Can AI be an Unbiased Proxy Advisor?

Every spring, millions of shareholder votes are cast at annual meetings of U.S. public companies. For many of those votes, the decision isn't really made by the investor, but rather by a proxy advisor. Firms like Institutional Shareholder Services (ISS) and Glass Lewis wield enormous influence over corporate governance by issuing voting recommendations that institutional investors routinely follow. ISS is used by the vast majority of institutional investors, and roughly 25% of mutual funds outsource their voting entirely to its recommendations. Yet the industry has long faced a troubling question: how do we know those recommendations are actually unbiased?

The concern isn't abstract. ISS and Glass Lewis both sell consulting services to the same corporations whose governance proposals they later evaluate. Both firms maintain that their advisory and consulting operations are walled off from each other; ISS has even described a formal "firewall" between the two businesses. But the lack of disclosure about client identities makes that claim nearly impossible to verify. Regulators and investors have complained for years. The debate continues, largely in circles, because nobody has had a clean benchmark against which to assess whether recommendations deviate from what the guidelines themselves would predict.

That's the gap our recent working paper tries to fill, and we do it by hiring a new kind of proxy advisor: artificial intelligence.

The AI Analyst

Our approach is straightforward in concept, if not in execution. We take the publicly available voting guidelines that ISS publishes each year, and we feed them to a large language model alongside the actual text of shareholder proposals: the sponsor's supporting statement and management's opposition statement. We then ask the AI to issue a binary "For" or "Against" recommendation, just as ISS would. Because the AI has no financial relationship with any firm, no consulting revenue at stake, and no incentive to deviate from the stated guidelines, its recommendations represent a kind of rules-based benchmark driven by the ISS guidelines.

We use Meta's Llama-3.3-70B-Instruct model and run it in a deterministic setting, meaning the model always produces the same output for the same input. The resulting recommendation is therefore both transparent and replicable. 

What We Find

Over a sample of more than 5,000 shareholder proposals spanning 2009 to 2021, the AI agrees with ISS about 74% of the time when given both the proposal text and the relevant ISS guideline. That figure rises when the guidelines are more explicit and falls when they are more subjective, which is itself an interesting finding, because ISS's guidelines have become dramatically more subjective over time. In 2010, fewer than 15% of applied guidelines were classified as "Case-by-Case." By 2021, that figure had risen to nearly 45%. The shift toward open-ended, judgment-intensive criteria is striking, and it opens the door to the kind of discretion that critics have long worried about.

We next analyze how the AI recommendations relate to shareholder support. ISS recommendations are unsurprisingly associated with higher shareholder voting support. But controlling for the impact of the ISS recommendation, we find something interesting: the AI recommendation is more predictive of eventual shareholder support than a positive ISS recommendation alone. The AI appears to be capturing something real about shareholder preferences. 

The Consultant Effect

The most pointed finding concerns governance consultants. Many firms disclose in their proxy statements that they have retained a third-party governance consultant. These consultants advise management on how to engage with proxy advisors and position their governance practices favorably. We find that when a firm has such a consultant, the probability of the AI recommending "For" a shareholder proposal while ISS recommends "Against" rises significantly. Keep in mind, managers are nearly universally opposed to shareholder proposals, so an “Against” recommendation from a proxy advisor aligns with their preferences. 

To be clear about what this does and doesn't show: we are not claiming that consultants buy favorable ISS recommendations. We can't establish that, and the consultants we observe may not even be ISS or Glass Lewis themselves. What we can say is that when firms bring in external governance consultants, ISS recommendations become systematically less supportive of shareholder proposals than what ISS's published guidelines, as applied by the AI, would predict. Something changes in the recommendation environment when consultants are present. Whether that's the result of legitimate engagement, information provision, or something more troubling is a question the data cannot fully resolve. 

A Benchmark, Not a Replacement

We are not arguing that AI should replace proxy advisors. Proxy advisory work requires institutional knowledge, relationship-building, and nuanced contextual judgment that no model fully replicates today. What AI can do is serve as a transparency-enhancing benchmark, a way to check whether recommendations drift from the stated guidelines that advisors publish as their supposedly objective framework.

There's also a practical implication for the ES risk debate. We find that AI recommendations on environmental and social proposals are predictive of subsequent ES risk incidents at the firm, even after controlling for ISS recommendations, which show no such predictive power. When the AI says a firm's ES practices are problematic enough to merit a shareholder "For" vote, ES incidents tend to follow. That signal has real economic value.

The proxy advisory industry is not going away, nor should it. But transparency demands more than annually published guidelines that become increasingly unverifiable as they grow more discretionary. Our results suggest that AI-based auditing of advisory recommendations is not just feasible; it's also informative. As the regulatory debate over proxy advisor oversight continues, that seems worth taking seriously.

______________

Matthew Souther is an Associate Professor of Finance at the Darla Moore School of Business. 

Choonsik (Chris) Lee is an Associate Professor of finance in the College of Business at the University of Rhode Island.

This blog is based on a paper presented at the 2026 Corporate Governance Symposium and John L. Weinberg/IRRCi Research Paper Award Competition on 6th March 2026. Visit the event page to explore more conference-related blogs.

The ECGI does not, consistent with its constitutional purpose, have a view or opinion. If you wish to respond to this article, you can submit a blog article or 'letter to the editor' by clicking here.