Generative AI–Based Readability as a Measure of Disclosure Quality: Evidence from ESG Rating Divergence

ESG ratings have become an increasingly important tool for evaluating firms’ non-financial performance. However, substantial divergences across rating agencies weaken their credibility and exacerbate information asymmetry in capital markets. Using CSR reports of Chinese listed companies from 2015 to 2023, this study develops a novel readability metric based on GPT-4.1, fine-tuned with domain-specific corpora. The fine-tuned model outperformed traditional lexical measures and the non-fine-tuned model in scoring stability, discriminative power, and contextual adaptability, thus advancing from simple “model application” to “framework innovation.” In addition, AI-assessed readability was significantly and negatively associated with ESG rating divergence, indicating that higher disclosure quality helps reduce inter-agency disagreement and information asymmetry. In contrast, conventional lexical indicators exhibited no significant effect. Further analysis demonstrated that corporate greenwashing weakened this mitigating role. Mechanism tests provide evidence that readability reduces divergence primarily by improving disclosure quality and limiting negative media sentiment. The effect was stronger among firms with green investors, higher-quality firm–investor interactions, and Global Reporting Initiative-based reporting. Sub-dimension analysis revealed that readability mainly alleviated divergence in the social and governance dimensions, while showing a limited impact on the environmental dimension. This study highlights the essential role of non-financial disclosure quality in enhancing the information environment and improving ESG rating consistency, offering novel financial evidence for investors’ risk assessment and regulatory policy refinement. • Fine-tuned GPT-4.1 captures CSR readability better than traditional measures. • Higher AI-readability scores reduce ESG rating divergence across agencies. • Greenwashing weakens the role of CSR readability in reducing ESG rating divergence. • Effects are stronger with green investors, GRI reporting, and voluntary disclosure. • Disclosure quality and media sentiment explain the readability effect.