What Sustainability Disclosures Actually Disclose

Society has invested heavily in voluntary corporate sustainability reporting. In theory, these disclosures do real work: they could help civil-society groups, analysts, and other stakeholders hold firms accountable for externalities that regulation leaves untouched, and they could help markets price risks that financial statements miss. In practice, critics dismiss the reports as mere marketing. They are often unassured and therefore not credible, not comparable across firms or over time, vague rather than verifiable, and cherry-picked to shield bad news.

An entire industry has grown up around fixing these disclosures. Nonprofits and shareholders press companies to say more. Standard-setters have built an alphabet soup of voluntary frameworks, including GRI, SASB, TCFD, CDP, and SBTi. Firms increasingly pay for external assurance, and newer mandatory regimes often piggyback on the voluntary frameworks. But this entire enterprise has proceeded without basic facts about what the reports actually contain, or whether the frameworks firms adopt track better disclosure. The reason is simple: reading thousands of heterogeneous PDFs at scale has been prohibitively expensive.

In our new working paper, we use large language models to close that gap. We analyze more than 15,000 disclosure documents issued by over 2,100 Russell 3000 firms, covering more than 11,000 company-year observations spanning 1998 to 2023. We do two things. First, we pin down metadata that reports present inconsistently and prior datasets largely lack: each document’s reporting period, whether it was externally assured, which voluntary frameworks it adopts, and how many data tables and figures it contains. Second, we build sentence-level measures of the characteristics stakeholders say they care about. Because no objective yardstick of report quality exists, we translate the standard critiques into measurable proxies: specificity, quantitative evidence, promotional language (which we model on the legal concept of puffery and call “fluff”), and disclosure of negative, against-interest news.

Five patterns stand out.

First, sustainability reporting has mainstreamed. Report issuance and the adoption of voluntary frameworks surged after 2015, rising in parallel. Reports also became much longer, both as companies report for longer and in association with framework adoption.

Second, framework adoption is sticky and overlapping. Firms that take up a framework tend to keep it, and frameworks cluster rather than substitute for one another. More than 80 percent of GRI-aligned reports in 2022 were also SASB-aligned, and most adopters hold several frameworks at once.

Third, on the dimensions stakeholders demand, the average report moved the wrong way. As reporting spread, reports became less specific, less quantitative (especially after 2015), and fluffier, even as the use of data tables and figures grew. Part of this reflects a wave of thinner first-time reporters. But for specificity and quantitative content, it also reflects declines within firms that kept reporting. As Figure 1 illustrates, the trends hold among incumbents, not just newcomers.

Figure 1. Disclosure characteristics over time

Figure Note. Annual unweighted average of each disclosure measure, for all reporting companies (solid black line with dots), companies in their first year of reporting (dashed orange line with triangles), and incumbent reporters (those that first reported in 2015 or earlier, dashed green with square markers). The vertical line marks 2015.

Fourth, we find little evidence of learning by doing. Companies that have reported for longer do produce more concrete reports: more specific, more quantitative, less fluffy, with more tables and more negative news. But that gap largely reflects which firms started reporting early, not what firms learn from experience. Once we isolate change within a given firm and net out calendar trends, most of the association disappears. What survives is confined to the first few years of reporting and to a few measures (less fluff, more tables, and slightly greater specificity).

Fifth, voluntary frameworks show measurable but mixed associations with report content. In our main specification, the clearest associations concentrate in the three frameworks that directly govern how a report is composed, as Figure 2 illustrates. GRI adoption is associated with more tables but lower quantitative density; SASB with less fluff but also less quantitative density; TCFD with less fluff and more negative-news disclosure. Other associations are weaker and lose precision under robustness checks (for example, GRI’s association with less specificity). The associations do not point in a single direction, and we find no consistent evidence that adopting a framework leads to improvement on the dimensions we study. Nor can we rule out selection: firms that adopt these tools may simply report differently to begin with. These are associations, not causal effects. Still, a framework whose adopters score low on a dimension it is meant to promote may learn something useful about its reports.

Figure 2. Governance frameworks and disclosure characteristics

Figure Note. Estimated association between adopting each voluntary framework and each disclosure characteristic, expressed as a percentage change relative to the adopting firms’ own pre-adoption average. Estimates come from stacked difference-in-differences models comparing adopters to firms that had not yet adopted; lines are 95% confidence intervals and filled circles indicate associations that survive correction for multiple comparisons. These are associations, not causal effects.

What should directors, standard-setters, and policymakers take from this?

We think the results reframe the familiar voluntary-versus-mandatory debate as something closer to a rules-versus-standards one. In our data, firms rapidly crossed the rule-like margins of sustainability reporting: whether to publish a report at all, and whether to adopt a named framework. These are bright lines. Silence on them is legible and reputationally costly, and voluntary pressure unraveled them quickly. But firms stalled on the standard-like substance of what a report actually says: how specific it is, how quantitative, how forthright about negative news. On those margins, consensus on what counts as good disclosure is uneven—settled for some topics, still forming for others as new ones surface—so reputational enforcement has less to grab onto.

As a result, voluntary regimes that want to move substance, not just adoption, may need firmer agreement on what specific, high-quality disclosure looks like, topic by topic. That agreement is also what lets a standard-like ask harden into a rule, and the voluntary regime we study may have been doing exactly that work, surfacing topics and testing metrics that can later be codified. But codifying an ask is a separate choice from who enforces it (GRI and the ISSB’s IFRS S1 and S2, building on SASB, already write topic- and industry-specific standards this way; the ISSB standards are now being taken up in mandatory regimes as well).

The rapid growth, stickiness, and overlap in framework adoption might look like more like badge-collecting than genuine framework influence. GRI on the one hand and SASB and TCFD on the other differ substantially in audience and purpose: GRI targets double materiality, while SASB and TCFD focus on investors and financial materiality. But adoption overlaps heavily, and few firms dropped GRI when the investor-facing frameworks arrived. Those patterns are consistent with firms accumulating disclosure credentials. But the fact that GRI, SASB, and TCFD are associated with diverging patterns of report content, despite high co-adoption, suggests the frameworks may exert real influence on what companies write.

For standard-setters and regulators, one practical lesson is immediate. Clearer specifications, including required alignment language (as GRI does, distinguishing reports “in accordance” with as opposed to “with reference to” its standard) and data tables released in machine-readable rather than text-embedded formats, would sharply improve both automated and human comparison of reports. For boards and disclosure committees, the practical upshot is that reports can now be read at scale. The questions we ask of 15,000 documents can be asked of any dozen, to see how one company’s reports score on specificity, quantitative content, fluff, and disclosure of negative, against-interest news, and how that compares to its peers.

Our methods have limits. The measures are proxies for what stakeholders have demanded, and we do not know which dimensions matter most to which stakeholders. Our results are associations, not causal effects. But the project provides a factual foundation on which future research and standard-setting can build. To that end, we will make the translated text of all 15,000-plus documents and our full dataset publicly available.

The working paper is available for download here.