A 2025 cross-sectional study of 828 U.S.-listed firms used four fine-tuned large language model (LLM) classifiers (sentiment, commitment, specificity, and quantitative-target ambition) to score corporate sustainability and annual reports, finding that quantitative targets are decoupled from narrative tone, that disclosure intensity rises with size and emissions but inconsistently with targets, and that disclosure styles converge across firms (mimetic behavior). The study leaves open the natural next question: do these narrative indicators carry information about realized climate performance, and do the observed decoupling and mimicry persist, dissolve, or strengthen over multi-year windows? Answering requires a longitudinal panel linking the classifier-derived indicators to realized emissions trajectories, with robustness checks on the threshold rules used to aggregate paragraph-level labels.

Background

Corporate climate disclosures — the sections of sustainability and annual reports in which firms describe climate risks, commitments, and targets — are a central input for investors and regulators assessing the transition to a low-carbon economy. A long-standing concern from legitimacy theory and voluntary-disclosure theory is that such disclosures can be symbolic rather than substantive: firms may engage in greenwashing (conveying a misleadingly positive environmental impression) or copy the reporting style of peers, a behavior called mimetic isomorphism in institutional theory, which erodes the informational value of disclosures. Emissions are conventionally reported in three scopes: Scope 1 (direct operational emissions), Scope 2 (emissions from purchased energy), and Scope 3 (value-chain emissions upstream and downstream, which are the hardest to measure and the most often missing).

Machine measurement of disclosure content became practical with ClimateBERT (Webersinke et al., 2021), a transformer language model — the neural-network architecture underlying modern large language models (LLMs) — adapted to climate-related text. Bingler et al. (2022) used paragraph-level ClimateBERT classifiers to distinguish non-committal forward-looking statements ("cheap talk") from concrete commitments in corporate climate risk disclosures, and a follow-up study (Bingler et al., 2024) related cheap talk to climate initiatives, corporate emissions, and reputation risk. Coen, Herman and Pegram (2022) tested the "talk–walk" hypothesis on 725 corporate sustainability reports against greenhouse-gas changes over a ten-year period and found mixed results: talk about operational improvements predicted performance improvement, while many climate commitments amounted to symbolic greenwashing. Colesanti Senni et al. (2025) built a 64-indicator large-language-model tool to assess corporate transition-plan disclosures and found that carbon-intensive companies disclose more about target setting ("talk") than about implementation ("walk"), with higher disclosure associated with lower emissions.

Hassani, Bahini and Mushtaq (2025) contribute a cross-sectional study of 828 U.S.-listed firms. Four fine-tuned ClimateBERT-family classifiers score climate-relevant paragraphs of one representative report per firm: sentiment (risk, neutral, or opportunity framing), commitment (explicit mitigation/adaptation commitment), specificity (general versus specific, verifiable actions), and target ambition (emission-reduction or net-zero targets). Paragraph-level labels are aggregated to firm-level labels by thresholds — a firm is labeled, for example, as committed when more than 40%40\% of its climate-relevant paragraphs are positive for commitment, with tone and target labels using 30%30\% cutoffs. The resulting firm-level profile is the firm's disclosure maturity — the overall sophistication of its climate communication as captured by the four indicators. Linking these indicators to emissions data from Refinitiv (a financial-market data provider; Scopes 1–3), market capitalization, and sector, the study reports that (1) risk-oriented tone co-occurs with explicit commitments, but quantitative targets such as net-zero pledges are decoupled from tone; (2) larger and higher-emitting firms disclose more commitments and actions, though inconsistently with quantitative targets; and (3) disclosure styles converge strongly across firms, indicating mimetic behavior. Because the design observes each firm once, the authors themselves flag in their limitations section that the analysis cannot capture how disclosure strategies evolve over time, that the aggregation thresholds may introduce measurement bias, and that narrative classifications have not been compared with actual climate performance. Their stated future work is to extend the framework longitudinally, considering longer observation windows (they suggest four years or more, since corporate narratives often take substantial time to change), to compare narrative-based classifications with actual climate performance, and to apply richer embedding-based text analysis — representing paragraphs as dense numeric vectors rather than the predefined categorical labels — so that disclosure nuance beyond those labels can be captured.

