Climate talk in annual reports tracks higher carbon emissions at Chinese firms

When Chinese listed companies fill their annual reports with climate and carbon language, their factories and facilities tend to emit more, not less, according to a new study of more than 13,000 firm-year observations from China’s A-share market. The finding, published in Carbon Research, upends the intuitive expectation that companies that talk more about climate risk are the ones cutting their carbon footprints fastest. Instead, denser climate-related narrative attention in corporate reporting is associated with higher firm-level carbon dioxide equivalent emissions, a pattern the authors interpret through the lens of principal–agent conflicts between shareholders and managers.

The research team, led by Mohammad Haseeb with colleagues including Miguel Angel Esquivias and Wisnu Wibowo, analyzed Chinese A-share firms listed in Shanghai and Shenzhen from 2011 to 2024. After filtering out financial firms, distressed special-treatment issuers, and companies with incomplete data, the baseline estimation sample contained 13,336 firm-years covering 2,184 unique firms. Emissions data came from the corporate carbon-emission modules of the CSMAR and Wind databases, constructed on IPCC-consistent inventory principles and covering a Scope 1-type total: combustion and fugitive emissions from owned operations, industrial process emissions, waste-related emissions, and firm-attributable land-use conversion. Purchased electricity and value-chain emissions were excluded from the baseline measure.

To measure how much attention managers devote to climate threats, the researchers built a text-mining dictionary from full-text annual reports. They started with seed terms drawn from a manual reading of more than 600 reports, meteorological disaster terminology, and published Chinese climate-risk lexicons, then expanded the list using a word-embedding algorithm trained on the cleaned report corpus. Three domain experts independently vetted the candidates, and the final dictionary contained 98 terms, including phrases such as climate risk, extreme weather, carbon emission, carbon neutrality, and green transition. The index, called Risk, counts the share of climate-related words in the narrative body of each report, expressed in percentage points. In the estimation sample it averaged 0.286 with a standard deviation of 0.241.

The headline result is a robust positive association. In the preferred specification with firm and year fixed effects, a one percentage-point increase in climate-risk language intensity is associated with roughly a 7.95 percent increase in firm-year CO2e. A more realistic movement of one standard deviation in the index corresponds to about a 1.86 percent rise in emissions. For a firm emitting one million tons of CO2e per year, that translates into roughly 18,000 to 19,000 additional tons, other things equal. The association survives when emissions are scaled by operating revenue, when industry-by-year, province-by-year, and city-by-year fixed effects are added, and under a double machine learning procedure that flexes the control function with random forests and gradient boosting.

The authors are unusually explicit that this is a conditional correlation, not a causal effect. Reverse causality remains a first-order threat: high-emitting firms face more regulatory pressure, investor scrutiny, and reputational risk, and may simply write more climate language into their reports. The team tested instrumental-variable designs using lagged non-local industry peer attention and the firm’s own lagged language intensity. The positive association survived and even grew larger under two-stage least squares, but the researchers declined to claim causal identification, because national industry regulation, technological change, and decarbonization policy could jointly move peer language and focal-firm emissions, violating any credible exclusion restriction. Strong first-stage statistics and a non-rejection of Hansen’s J test, they note, cannot repair that defect.

Why might more climate talk accompany more carbon? The paper organizes the pattern with principal–agent theory. Managers evaluated on near-term accounting performance may respond to salient climate risks by protecting operational continuity rather than pursuing capital-intensive abatement that pays off slowly. Maintaining high-utilization carbon-intensive assets, adding energy-intensive buffers such as cooling and backup power, and deferring low-carbon retrofits all stabilize earnings in the short run while raising emissions. Narrative attention, meanwhile, is cheap: discussing climate risk helps manage external audiences without requiring operating choices to turn green. The authors stress that they did not directly measure short-termism, incentive horizons, executive compensation, or continuity-preserving expenditure, so the agency mechanism remains an organizing reading rather than a tested channel.

Two moderating conditions flatten the positive slope. First, stronger internal control, measured with the DIB internal control index that aggregates control environment, risk assessment, control activities, information and communication, and monitoring, significantly weakens the association between climate language and emissions. Better-governed firms emit less on average, and the interaction term is negative and significant, though the authors caution that the index may also proxy reporting quality, organizational capacity, or general management quality. Second, in a subsample of coal-relevant and energy-intensive industries including mining, power and heat production, metals, chemicals, and building materials, higher assigned province–industry coal productivity co-moves with a flatter narrative-attention–emission slope. Because this efficiency measure is matched by headquarters province and two-digit industry rather than measured at the firm’s own plants, it captures a sectoral condition rather than firm-level technology.

Green innovation shows a more nuanced pattern. Green patent applications, identified from China National Intellectual Property Administration filings, are positively associated with climate-risk narrative attention and negatively associated with same-year CO2e, yielding a negative product of the two paths. But the authors insist this is not causal mediation: patents generally affect production technologies with lags, and risk language, patenting, and emissions may all respond jointly to regulation, expansion, scrutiny, financing conditions, and technological opportunity within the same firm-year. The Sobel statistic and bootstrap interval quantify a product of coefficients, not a validated sequential mechanism.

The association is also unevenly distributed. It is statistically steeper among firms with low digital transformation, measured through keyword frequencies in annual reports; among state-owned enterprises, which face stronger policy and employment objectives that can favor output continuity; among non-high-tech firms that rely on more energy-intensive routines; and among firms in less marketized regions, classified using the National Economic Research Institute’s index of marketization. Formal interaction tests with seemingly unrelated regression comparisons confirm that these slope differences are statistically significant, not artifacts of informal point-estimate comparisons.

The study’s limitations are candid and substantial. The dictionary includes terms conceptually close to the outcome itself, such as carbon emission and carbon trading, so a dirty or heavily regulated firm may score high on narrative attention simply because it writes about carbon more often. A restricted risk-and-threat dictionary that drops promotional green-marketing phrases preserves the positive association, but the authors acknowledge they have no survey, earnings-call, or internal-document benchmark against which to validate the index as a measure of private managerial beliefs. A descriptive contrast around China’s 2020 dual-carbon announcement shows relative post-2020 emission increases in high-carbon-intensity industries, but the authors explicitly refuse to read it as evidence about firm-level climate attention, given confounding from energy-market shocks, the national emissions trading scheme, and COVID disruptions. The evidence, they conclude, supports at most a positive conditional association between climate-related annual-report language and estimated firm emissions in China’s distinctive institutional setting, and policy recommendations about disclosure mandates or executive compensation would require causal evidence the study does not provide.

Subject of Research: The association between climate-risk narrative attention in annual reports and firm-level carbon emissions among Chinese A-share listed firms

Article Title: Climate-risk narrative attention and carbon emissions: internal governance and coal efficiency as moderators, green innovation as a mitigating channel

Article References: Haseeb, M., Esquivias, M. A., Hossain, M. E., Al-Mu’ani, L., Khudoykulov, K., & Wibowo, W. (2026). Climate-risk narrative attention and carbon emissions: internal governance and coal efficiency as moderators, green innovation as a mitigating channel. Carbon Research, 5(1), Article 64. https://doi.org/10.1007/s44246-026-00306-z

Image Credits: AI Generated

DOI: 10.1007/s44246-026-00306-z

Keywords: climate risk, carbon emissions, annual reports, Chinese A-share firms, principal-agent theory, internal control, green patents, coal efficiency, text mining, corporate governance, double machine learning, state-owned enterprises