Proxy Advisors Under Artificial Intelligence: Unverified Reasoning in Shareholder Voting Recommendations

2.1 Lack of Transparency in the Recommendation Process

One of the most persistent criticisms of proxy advisory firms is their lack of transparency regarding how voting recommendations are formulated. While institutional investors rely on these recommendations for proxy voting decisions, the specific methodologies used to assess governance issues remain unclear. Proxy advisors provide benchmark voting policies that specify general criteria for evaluating common shareholder meeting proposals, but they are not required to disclose in detail how individual recommendations are determined, including the weight assigned to different governance factors or the decision-making models used. This limited disclosure raises broader concerns about the reliability and objectivity of proxy advice, making it difficult to assess whether recommendations genuinely contribute to improving corporate governance and shareholder value.

This issue is further complicated by the fact that proxy advisors do not merely aggregate investor preferences but also exercise considerable discretion in shaping governance decisions. Thomas et al (2012) suggest that proxy advisors act as intermediaries that coordinate institutional investors’ voting behavior, thereby improving corporate governance. However, Larcker et al. (2013b) argue that proxy advisors go beyond mere aggregation, instead imposing their own governance standards through their recommendations. This distinction is crucial because if proxy advisors actively influence governance structures rather than simply reflecting investor sentiment, the lack of transparency surrounding their methodologies becomes even more problematic. Without clear disclosure, it is unclear whether their recommendations genuinely serve shareholder interests or introduce governance norms driven by their own biases and incentives.

Researchers have sought to evaluate whether proxy advisor recommendations contribute to shareholder value, as proxy advisors themselves do not systematically disclose assessments of their recommendations’ effectiveness. The findings, however, remain inconclusive. Some studies suggest that proxy advisors assist institutional investors in improving governance, leading to positive shareholder outcomes. For instance, Alexander et al. (2010) show that recommendations by Institutional Shareholder Services (ISS) in contested director elections tend to align with positive shareholder returns, indicating that proxy advisors may enhance governance effectiveness in high-stakes scenarios. Similarly, Dey et al (2024) find that firms receiving low say-on-pay support and subject to ISS scrutiny exhibit increased and sustained shareholder engagement, and that markets respond positively when such engagement is anticipated, suggesting that ISS influences governance not only through standardized recommendations but also by promoting firm-specific dialogue with investors.

However, other studies raise concerns about the potential misalignment between proxy advisor guidelines and long-term corporate performance. Larcker et al. (2013a) find that stock option repricing plans conforming to ISS recommendations are associated with lower returns, diminished operating performance, and increased employee turnover, indicating that proxy advisor preferences may not always align with optimal firm strategies. Additionally, Daines et al (2010) conclude that ISS governance ratings do not reliably predict future operating or stock returns, casting doubt on the extent to which proxy advisor recommendations serve as reliable indicators of shareholder value. Further, Larcker et al (2015) report that companies adjusting executive compensation programs to align with ISS’s say-on-pay guidelines experience negative shareholder reactions, suggesting that rigid adherence to proxy advisor standards may sometimes conflict with market preferences.

These mixed findings underscore the need for further research into the mechanisms by which proxy advisors formulate recommendations and how their methodologies influence long-term corporate performance. Given the nature of their services, a natural question arises: as generative AI advances, to what extent are proxy advisors integrating AI into their decision-making processes, and how are the associated risks being managed? While studies have explored the consequences of proxy advisory guidance, little attention has been given to the role of AI in shaping these recommendations. Indeed, similar concerns have already emerged in other advisory services, such as credit rating agencies, where AI-driven decision-making has raised issues of transparency, fairness, and accountability, prompting regulatory scrutiny and debate (ESMA 2024). Given these precedents, it is essential to examine whether proxy advisory services face comparable risks. Section 4 explores these questions in greater detail.