Problem Statement

Determine whether corporate climate narratives measured by the four-dimension LLM framework of Hassani, Bahini and Mushtaq (2025) — sentiment, commitment, specificity, and quantitative-target ambition, aggregated from paragraph-level classifier labels to firm-level labels — carry decision-useful information about realized climate performance, and how they evolve over time. Concretely, construct a multi-year panel of the same firms' sustainability and annual reports scored with this framework, matched to emissions trajectories (Scopes 1, 2, and 3), and answer three linked questions: (1) temporal dynamics — do disclosure maturity, the decoupling between narrative tone and quantitative targets, and the mimetic convergence of disclosure styles persist, strengthen, or dissolve across years; (2) narrative-to-performance mapping — do current narrative classifications predict, or at least systematically relate to, subsequent realized emissions changes, controlling for obvious confounders such as firm size, sector, and emissions level; and (3) measurement robustness — are the answers stable under the paragraph-aggregation thresholds (e.g., the 30%30\% and 40%40\% cutoffs) and under at least one richer text representation beyond the categorical labels. The problem does not require demonstrating indirect estimation of emission levels from narrative features: the source authors report that this appeared infeasible without performance-based data, and an answer finding no reliable narrative-to-performance mapping is as valid as one finding it.

References

  1. Hassani, Bertrand Kian; Bahini, Yacoub; Mushtaq, Rizwan (2025). Discourse vs emissions: Analysis of corporate narratives, symbolic practices, and mimicry through LLMs. arXiv:2510.01222. https://arxiv.org/abs/2510.01222
  2. Webersinke, Nicolas; Kraus, Mathias; Bingler, Julia Anna; Leippold, Markus (2021). ClimateBert: A Pretrained Language Model for Climate-Related Text. arXiv:2110.12010. https://arxiv.org/abs/2110.12010
  3. Bingler, Julia Anna; Kraus, Mathias; Leippold, Markus; Webersinke, Nicolas (2022). Cheap talk and cherry-picking: What ClimateBert has to say on corporate climate risk disclosures. Finance Research Letters, 102776. https://doi.org/10.1016/j.frl.2022.102776
  4. Bingler, Julia Anna; Kraus, Mathias; Leippold, Markus; Webersinke, Nicolas (2024). How cheap talk in climate disclosures relates to climate initiatives, corporate emissions, and reputation risk. Journal of Banking & Finance. https://doi.org/10.1016/j.jbankfin.2024.107191
  5. Coen, David; Herman, Kyle; Pegram, Tom (2022). Are corporate climate efforts genuine? An empirical analysis of the climate 'talk–walk' hypothesis. Business Strategy and the Environment. https://doi.org/10.1002/bse.3063
  6. Colesanti Senni, Chiara; Schimanski, Tobias; Bingler, Julia; Ni, Jingwei; Leippold, Markus (2025). Using AI to assess corporate climate transition disclosures. Environmental Research Communications. https://doi.org/10.1088/2515-7620/ad9e88
  7. Schimanski, Tobias; Bingler, Julia; Hyslop, Camilla; Kraus, Mathias; Leippold, Markus (2023). ClimateBERT-NetZero: Detecting and Assessing Net Zero and Reduction Targets. arXiv:2310.08096. https://arxiv.org/abs/2310.08096
  8. Hang, Haotian; Shen, Yueyang; Zhu, Vicky; Cruz, Jose; Li, Michelle (2025). Chitchat with AI: Understand the supply chain carbon disclosure of companies worldwide through Large Language Model. arXiv:2511.00024. https://arxiv.org/abs/2511.00024
  9. Chuang, Marianne; Chuang, Gabriel; Chuang, Cheryl; Chuang, John (2025). Judging It, Washing It: Scoring and Greenwashing Corporate Climate Disclosures using Large Language Models. arXiv:2502.15094. https://arxiv.org/abs/2502.15094
  10. Herman, Kyle S. (2025). Climate Talk, Performance Walk? A Meta-Review of Corporate Climate Commitment and Performance. SSRN. https://doi.org/10.2139/ssrn.5282573
  11. Dong, Yuting; Sun, Ziyuan; Zhou, Yiqiang (2026). Walking the Walk or Just Talking the Talk? Corporate Climate Strategies Under Extreme Weather. Business Ethics, the Environment & Responsibility. https://doi.org/10.1111/beer.70096
  12. Acheampong, Albert; Danso, Albert; Adu-Ameyaw, Emmanuel; Sakariyahu, Rilwan (2026). R&D Disclosure and Carbon Performance: A Machine Learning Analysis of Carbon-Intensive Firms. Business Strategy and the Environment. https://doi.org/10.1002/bse.70594