AI-driven ESG assessment: How large language models change corporate sustainability transparency

Introduction: The Dilemma of Corporate Sustainability Monitoring and AI as a Breakthrough

Amid the global net-zero targets and the wave of ESG regulations, corporations face unprecedented disclosure pressure. However, traditional ESG data collection relies on manual review and commercial rating agencies, suffering from narrow coverage, poor timeliness, and low standardization. A recent study published in *Nature Communications* shows that a systematic analysis covering 600 large European companies (2014–2023) reveals a significant "transparency gap" in corporate ESG disclosures—firms in the top rating percentile disclose 22% more indicators than those in the bottom percentile, although this gap is narrowing over time. The study uses a large language model (LLM) framework to automatically extract structured ESG indicators from annual reports and sustainability reports, constructing an open dataset containing 2.9 million observations, providing a new paradigm for global corporate sustainability monitoring.

AI-Driven ESG Data Extraction: From Text to Quantifiable Governance

Traditional ESG research often focuses on a single dimension (e.g., carbon emissions) or relies on qualitative narrative analysis, making cross-industry, cross-period, and multi-dimensional comparisons difficult. The machine learning framework developed in this study can align disclosure indicators with the European Sustainability Reporting Standards (ESRS) and automatically extract quantitative data covering environmental, social, and governance aspects. For example, the model can identify numerical information in reports regarding "proportion of female executives," "employee turnover rate," "water consumption," and "Scope 3 carbon emissions." The breakthrough of this technical approach lies in converting unstructured ESG disclosures into structured, traceable governance information, providing real-time, granular monitoring tools for corporate boards, investors, and regulators.

Key Findings: Imbalance Between Transparency and Performance

The study reveals three core trends: First, the transparency gap is narrowing but persists. Between 2014 and 2023, the rate of disclosure improvement among low-rated firms exceeded that of high-rated firms, but leading companies still occupy a "high-density zone" of disclosure. This means that for latecomer firms, AI-driven disclosure diagnostics can help quickly identify gaps and accelerate compliance. Second, environmental performance improves while social performance stagnates. In the environmental dimension, indicators such as energy intensity and waste recycling rates have improved, but Scope 3 emissions appear "inflated" due to increased supply chain data coverage, so the true emission reduction effect still needs to be assessed by stripping out the disclosure effect. Third, social indicators are "stuck in neutral." Apart from board gender diversity, core social issues such as employee pay equity, occupational health and safety, and community relations have not made substantial progress. This finding warns companies: ESG strategy cannot remain solely on environmental "visible indicators"; deficiencies in social governance capabilities may become a long-term competitive weakness.

Strategic Implications: Three Pathways for AI to Reshape Sustainability Governance1. From “Passive Compliance” to “Active Disclosure Optimization”: With the mandatory enforcement of the EU’s Corporate Sustainability Reporting Directive (CSRD), companies must disclose over a thousand indicators in accordance with the ESRS standards. AI systems can automatically compare existing reports against the standards, generate customized improvement roadmaps, and transform compliance costs from “manual labor” to “algorithm-driven” processes. 2. From “Rating Dependency” to “Internal Governance Engine”: Commercial ESG ratings are often questioned due to opaque methodologies. Enterprises can leverage LLMs to build their own ESG data warehouses, track the environmental and social performance of each business unit in real time, and convert external rating pressure into an internal management feedback loop. 3. From “Static Reporting” to “Dynamic Risk Early Warning”: Research shows that sharp increases in Scope 3 emissions actually reflect improved disclosure in supply chain data rather than substantive deterioration. AI systems can identify such “disclosure effects,” helping management avoid misjudging risk trends and focus on weak links that truly require intervention.