AI sharpens ESG scoring and enables real‑time tracking of environmental and social risks, but gains are uneven: poor data, algorithmic bias and weak governance hamper reliable, trustworthy sustainable finance.

Main Finding

AI (machine learning, deep learning, NLP, transformers/LLMs, reinforcement learning, GNNs, etc.) substantially improves ESG measurement, real‑time monitoring, climate‑risk forecasting, and sustainable‑portfolio construction relative to many traditional approaches, but widespread benefits are constrained by data quality/coverage, algorithmic bias, transparency gaps, and regulatory/governance shortfalls. The paper’s systematic PRISMA review (84 included studies) identifies rapid growth in transformer/LLM research (2019–2025) and highlights both high promise and urgent governance, data, and equity challenges for AI in sustainable finance.

Key Points

• Scope and trend
  • Systematic review of interdisciplinary literature on AI + sustainable finance + ESG using PRISMA 2020.
  • Rapid growth in publications 2019–2025; transformer/LLM work grew dramatically and overtook classical ML by 2024.
• Core AI techniques and applications
  • Supervised ML, ensembles and gradient boosting: improved ESG scoring, credit and risk prediction, and replicable cross‑firm scoring.
  • Deep learning (CNNs, RNNs): processing high‑dimensional data (satellite imagery, sensors), forecasting climate impacts on markets and dynamic ESG trends.
  • NLP and transformers/LLMs: automated analysis of corporate reports, news, regulatory filings, social media for disclosure quality, controversy detection, and narrative‑based ESG signals.
  • Explainable AI (XAI) & responsible AI: feature importance, surrogate rules, fairness audits to increase trust and regulatory compliance.
  • Big‑data analytics & data mining: integration of structured and unstructured alternative data (satellite, supply‑chain, media) for real‑time ESG monitoring.
  • Reinforcement learning: sustainable portfolio optimization and dynamic allocation under ESG constraints.
  • Graph neural networks & network analysis: modeling firm‑investor‑supply‑chain relationships and contagion of sustainability risks.
• Benefits identified
  • Higher predictive accuracy for ESG outcomes and forward‑looking risk metrics compared with many manual/rule‑based systems.
  • Real‑time or near‑real‑time monitoring enabling faster detection of controversies, climate exposures, and disclosure gaps.
  • Potential to reduce subjectivity and inconsistency in ratings and improve regulatory surveillance (RegTech).
• Principal concerns and gaps
  • Data: sparse, noisy, inconsistent ESG labels; regional imbalance (developed‑market bias); lack of standardized ground truth.
  • Algorithmic bias and opacity: black‑box models can reproduce or amplify inequities or misclassify ESG performance.
  • Governance & regulation: fragmented regulatory regimes, unclear liability/accountability for model outputs, and limited standards for model auditing.
  • Emerging technologies underexplored: generative AI, blockchain integration, long‑horizon climate tail‑risk stress testing.
  • Access & equity: data‑rich incumbents may gain outsized advantage; emerging markets underrepresented in studies.
• Practical recommendations noted
  • Invest in standardized, high‑quality ESG datasets and public benchmarks.
  • Integrate XAI, fairness testing, and domain‑specific audits in model pipelines.
  • Develop RegTech/AI governance frameworks and interoperable disclosure standards to reduce greenwashing and misallocation.

Data & Methods

• Review method: PRISMA 2020 systematic literature review.
• Databases searched: Scopus, Web of Science, IEEE Xplore, PubMed (Boolean/subject‑adapted queries combining AI terms and ESG/sustainable finance terms).
• Search yield and selection:
  • Initial records: 1,160 (Scopus 450; WoS 380; IEEE Xplore 210; PubMed 120).
  • After duplicate removal: 875 unique records.
  • Title/abstract screening removed 618 records → 257 full‑text requests.
  • 32 full texts inaccessible; 225 full texts screened.
  • Exclusions after full‑text review: 141 (reasons: off‑topic, non‑empirical/methodological, non‑peer‑reviewed, out‑of‑date).
  • Final included studies: 84 (qualitative synthesis).
• Inclusion criteria: peer‑reviewed empirical, methodological, or review articles that apply AI techniques to ESG analysis, sustainable finance, climate risk modelling, or responsible investment.
• Exclusion criteria: conference abstracts without full text, grey literature, book chapters, editorials, studies focused on AI or finance alone without substantive ESG/sustainability component.
• Evidence synthesized qualitatively (figures referenced include PRISMA flow diagram and publication trend stacked area chart). The literature surveyed spans up through 2025 (submission in early 2026).

Implications for AI Economics

• Market efficiency and capital allocation
  • Improved ESG signals and real‑time monitoring can reduce information asymmetries, potentially improving capital allocation toward sustainable firms and reducing greenwashing-related misallocation.
  • However, model errors or biased signals can systematically misprice ESG risk, shifting capital flows erroneously and creating welfare losses.
• Asset pricing, risk premia, and portfolio choice
  • AI‑driven ESG metrics that are predictive of future fundamentals or climate shocks could be priced into returns, altering expected returns and hedging strategies for climate risk.
  • Reinforcement learning and dynamic allocation under ESG constraints may change portfolio construction norms and risk‑return tradeoffs for institutional investors.
• Concentration, market power and inequality
  • Data‑rich institutions (large asset managers, Big Tech) that control high‑quality alternative data and advanced models may obtain persistent informational advantages, raising competitive and distributional concerns in financial markets—especially versus participants in emerging markets.
• Regulatory economics and policy design
  • Need for standards and audits: regulators should incentivize standardized ESG reporting, public benchmarks, and model transparency to limit market failures from poor information and externalities.
  • RegTech and algorithmic audits can lower compliance costs and improve enforcement—changing the cost structure of sustainable finance and potentially spurring wider adoption.
• Externalities, systemic risk, and model governance
  • Widespread reliance on similar AI models and datasets could create correlated model risk and systemic vulnerabilities (e.g., herd behavior in ESG allocations), implying a role for macroprudential oversight.
  • Model explainability requirements and stress‑testing for climate tail risks should be integrated into prudential frameworks to mitigate systemic mispricing of long‑term risks.
• Research and public‑good actions that would improve economic outcomes
  • Invest in public, high‑quality ESG datasets and open benchmarks to reduce market concentration and increase replicability.
  • Fund interdisciplinary research on model fairness, causal inference in ESG signals, and valuation of climate risks to inform policy.
  • Encourage policies that require disclosures of model usage in investment products so investors can price model risk and governance quality.
• Labor and skills
  • Demand for quantitative and domain‑knowledge hybrids (AI + sustainability expertise) will rise; retraining and education are economically important to realize the benefits of AI in sustainable finance.

Summary conclusion: AI materially improves ESG analytics and sustainable‑finance tools and has meaningful economic consequences for asset pricing, capital allocation, and risk management—but realizing net social benefits depends on resolving data, fairness, transparency, and regulatory challenges via standards, public data, XAI, and RegTech.