Explainable AI is supplanting black‑box models in ESG analytics, increasing transparency and stakeholder trust, but rigorous evidence that explainability changes asset prices, investor behavior, or regulatory outcomes is largely missing.

Main Finding

The systematic review finds a clear and growing transition from conventional machine learning toward Explainable AI (XAI) in ESG analytics. XAI techniques (notably SHAP, LIME, and inherently interpretable models) are increasingly used to improve transparency, accountability, and stakeholder trust in ESG prediction, disclosure assessment, and sustainable investment decision‑making. However, research is methodologically concentrated in quantitative approaches, with limited qualitative and mixed‑method work and important gaps around governance, evaluative standards, and economic impacts of explainability.

Key Points

• Transition to XAI: Literature shows an ongoing shift from black‑box ML models toward methods that produce human‑interpretable explanations (post‑hoc explainers like SHAP/LIME, attention/feature‑importance, rule‑based and sparse models).
• Domains: XAI is applied across ESG rating harmonization, automated disclosure assessment (e.g., text mining of sustainability reports), green bond and sustainable product screening, and portfolio construction with ESG objectives.
• Drivers: Explainability is pursued to increase stakeholder trust, meet regulatory and auditability demands, reduce model risk, and support decision justification in investment processes.
• Methodological pattern: Predominance of quantitative, model‑centric studies (algorithm development, performance + explanation evaluation). Few studies engage stakeholders or use qualitative methods to assess how explanations affect human decision‑making.
• Emerging themes and gaps:
  • Need for standardized evaluation metrics for XAI in ESG (beyond technical fidelity/accuracy to include usefulness, actionability, and trust).
  • Limited work on causal explanations, fairness, and distributional impacts of ESG models.
  • Sparse research on governance structures, regulatory frameworks, and industry adoption costs/benefits.
  • Very little empirical evidence on economic outcomes (asset prices, capital allocation, investor behavior) arising from explainability in ESG tools.
• Limitations of extant literature: Many studies are dataset/model specific, with limited cross‑context validation and few longitudinal or field studies assessing real‑world impacts.

Data & Methods

• Review protocol: Combined TCCM (Theory, Context, Characteristics, Methodology) framework with SPAR‑4‑SLR procedures for a systematic literature review.
• Data sources: Peer‑reviewed articles indexed in Scopus and Web of Science (search and selection following SPAR‑4‑SLR steps).
• Analytic approach:
  • Bibliometric analysis to map publication trends, influential venues/authors, and thematic clusters.
  • Content analysis to code theoretical foundations, contextual settings (finance, corporate disclosure, ratings agencies), methodological choices (algorithms, explainability techniques), and identified gaps.
• Typical methods in the reviewed empirical papers:
  • Machine learning models for ESG prediction: tree ensembles, neural nets, NLP models for disclosure text.
  • XAI techniques: SHAP, LIME, partial dependence plots, attention visualization, decision rules, feature‑importance ranking.
  • Evaluation: model accuracy/ROC, explanation fidelity, case studies, and occasionally user studies; few papers use experimental or causal inference designs.
• Review limitations: scope restricted to indexed peer‑reviewed literature (Scopus/WoS); potential language or publication‑type biases; heterogeneity in datasets and evaluation metrics across studies.

Implications for AI Economics

• Information asymmetry & market efficiency: XAI in ESG reduces opacity of algorithmic signals, lowering informational asymmetries between algorithmic ESG scorers and market participants. This can improve price discovery for sustainability‑linked assets, reduce mispricing of green claims, and alter returns/risk premia tied to ESG.
• Asset pricing & valuation: Explainability may affect investors’ willingness to pay for ESG‑labeled securities. Transparent explanations could create an “explainability premium” (or discount if explanations reveal weaknesses), changing demand and thus pricing dynamics for green assets.
• Capital allocation & real effects: More trustworthy XAI tools can shift capital allocation toward genuinely sustainable firms if explanations enable better identification of real sustainability performance, amplifying real economy impacts of finance.
• Agency, governance & regulatory costs: Regulators and auditors are more likely to accept AI‑driven ESG assessments that are explainable; this may lower compliance costs and accelerate institutional adoption. Conversely, demands for explainability impose development, documentation, and auditing costs affecting incumbents and entrants differently.
• Model risk & systemic risk: Explainability reduces model risk at the firm level but may have ambiguous systemic effects—widespread adoption of similar explainable models could increase herd behavior or model correlation across institutions.
• Labor and market structure: Improved interpretability may change the division of labor between AI systems and human analysts in sustainable finance, affecting demand for specialized human expertise and altering market competitiveness among rating providers.
• Research & policy agenda for AI economics:
  • Quantify economic value of explainability: measure effects of XAI adoption on asset returns, flows into ESG funds, cost of capital, and firm investment.
  • Experimental and quasi‑experimental designs: RCTs with investors, difference‑in‑differences around disclosure/XAI regulation, event studies on adoption announcements.
  • Standardization and incentives: study optimal regulatory designs (e.g., disclosure standards for model explanations), certification/audit markets, and incentive structures encouraging robust, interpretable ESG models.
  • Welfare and distributional analysis: assess who benefits from improved explainability (retail vs institutional investors, high‑ vs low‑ESG firms) and potential unintended consequences (greenwashing reclassification, correlated strategies).
  • Metrics & benchmarking: develop economic‑oriented XAI evaluation metrics (actionability, impact on capital allocation efficiency, reduction in mispricing).

Concluding note: The reviewed literature establishes that XAI is a promising means to increase the credibility and usability of AI in ESG analytics, but the field needs more interdisciplinary work—especially economic analyses, causal and experimental designs, governance studies, and stakeholder‑centered evaluations—to fully understand and harness explainability’s macro‑ and microeconomic effects in sustainable finance.