narrative tone of a report — the positivity or negativity of its language — while substance captures the reporting density of specific environmental strategies, such as carbon emission reduction, renewable energy adoption, or waste management. The authors ground this distinction in three complementary theoretical perspectives. Signaling theory views corporate disclosure as a mechanism for reducing information asymmetry between companies and stakeholders. Impression management theory highlights how firms frame and emphasize information to shape perceptions. Disclosure-quality perspectives emphasize that the usefulness of reporting depends not on volume alone but on specificity and relevance. Narrative–substance divergence, the researchers stress, is not evidence of deception; it may arise from reporting conventions, sector characteristics, regulatory requirements, or communication practices.

To build the indicator, the team assembled a purposive sample of twenty large, high-visibility Chinese corporations, four each from five contrasting sectors: Finance, Energy, Food and Beverage, Technology, and Automotive. The selection criteria reflected corporate scale, market relevance, and environmental relevance — including asset size, revenue, market capitalization, and sales volume, complemented where available by MSCI ESG ratings. The Finance group included giants such as the Industrial and Commercial Bank of China and Ping An Insurance; the Energy group included Sinopec, PetroChina, China Shenhua and CNOOC; the Technology group spanned Tencent, Alibaba, JD.com and Baidu; and the Automotive group featured BYD, SAIC Motor, Geely Holding and NIO. For each corporation, the researchers analyzed the environmental section of the 2024 fiscal-year ESG or Sustainability Report, excluding front matter, executive statements, governance sections and appendices to ensure consistency across the corpus.

The resulting dataset was substantial: 520,071 words and 675,304 tokens of environmental text. The NLP pipeline began with preprocessing steps — tokenization, normalization, stop-word removal, and lemmatization — with Chinese-language segmentation performed using the Jieba library in Python, while English versions of reports were consulted only to verify terminology. Because general-purpose sentiment dictionaries often miss the specialized vocabulary of environmental reporting, the team built a domain-specific lexicon combining established Chinese sentiment dictionaries, sector-specific terminology from finance, environmental science, technology and manufacturing, and ESG-specific terms identified during corpus preparation. A Sentiment Score was then calculated for each report, normalized to a range from −1 (strongly negative) to +1 (strongly positive), with token-count normalization reducing the influence of differences in report length.

Substance was measured through normalized keyword-frequency analysis across seven environmental strategy categories: carbon emission reduction, energy consumption improvement, water recycling and reduction, waste management, renewable energy adoption, carbon neutrality achievement, and Nature-based Solutions adoption. Keyword frequencies were expressed as mentions per 1,000 tokens, drawing on themes common to established frameworks such as the Global Reporting Initiative, SASB, and MSCI, while also reflecting China’s Dual Carbon policy context. The Sentiment–Substance Indicator then compared the two dimensions relatively within the sample: reports combining comparatively high sentiment with lower substance were interpreted as showing greater narrative–substance divergence, while the reverse pattern indicated closer alignment. Crucially, the researchers emphasize that the indicator is a diagnostic screening tool, not a verdict on corporate behavior or actual environmental performance.

The results are revealing. The Technology sector showed the lowest average Sentiment Score (0.02) and the highest Substance Metric (72.0), corresponding to relatively higher disclosure alignment — a near-neutral narrative tone paired with dense reporting of specific strategies, including high mention frequencies for carbon emission reduction (15.2) and waste management (13.8). The Finance group presented the contrasting picture: the highest average sentiment (0.16) alongside the lowest substance (54.0), indicating relatively lower alignment within the sample. The authors suggest a plausible explanation: financial institutions report environmental strategies through financed emissions, green finance, portfolio exposure and climate-related investment rather than direct operational measures, which may naturally produce different disclosure profiles. Energy, Automotive, and Food and Beverage occupied intermediate positions, with sentiment scores between 0.11 and 0.13 and substance values between 61.2 and 64.2. The Food and Beverage sector, for instance, emphasized water-related strategies (11.2 mentions per 1,000 tokens), reflecting water’s role as a key production input, while showing comparatively low reporting of renewable energy (7.2).

Reliability was tested against a manually annotated dataset of 500 sentences, 100 from each sectoral group, each classified as positive, neutral or negative. The automated classification performed consistently across sectors, with accuracy ranging from 82.9 to 88.5 percent and F1 scores from 82.3 to 85.3 percent, providing an empirical basis for using the Sentiment Score as the narrative component of the indicator. The study also examined Nature-based Solutions — afforestation, ecosystem restoration, and land management practices that enhance carbon sequestration — as reported by the corporations. The Energy group reported the highest aggregated values: 7.56 billion Yuan in NBS investment, 10.0 million trees planted, and 3.252 million tons of CO2-equivalent in reported carbon offsets. PetroChina alone reported 3.41 million trees planted in 2024, including 1.624 million trees under an ‘Afforestation Campaign for 10,000 Well Sites’. The Technology group reported the lowest aggregated planting (1.20 million trees) and offsets (0.327 million tons CO2e). The authors caution that these figures are corporation-reported and were not independently verified, and that reported offsets do not address issues of additionality, permanence, or long-term sequestration.

The implications extend to both assessment practice and policy. For investors and regulators, the indicator offers a complementary screening layer: reports showing relatively greater narrative–substance divergence could be flagged for closer examination using independent environmental-performance data. The computational structure is suitable for scaling to much larger collections of ESG reports, although the authors stress that performance and interpretability require validation on broader and more representative datasets. On the policy side, the findings suggest that standardized disclosure requirements — particularly regarding specific environmental strategies and measurable information — could improve comparability and reduce ambiguity, and that the goal should be greater specificity and transparency rather than simply more disclosure volume.

The limitations are candidly acknowledged. The sample of twenty corporations is an exploratory benchmark, not a statistically representative slice of Chinese business; the analysis covers a single reporting year; the Substance Metric is sensitive to the choice of keyword taxonomy; and the NBS figures rest on unverified corporate claims. Yet as a proof of concept, the study marks a meaningful step forward. By rigorously separating how companies talk about the environment from what they actually report doing, the Sentiment–Substance Indicator gives regulators, investors and researchers a transparent, replicable instrument for interrogating the fastest-growing genre of corporate communication on Earth — and a reminder that in the age of green narratives, the most important question is not how confident a company sounds, but how concretely it reports.

Subject of Research: A text-mining indicator measuring narrative–substance alignment in corporate ESG environmental reports of Chinese firms

Article Title: An exploratory indicator of disclosure consistency in ESG environmental reports: Application to Chinese corporations

Article References: Zhou, P., Valente, D., Giannico, V., Elia, M., & Lafortezza, R. (2026). An exploratory indicator of disclosure consistency in ESG environmental reports: Application to Chinese corporations. Environmental and Sustainability Indicators, 32, Article 101561. https://doi.org/10.1016/j.indic.2026.101561

Image Credits: AI Generated

DOI: 10.1016/j.indic.2026.101561

Keywords: ESG reporting, greenwashing, natural language processing, sentiment analysis, China, corporate sustainability, Nature-based Solutions, carbon neutrality, disclosure consistency, Dual Carbon targets, text mining, sustainability indicators