2.2 Undue Influence Over Corporate Decision-Making

Proxy advisory firms are designed to influence institutional voting behavior by providing recommendations on governance matters. The critical question is not whether they exert influence – this is their intended function – but rather how extensive that influence is and whether it reflects sound governance principles or creates distortions in corporate decision-making. If proxy advisors simply provide research that informs independent investor decisions, their role may be unproblematic. However, empirical evidence suggests that their recommendations have a substantial impact on voting outcomes, often in ways that raise concerns about investor autonomy and the broader implications for corporate governance.

Empirical research has sought to quantify the extent of this influence. Brav et al (2022) find that an ISS negative recommendation decreases institutional investors’ support for management proposals by 50.7 percentage points, compared to a 1.8 percentage point decline among retail investors. Similarly, Malenko and Shen (2016) estimate that an ISS negative recommendation on say-on-pay proposals leads to a 25 percentage point drop in shareholder support. Further studies report similar trends. Based on aggregate voting data, Copland et al (2018) show that institutional investors are substantially more likely to vote in favor of management when ISS issues a favorable recommendation. Say-on-pay proposals receive 27.7 percentage points more support, equity-plan proposals 17.3 points more, and uncontested director elections 13.8 points more, compared to cases in which ISS recommends against. These findings indicate that proxy advisors do more than just provide advisory input – they significantly shape voting patterns among institutional investors. The key issue is whether this influence results from the intrinsic quality of their recommendations or if it arises due to other institutional or regulatory factors that distort independent decision-making.

Ideally, proxy advice would serve as one input among many, allowing institutional investors to make independent, well-informed voting decisions. However, in practice, some investors adopt proxy recommendations with minimal scrutiny, raising concerns about over-reliance on these firms. Rose (2021) defines “robo-voting” as the practice of institutional investors automatically voting in near-total alignment with proxy advisor recommendations, typically without independent analysis. He finds that in 2020, 114 institutional investors engaged in this practice – 86 % with ISS and 14 % with Glass Lewis – collectively managing over $5 trillion in assets.^[6] Similarly, Iliev and Lowry (2015) find that over 25 % of mutual funds vote in near-complete alignment with ISS recommendations across all portfolio firms, suggesting that some funds may rely heavily on proxy advisors instead of conducting independent analysis. This heavy reliance on proxy advisors suggests that in some cases, their recommendations are not merely guiding investor judgment but effectively substituting for it.

Beyond influencing shareholder voting, proxy advisors’ influence extends even further – directly shaping corporate decision-making. Companies, anticipating negative proxy advisor recommendations, may preemptively alter governance practices to align with proxy firm preferences, even when these changes may not be optimal for long-term performance.^[7] Nowhere is this dynamic more evident than in executive compensation decisions. Larcker et al (2015), cited earlier, reveal firms’ preemptive actions to adjust pay structures in anticipation of negative proxy advisor recommendations – highlighting how proxy advisor influence can shape governance decisions ex ante, not merely through voting outcomes.

Edmans et al (2023) find that 53 % of directors report having offered lower CEO pay than they otherwise would have, specifically to avoid the risk of a negative recommendation from a proxy advisor. Jochem et al (2021) find that proxy advisor influence contributes to reduced variation in CEO pay across firms, potentially limiting firms’ ability to tailor compensation to their specific needs and weakening incentive structures. Cabezon (2025) similarly finds that proxy advisor pressure has led to standardized pay structures, which is associated with lower shareholder value. These findings raise concerns that proxy advisors may not only influence voting outcomes but also act as de facto standard setters, dictating governance policies in a way that constrains managerial discretion and limits firm-specific flexibility.

Institutional investors have a natural incentive to minimize the costs associated with researching and analyzing governance issues for proxy voting. This has created a structure where proxy advisors serve as a complementary service, helping investors navigate complex governance decisions. However, over-reliance on these services becomes problematic when proxy advisor recommendations do not necessarily contribute to shareholder value. Empirical research suggests that there may be issues with the quality of proxy advice, raising concerns about whether their influence aligns with sound corporate governance principles. Given the extent of their impact, further scrutiny is necessary to assess how proxy advice is generated. One such avenue of inquiry is the potential use of artificial intelligence (AI) in formulating recommendations. If proxy advisors rely on AI models that systematically produce biased or erroneous outputs – without sufficient human oversight to correct these flaws – this could lead to distortions in governance practices across the market.

