The correlation between ESG ratings from major providers is approximately 0.54. For context, the correlation between credit ratings from Moody's and S&P sits at roughly 0.99. That single comparison, drawn from Berg, Koelbel, and Rigobon's widely cited 2022 study in the Review of Finance, tells you almost everything you need to know about the state of sustainability ratings: two agencies evaluating the same company on the same broad dimension can reach almost opposite conclusions.
I spend a lot of time analyzing how AI companies are trying to fix this problem, and I think the honest answer is mixed. Machine learning can meaningfully improve the data quality underlying ESG assessments, but it cannot resolve what is fundamentally a disagreement about values and definitions.
Why ESG Ratings Are Unreliable
The ratings divergence problem has several root causes, and they compound each other in ways that make the overall system less trustworthy than any individual weakness might suggest.
Methodological inconsistency. MSCI, Sustainalytics, ISS, and Refinitiv each define ESG using different category structures, different materiality assessments, and different aggregation methods. One provider might weight carbon emissions heavily while another emphasizes board diversity; one might penalize a company for operating in fossil fuels regardless of its transition plans while another rewards the same company for investing in renewables. The frameworks are internally coherent but mutually contradictory.
Self-reported data. The majority of ESG data still originates from companies themselves, through sustainability reports, CDP questionnaires, and regulatory filings. Companies choose what to disclose and how to frame it. A firm might report Scope 1 emissions with precision while conveniently omitting Scope 3 supply chain figures that would triple its carbon footprint. Rating agencies working from the same disclosures may interpret these omissions differently.
Subjective materiality. What counts as material to ESG performance varies by industry, geography, and philosophical stance. Is a tech company's energy consumption more material than its labor practices in content moderation? Reasonable analysts disagree, and their disagreements cascade through the final ratings.
The result is that a portfolio manager trying to build an ESG-compliant fund faces a paradox: the screening tool she uses will produce a materially different portfolio depending on which rating provider she subscribes to. That's not a minor inconvenience; it's a structural failure in the information supply chain.
Three Ways AI Improves ESG Data
Despite the definitional challenges, machine learning is making real progress on the data quality problem. I group the applications into three categories.
1. NLP Extraction from Unstructured Sources
The most mature application of AI in ESG is using natural language processing to extract signals from sources that traditional rating agencies either ignore or process manually. News articles, NGO reports, regulatory enforcement actions, court filings, employee reviews, and social media all contain ESG-relevant information that updates faster than annual sustainability reports.
Clarity AI, which serves institutional investors managing over $30 trillion in assets, processes more than 100 million data points across these unstructured sources. Their models identify specific ESG events (a chemical spill, a labor dispute, a governance controversy) and link them to companies and issue categories in near real-time. This is a genuine improvement over the traditional approach of waiting for a company's next sustainability report to learn about a controversy that happened nine months ago.
Arabesque applies ML across more than 8,000 publicly listed companies, combining financial and ESG signals into what they call an "S-Ray" sustainability score. Their approach is interesting because it treats ESG as a financial variable rather than a moral one; their models attempt to identify which ESG factors are predictive of stock performance, which sidesteps some of the normative weighting debates.
Briink takes a different angle entirely: using NLP to compare what companies say against what they actually do. Their models analyze corporate sustainability claims in marketing materials, press releases, and annual reports, then cross-reference them against concrete disclosures, third-party data, and regulatory records. This is essentially automated greenwashing detection, and I think it may be one of the highest-value applications of NLP in the ESG space. The gap between corporate rhetoric and corporate action is wide, and humans reading sustainability reports are poorly equipped to catch systematic patterns of exaggeration across hundreds of pages of carefully crafted prose.
2. Satellite and Sensor Verification
The most exciting development in ESG data is the shift from reported to observed. Satellite imagery, IoT sensors, and remote sensing technologies allow independent verification of environmental claims that previously relied entirely on self-disclosure.
Sylvera has built its business around this principle, using satellite data to verify carbon offset claims. The voluntary carbon market has been plagued by low-quality credits from forest conservation projects that either don't deliver the promised sequestration or would have been conserved anyway (the "additionality" problem). Sylvera's models analyze satellite imagery to assess actual forest cover changes, biomass density, and deforestation risk, then rate offset projects on a standardized scale. When they find that a supposedly protected forest has lost 30% of its canopy since the offsets were issued, that's a data point no amount of corporate reporting would have surfaced.
