At the start of the 21st century, almost no companies released ESG-related disclosures, but by 2021, most large publicly traded U.S. firms had converged around voluntary standalone ESG reports as a primary means of documenting their ESG activities.

ESG reports are unstandardized in form and content, since they are not audited, mandated, or regulated in the United States (and most other jurisdictions), and the content of these reports continues to vary widely by firm and industry, as well as over time. ESG reports are among the fastest growing voluntary disclosures in history. Given the demand for this information and the current lack of uniformity, regulators are considering reporting mandates to provide frameworks for disclosing ESG activities.

Our Findings

In our paper, we examine the evolution of ESG reports for S&P 500 companies and explore how the content of ESG reports has evolved in the absence of regulation. We also document how this content changed around the introduction of voluntary disclosure standards that defined a comprehensive set of financially material ESG issues. We study ESG reports at two units of analysis: the document-year level and document-topic year level (“the topic level,” going forward). At the document-year level, we find that the percentage of firms releasing these voluntary disclosures increased from 35% to 86% during this period, although the length of these documents experienced more modest growth. We also find evidence that firms in the same sector increasingly use similar language over time, as do firms across sectors, meaning that firms may be coalescing around a common ESG vocabulary.

We next study the content in these reports (i.e., topic-level analysis) to understand the rich heterogeneity within these documents. We use a semisupervised machine-learning approach that is guided by the Sustainability Accounting standards Board (SASB) standards that identified material ESG issues to learn the meanings of all words and phrases within ESG reports. Specifically, we define topics based on the full set of ESG activities defined by SASB and provide a group of unique seed words for each topic from the standards to train a neural network model that learns the words and phrases used to describe each topic. The algorithm allows us to quantify the semantics, as opposed to the syntax, of each document and produce our topic scores, calculated as a weighted-frequency count of all words on that topic within the ESG report, the term frequency-inverse document frequency.

We find that the language used within topics has evolved as the operating environment has changed. For example, when discussing customer welfare in ESG reports, terms like “vaccine” were uncommon in 2010 but were identified by the algorithm as among the most important terms in 2020, a reflection of the COVID-19 pandemic. In addition to providing unique descriptive evidence of how disclosure language changes, this analysis shows the flexibility of the machine-learning approach in measuring text as language evolves.

We also find that firms devote most of their reports to topics that are material to their sector. We find that, on average, firms disclose 48% more on material topics relative to immaterial topics. We also exploit the staggered sector-level introduction of these standards and use a difference-in-differences empirical specification to uncover whether the amount of information related to material and immaterial topics changes as SASB provides voluntary disclosure guidance on ESG’s financial materiality. We find that the relative amount of material information increased by 11% after the release of voluntary standards.

Given the voluntary nature of both ESG disclosures and the adoption of SASB standards, understanding how firms converge toward material disclosures remains an outstanding but important question. To provide insights into the question of how convergence arises, we examine a subset of firms that were involved in the standard-setting process (i.e., potential disclosure leaders) and compare their disclosure practices to those of other early disclosers. We find that firms that helped develop the standards increased material disclosures at similar rates while the standards were being developed. On the other hand, the material information in the reports of other early disclosers was unchanged in the years leading up to the release of the standards but increased in the post-period in a way that resembled the IWG firms in the pre-period. These results provide evidence that standards, even voluntary ones, can serve as powerful guidance when a large sample of firms chooses voluntary disclosure in the absence of regulation. The analysis also suggests that disclosure leaders (i.e., IWG firms) can learn while doing, while other firms may be equally quick to respond once standards are known.

Concluding Remarks

We are among the first researchers to use state-of-the-art machine learning to clean, parse, quantify, and investigate not just the disclosure choice but also the heterogeneous content of ESG reports. We provide evidence that the content within ESG reports can converge toward material information in the absence of regulation, and that both firms and standard setters can play important roles in this process. Our findings suggest that well-defined voluntary standards and guidance can help improve ESG disclosures and should be of interest to investors and regulators. Our findings have implications on how firms’ disclosures may converge in response to a mandate. This is an important insight as regulators are considering ESG disclosure mandates (e.g., such as the ESG Disclosure Simplification Act of 2021 proposed by the U.S. Congress and the founding of International Sustainability standards Board).