2.3 Conflicts of Interest

A key concern surrounding proxy advisory firms is the potential for conflicts of interest in their business models. ISS, beyond providing proxy voting recommendations, offers consulting services on governance matters – including executive compensation and shareholder engagement – to corporate issuers through a separate subsidiary. This dual role raises concerns that ISS may issue favorable recommendations to retain or attract corporate clients, compromising its objectivity. To address this, ISS has implemented firewalls to separate its proxy advisory and corporate consulting entities (ISS 2023). However, the effectiveness of these safeguards remains unclear, as independent oversight is limited. Law firms and other professional service providers handling sensitive client relationships also use firewalls, but they operate under strict ethical rules and external regulation. In contrast, proxy advisors lack comparable industry-wide standards or oversight, making it difficult to assess whether their firewalls effectively prevent undue influence. Given ISS’s significant role in corporate governance, greater transparency in managing these conflicts is essential to ensure the integrity of its recommendations.

Theoretical and empirical studies highlight the existence and impact of conflicts of interest within proxy advisory firms. Ma and Xiong (2021) employ theoretical modeling to demonstrate that such conflicts can distort voting recommendations and subsequently reduce firm value. Their analysis suggests that when proxy advisors have business relationships with the firms they evaluate, their recommendations may be biased, leading to suboptimal governance outcomes. Empirically, Li (2018) examines voting recommendations and finds that ISS becomes less favorable toward firms that are likely to be corporate clients – proxied by large companies – when Glass Lewis, a competing proxy advisor, initiates coverage of the same firms. The author interprets this behavioral shift as evidence that commercial incentives may compromise the independence of proxy advice.

Beyond these conflicts, Malenko et al (2025) argue that proxy advisors may have economic incentives to create controversy, as closely contested votes increase the perceived value of their recommendations and, consequently, the demand for their services. If proxy advisors systematically frame governance issues in ways that amplify division among shareholders, they may not only introduce bias into voting outcomes but also drive corporate governance changes that serve their business interests rather than those of investors. Hayne and Vance (2019) further suggest that proxy advisors do not merely function as information intermediaries but act as de facto standard setters, as their voting guidelines compel widespread corporate conformity. This creates an environment where firms feel pressured to align with proxy advisor expectations, often prioritizing compliance over governance strategies tailored to their unique circumstances. The combination of agenda-setting power and financial incentives to cultivate controversy raises fundamental questions about whether proxy advisory firms serve as neutral governance analysts or profit-driven entities shaping corporate policy to sustain their influence.

Concerns over conflicts of interest have become more pressing with the increasing reliance on AI in decision-making processes. If proxy advisors utilize AI trained on past recommendations, there is a risk that biases – if present – stemming from business relationships or other incentives could be systematically reinforced. Moreover, AI’s opacity makes it difficult to assess whether recommendations are genuinely objective or whether commercial interests might either advertently or inadvertently shape outcomes. As AI plays an increasingly prominent role in governance-related decision-making, ensuring that its use in proxy advisory services does not introduce or exacerbate conflicts of interest remains a critical issue.

3.1 United States

Since proxy advisors are typically retained by institutional investors acting as investment advisers, it is useful to begin with the regulatory framework that governs how investment advisers may use proxy advisors. Historically, regulatory attention by the Securities and Exchange Commission (SEC) has focused first on how investment advisers fulfill their duties when relying on proxy advisors, rather than on proxy advisors as independent subjects of regulation.

