"Searching for Signals of Credible Intent in Corporate Climate Reports" presented by Aldís Elfarsdóttir at the 2026 Sustainability Data Science Conference

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Novel NLP methods and large-scale empirical evidence on the credibility of corporate climate disclosures, including private firms.

🏢実務担当者:Findings help design more credible climate reports by focusing on specific language, consistent reporting, and third-party verification.

🏛政策担当者:Credibility signals (e.g., consistency, specificity) can inform disclosure standards and greenwashing detection.

Corporate sustainability disclosures abound, yet their informational content varies considerably. We ask whether firms are likely to do what they say they will do to reduce their greenhouse gas emissions. To answer this question, we constructed a novel dataset linking voluntary survey responses with emissions estimates and financial controls across over 118,724 firm-year observations including both public and private companies, and we applied natural language processing to identify five empirically grounded credibility signals in how firms communicate about their climate strategies. The signals are emissions data parity, third-party verification, longitudinal reporting consistency, specific language, and specific initiative types. Third-party verification and parity between data sources on emissions add credibility to baseline reports intuitively. As for consistency, we documented 613 distinct annual reporting-pattern sequences and classify firms as consistent (reporting every year from first year of reporting to the end of the sample) or inconsistent. Using an event-study design centered on target adoption and initiative implementation and interpreting results as controlled group differences over time, we saw post-adoption trends in Scope 1 and 2 emissions were significantly more pronounced for consistent reporters. Consistency therefore seems to amplify the signal value of disclosed targets and initiatives. To identify specific language, we applied ClimateBERT, a fine-tuned LLM trained on climate finance reports. Specific passages tend to carry quantitative information, proper nouns such as facility names, and more details as opposed to generic statements. Our central finding reverses a prior literature result: science-based target adoption alone is associated with higher three-year forward emissions. However, for firms that adopt targets and are highly specific in their reports, this association becomes significantly negative, most noticeably among private companies. Firms can also be specific about the types of initiatives they implement. Transportation, behavior change, building energy efficiency, and low-carbon electricity purchases are associated with lower absolute emissions. Process efficiency and low-carbon installation show positive absolute associations, though these firms may still credibly intend to reduce emissions intensity while growing. Firms reporting “other” initiatives show no reliable improvement. We additionally identify an inverse credibility signal: high use of climate language. Climate language aggregates all ClimateBERT-classified text referencing emissions reduction, net-zero, renewables, TCFD, environmental claims, climate commitments, climate risks, and climate sentiment across all survey fields. Firms using higher than median amounts of climate language when reporting to their preferred audience (investors for public companies and supply chain buyers for private companies) exhibit higher forward emissions, suggesting possible greenwashing tendencies. Together, our findings establish that it is not just what firms say, but how they say it that conveys credible intent: how specifically, consistently, and verifiably they report on their climate goals and progress. Using state-of-the-art LLMs to decode complex communication structures, we demonstrate how nuances in climate language predict intent about climate emissions. We measure the interaction between targets, language, and audience to demonstrate heterogeneous effects, while accounting for a richer set of information that was omitted in prior studies, such as reporting history, audience, and under- or over-reporting emissions tendencies, which affect main conclusions. Overall, we contribute the first systematic empirical analysis of private and public disclosure heterogeneity using one of the largest datasets yet.

• openalex https://doi.org/10.25740/dh246gs1160first seen 2026-05-05 08:14:23