Skip to main content
ESG Ratings and Their Disagreements

AI in ESG Ratings: Machine Learning and the Future of ESG Data

Pomegra Learn

How Is Artificial Intelligence Changing ESG Ratings?

Artificial intelligence and machine learning are transforming the ESG data industry. From natural language processing (NLP) tools that extract ESG disclosures from unstructured documents to machine learning models that predict ESG risks from satellite imagery and alternative data sources, AI is expanding the scope, coverage, and timeliness of ESG information. For an industry struggling with data gaps, inconsistent methodologies, and slow annual update cycles, AI offers compelling solutions. It also introduces new risks: algorithmic bias, opacity, data quality propagation errors, and the automation of imprecise measurements that appear more precise than they are.

Quick definition: AI in ESG ratings refers to the application of machine learning, natural language processing, computer vision, and other AI techniques to ESG data collection, processing, and analysis — enabling automated extraction of ESG information from text, satellite imagery, and alternative data sources at scales that human analysts cannot achieve.

Key takeaways​

  • NLP (natural language processing) is the most widely deployed AI technology in ESG — used to extract ESG data from sustainability reports, earnings calls, news articles, NGO reports, and regulatory filings.
  • Alternative data sources enabled by AI — satellite imagery for detecting methane leaks and deforestation, supply chain mapping from shipping data, social media sentiment for worker conditions — extend ESG assessment beyond what companies voluntarily disclose.
  • AI expands ESG coverage: automated systems can process tens of thousands of companies that analyst-led assessment cannot cover, reducing the large-cap bias of traditional ESG ratings.
  • AI introduces new risks: training data biases, model opacity (making validation difficult), false precision (AI-generated numbers that appear precise but reflect noisy underlying data), and the industrialization of errors at scale.
  • The most promising AI applications complement rather than replace human ESG judgment — providing faster information from wider sources, while analysts assess context, materiality, and significance.

Natural Language Processing Applications​

NLP is the most mature AI application in ESG, with several distinct use cases:

Sustainability report extraction: Automated systems extract specific ESG data points (emissions figures, diversity statistics, policy statements, target commitments) from sustainability reports and annual reports, using trained models to identify and normalize relevant information. This enables systematic coverage of companies that would otherwise require human reading of hundreds of pages per company.

News and media monitoring: NLP systems monitor global news sources in multiple languages for ESG-relevant content — environmental incidents, labor disputes, governance controversies, regulatory actions. These systems can process thousands of articles per hour and flag relevant items for analyst review. This is the backbone of controversy monitoring at scale.

Earnings call analysis: AI systems analyze the transcripts of earnings calls for ESG-relevant language — noting whether management discusses climate risk, workforce issues, or governance changes, and tracking sentiment changes over time. Earnings call language can serve as an early signal for ESG developments before formal sustainability report updates.

Regulatory filing analysis: NLP systems extract ESG-relevant content from 10-K, 20-F, proxy statements, and other regulatory filings — identifying climate risk disclosures, executive compensation ESG linkages, and litigation disclosures that may not appear in sustainability reports.

Alternative Data Sources​

AI enables ESG assessment from data sources that were previously impractical or impossible to analyze at scale:

Satellite imagery: Computer vision applied to satellite imagery enables:

  • Detection of methane leaks from oil and gas facilities (using infrared satellite bands)
  • Monitoring of deforestation in supply chain sourcing regions
  • Assessment of physical asset climate risk (flooding, wildfire proximity)
  • Monitoring of retail and industrial activity (parking lot density, ship traffic)

Companies like GHGSat, Carbon Mapper, and Planet Labs provide the satellite data; ESG analytics firms process it into investment-relevant information.

Supply chain mapping: AI systems can infer supply chain relationships from shipping data, procurement databases, customs filings, and business relationship data — mapping supply chain networks that companies do not voluntarily disclose. This enables supply chain ESG risk assessment that goes beyond what companies self-report about their suppliers.

Worker condition signals: Alternative data sources including job posting data (monitoring workplace quality descriptions), Glassdoor and Indeed review mining (worker satisfaction and safety signals), and social media monitoring can provide signals about labor conditions that traditional ESG assessments miss.

Physical climate risk mapping: AI-integrated climate models map physical climate hazard exposures (flood risk, wildfire risk, heat stress, sea level rise) at asset-level geographic coordinates — enabling company-level physical climate risk assessment based on the actual locations of facilities, not company-level disclosures.

AI applications in ESG assessment​

Opportunities: What AI Can Improve​

Coverage expansion: AI-enabled automated assessment can cover far more companies than analyst-led assessment at comparable cost. This is particularly valuable for small-cap and emerging market coverage where traditional ESG ratings are weakest.

Timeliness: AI systems can update ESG assessments more frequently than annual cycles — detecting controversies in near-real-time, incorporating new disclosures as they are published, and tracking satellite-observed environmental changes continuously.

Independent verification: Satellite-based emissions measurement and deforestation monitoring provide independent verification of company-disclosed environmental data — potentially catching discrepancies between what companies report and what satellites observe.

Scope expansion: AI enables assessment of ESG factors that are practically unmeasurable through traditional questionnaire and disclosure approaches — supply chain labor conditions through alternative signals, biodiversity through land use mapping, community relations through social media and local media analysis.

Risks: What AI Can Get Wrong​

Training data bias: AI models trained on historical ESG data inherit the biases of that data — including the large-cap developed-market bias, the disclosure-quality correlation with scores, and the Western cultural assumptions embedded in ESG frameworks. Models trained to identify "good ESG" from historical examples may perpetuate the same biases as the data they learn from.