This satellite-based verification extends beyond carbon. Methane emissions can be detected from space; water discharge patterns around industrial facilities can be monitored; construction activity at sites claiming to be nature reserves can be identified. The common thread is replacing self-reported numbers with independently observed physical reality.
3. Automated Carbon Accounting
A third category of AI-ESG companies focuses not on rating but on measurement, specifically helping companies calculate their own emissions with greater accuracy and less manual effort.
Persefoni has positioned itself as the "Bloomberg of carbon," building a platform that ingests financial and operational data to calculate Scope 1, 2, and 3 emissions using established accounting frameworks like the GHG Protocol. The ML component helps with data classification (mapping line items in procurement records to emissions factors) and estimation (filling gaps where direct measurement data is unavailable).
Watershed serves a similar function with a stronger emphasis on corporate decarbonization planning; its platform calculates current emissions and then models reduction pathways. Greenly targets small and mid-sized businesses that lack the resources for dedicated sustainability teams, automating emissions calculations by integrating with accounting software and bank feeds.
The carbon accounting layer matters because accurate measurement is a prerequisite for accurate rating. If companies themselves don't know their true emissions, the ratings built on top of that data will be unreliable regardless of how sophisticated the methodology is.
Regulatory Tailwinds
The regulatory environment is creating significant demand for machine-readable, auditable ESG data.
The EU's Corporate Sustainability Reporting Directive (CSRD), which began phasing in during 2024, requires roughly 50,000 companies to report detailed sustainability information using the European Sustainability Reporting Standards (ESRS). These reports must be digitally tagged and third-party assured, which means companies need systems that can produce structured ESG data at audit-grade quality.
The SEC's climate disclosure rules, though legally contested, signal a trajectory toward mandatory emissions reporting for public companies in the United States. The International Sustainability Standards Board (ISSB) has issued its first two standards (IFRS S1 and S2), creating a global baseline for sustainability disclosure.
All of these regulatory frameworks share a common requirement: companies need to produce specific, quantitative, verifiable sustainability data on a regular cadence. That's precisely what the carbon accounting platforms are built to deliver, and it's a structural tailwind for every company in this space.
The Skeptic's View
I want to be honest about the limits of what AI can solve here.
The ESG data quality problem has two layers. The first layer is informational: companies don't report enough data, the data they report is inconsistent across firms, and independent verification is sparse. AI is making genuine progress on this layer. NLP extraction, satellite verification, and automated carbon accounting all produce better inputs than what existed five years ago.
The second layer is definitional: what constitutes good ESG performance? How should environmental impact be weighted against social outcomes? Should a weapons manufacturer with excellent labor practices and low emissions score well? Should a solar panel company with poor governance score poorly? These are questions about values, not about data, and no amount of machine learning can resolve them.
I think the industry sometimes conflates these two layers, marketing AI-powered ESG tools as though better data will automatically produce consensus ratings. It won't. Two analysts with identical data will still produce different ESG scores if they disagree about what matters. The Berg, Koelbel, and Rigobon study decomposed the ratings divergence into measurement differences (different data about the same company), scope differences (different categories being evaluated), and weight differences (different importance assigned to each category). AI can reduce measurement divergence substantially. It has little to say about scope and weight divergence, because those reflect genuine normative disagreements.
Where This Leaves Us
My view is that AI-powered ESG tools are solving the right problem at the right time, but the industry needs to be more precise about what "solving" means. Better data is real and valuable. Clarity AI, Sylvera, Persefoni, and their peers are producing information that is more timely, more granular, and more verifiable than anything that existed a decade ago. Investors using these tools have a meaningful informational edge over those relying solely on traditional rating agencies.
But better data is not the same as better ratings, and better ratings are not the same as better outcomes. The question of whether a particular company is "sustainable" will remain contested as long as reasonable people disagree about what sustainability requires. AI sharpens the inputs to that debate; it does not settle it.
For practitioners, I think the implication is clear: use AI-powered ESG data as a supplement to, not a replacement for, your own materiality framework. The tools are strong enough now to trust the data layer. The judgment about what that data means is still yours to make.