Do AI and Lexicon-Based ESG Scores Converge? Evidence from IPO Prospectuses in an Emerging Market

Abstract

The growing use of artificial intelligence (AI) to evaluate Environmental, Social, Governance (ESG) scoring from corporate disclosures raises questions about alignment with conventional lexicon-based scores. Such alignment can be problematic as a lexicon-based approach quantifies the textual density using an ESG dictionary, while LLM-based scores are able to assess the thematic coverage of ESG themes. It remains unclear whether the two approaches produce comparable or distinct results when applied to the same corporate information. This paper examines the construct alignment between LLM-generated and lexicon-based ESG scores using 625 Indian IPO prospectuses issued between 2018 and 2024. A rebalanced dictionary of 630 terms measures disclosure intensity, while contextual scores are obtained by prompting a generative AI model. Both sets of scores are harmonised into quantile categories and compared using confusion matrices, correlation analysis, and chance-corrected agreement. The methods converge for the environmental dimension (EMR = 46.1%; weighted κ = 0.56) but diverge for social (28.2%) and governance (23.0%) disclosure, with only limited alignment overall (31.7%). The findings suggest the two approaches operationalise ESG differently, with implications for construct validity and transparency in ESG analytics.