Model opacity: Deep learning models that produce ESG assessments often do not provide interpretable explanations for their outputs — making it difficult for investors to understand why a company received a particular score and to identify potential errors. Regulatory requirements for explainability in AI-driven financial analysis create tension with the most powerful (but least interpretable) AI approaches.

False precision: AI-generated ESG numbers may appear precise (3 decimal places, continuous updates) while reflecting noisy underlying data. The precision of the number does not reflect the accuracy of the underlying measurement. Methane emission estimates from satellite data have significant uncertainty ranges; presenting them as precise figures can mislead.

Amplification of errors at scale: When an AI system makes an error — misinterpreting a text excerpt, incorrectly attributing an environmental incident, misclassifying a company's supply chain relationship — it typically applies that error consistently across all similarly processed data. Manual analysis errors are idiosyncratic; AI errors are systematic.

Gaming AI systems: As AI systems become better understood, companies may develop strategies to game them — crafting sustainability report language to score well on NLP systems, understanding which keyword combinations produce positive AI assessments, or structuring disclosures to satisfy automated extraction systems without the corresponding operational reality.

Real-world examples​

Kayrros satellite methane monitoring: Kayrros, an energy intelligence company, uses satellite imagery to independently measure methane emissions from oil and gas operations globally — comparing observed emissions to company-reported figures. Studies using Kayrros data have found that actual methane emissions from some operations significantly exceed company-reported figures, providing independent verification evidence.

Truvalue Labs NLP ESG: Truvalue Labs (now part of FactSet) pioneered NLP-based ESG controversy detection, analyzing thousands of unstructured data sources to produce momentum-based ESG signals. Its approach contrasts with traditional static annual scores by providing continuously updating ESG assessments that respond to breaking information.

Spatial Finance Initiative: The Spatial Finance Initiative (based at Oxford) applies satellite data and geospatial analytics to financial risk assessment, including climate physical risk mapping and supply chain monitoring. Its work demonstrates how AI-enabled geospatial analysis can extend ESG assessment into dimensions that questionnaire-based systems cannot reach.

Common mistakes​

Treating AI-generated ESG scores as more reliable than analyst-generated ones: The AI provenance of a number is not a quality signal. AI systems make systematic errors; analyst systems make idiosyncratic errors. The relevant question is which is more accurate for the specific application, not which sounds more "data-driven."

Underestimating the training data quality requirement: Machine learning models learn from their training data. ESG models trained on biased, incomplete, or low-quality historical ESG data will produce biased, incomplete, or low-quality outputs. Garbage in, garbage out is especially dangerous for AI because the scale of deployment amplifies input quality issues.

Ignoring the governance of AI models in investment processes: Investment committees and risk managers who accept AI-generated ESG inputs without understanding the model architecture, training data, and validation methodology are accepting significant model risk. AI ESG models need the same governance oversight as other quantitative models used in investment processes.

FAQ​

Are AI-generated ESG scores better than traditional human-analyst scores?​

The evidence is mixed and application-specific. For timeliness and coverage breadth, AI systems outperform traditional approaches. For accuracy on complex, context-dependent ESG assessments, human analysts with domain expertise often outperform AI. The best current ESG approaches combine AI data collection and preliminary analysis with human review for materiality, context, and significance assessment.

Can satellite data replace corporate ESG disclosure?​

Satellite data can independently measure some environmental factors (methane emissions, deforestation, surface temperature) but cannot replace disclosure for most ESG dimensions — governance quality, social conditions, employee practices, and strategic commitments require information that satellites cannot observe. Satellite data is most valuable as verification and supplementary evidence, not as a complete substitute.

Are there regulations governing AI use in ESG ratings?​

The EU's AI Act (2024) creates requirements for high-risk AI applications in financial services. ESG rating agencies using AI in their assessments may be subject to requirements for transparency, explainability, and human oversight depending on how their systems are classified. ESG-specific AI regulation is still developing; verify current requirements with legal counsel.

How does AI affect ESG data democratization?​

AI reduces the cost of ESG data collection and processing, potentially making broader ESG coverage available to investors who cannot afford comprehensive data from major providers. This democratization effect is real but partial — the most valuable AI-generated ESG signals (satellite emissions data, supply chain mapping) still require significant infrastructure investment.

What is the role of large language models in ESG analysis?​

Large language models (LLMs) are being applied to ESG analysis for document summarization, question-answering about corporate ESG positions, policy comparison across companies, and draft engagement letter generation. LLMs are effective at synthesizing unstructured information but prone to hallucination — generating plausible-sounding but incorrect information. Their use in ESG analysis requires careful validation to avoid acting on hallucinated ESG claims.

Summary​

AI is transforming ESG data collection through NLP, satellite imagery, and alternative data analysis — expanding coverage, improving timeliness, and enabling independent verification of company-disclosed ESG information. The most significant opportunities are in small-cap and emerging market coverage gaps, real-time controversy detection, and satellite-based environmental measurement that supplements company disclosure. The most significant risks are training data bias, model opacity, false precision, and systematic error propagation. The most effective AI applications in ESG combine automated data collection at scale with human judgment about context, materiality, and significance — using AI to do what it does best (process volume at speed) and humans to do what they do best (assess nuance and significance).

Next​

→ ESG Rating Conflicts of Interest