3.2 Europe

In Europe, the regulation of proxy advisors has emphasized transparency and disclosure, in contrast to more interventionist approaches such as the short-lived U.S. pre-review regime introduced in 2020. This reflects a broader regulatory philosophy within the European Union that favors proportionality, investor information access, and supervision through transparency rather than prescriptive control. In 2013, ESMA assessed the proxy advisory market and found no clear evidence of market failure warranting legislative intervention, instead recommending that the industry develop a code of conduct to improve practices around transparency and conflicts of interest (ESMA 2013). These findings informed the design of the provision on proxy advisors within the Shareholder Rights Directive II (SRD II).^[17] The directive was adopted in 2017 and required to be implemented by all EU member states by June 2019. While some countries experienced delays, the majority of member states had transposed the directive into national law by that date (ESMA and EBA 2023). Article 3j of SRD II, titled “Transparency of proxy advisors,” sets out the primary obligations applicable to proxy advisors.

Under Article 3j(1), they must publicly disclose the code of conduct they apply and explain how it is implemented. Where they do not comply with a code – whether by not applying one or by departing from specific provisions – they must provide a clear and reasoned explanation. This “comply or explain” disclosure must be made freely available and updated annually on their websites. A commonly referenced benchmark is the Best Practice Principles for Shareholder Voting Research (BPP), developed by the BPPG, previously discussed in Section 3.1.2.

In addition, Article 3j(2) mandates proxy advisors to disclose annually a range of information relevant to the preparation of their research, advice, and voting recommendations, so as to adequately inform clients about the accuracy and reliability of their services.^[18] This includes the essential features of the methodologies and models they apply, the main sources of information used, and the procedures they have in place to ensure the quality of their outputs as well as the qualifications of the staff involved. Proxy advisors must also indicate whether and how they take into account national market conditions, legal and regulatory frameworks, and company-specific factors. Furthermore, they are required to disclose the key features of the voting policies they apply in each market, whether they engage in dialogue with the companies they evaluate or with company stakeholders – and if so, the extent and nature of such engagement – and their policies for preventing and managing potential conflicts of interest. This information must be made publicly available on their websites, free of charge, and must remain accessible for at least three years from the date of publication.

Finally, Article 3j(3) requires proxy advisors to identify and promptly disclose to their clients any actual or potential conflicts of interest, or business relationships, that could influence the preparation of their research, advice, or voting recommendations, as well as the actions they have undertaken to eliminate, mitigate, or manage such conflicts.

A central component of the EU framework is the mechanism of monitored self-regulation built around the BPP. Developed in 2014 by the BPPG, the BPP were created in response to recommendations issued by ESMA in its 2013 report. In 2019, the BPP were revised to align with the requirements of SRD II, and an Independent Oversight Committee (IOC) was established. The IOC, composed of representatives of investors, issuers, and academics, conducts annual assessments of signatories’ statements and may issue recommendations for improvement.^[19]

While the IOC operates within the context of industry self-regulation, institutional oversight at the EU level is provided through ESMA, which combines private and public mechanisms. ESMA plays a coordinating role, including tracking SRD II implementation and evaluating its effectiveness. In its 2023 report, ESMA acknowledged that the current model of monitored self-regulation appears generally robust, but also identified areas where improvements are needed. The report recommended several improvements to the current framework, including greater specificity and clarity in disclosures, particularly regarding data sources such as ESG inputs; the potential for ESMA to play a role in addressing breaches of codes of conduct; and more effective disclosure of conflicts of interest.^[20]

A related but distinct approach can be observed in the United Kingdom. The United Kingdom transposed SRD II into domestic law through the Proxy Advisors (Shareholders’ Rights) Regulations 2019,^[21] which entered into force prior to Brexit. These regulations impose transparency obligations on proxy advisors similar to those under Article 3j and designate the Financial Conduct Authority (FCA) as the competent authority for oversight. The FCA has the power to investigate non-compliance and impose sanctions where necessary. Following the UK’s departure from the EU, this framework has remained in place without substantive change, and the FCA continues to monitor compliance through public disclosures and supervisory engagement.^[22]

In addition to this statutory framework, the UK Stewardship Code 2020 functions as a soft-law instrument that complements SRD II by promoting higher standards of conduct and transparency. The Code operates on an “apply and explain” basis, requiring signatories not only to adhere to its principles but also to explain how they have applied them in practice.^[23] For example, service providers such as proxy advisors are expected to explain how they have identified and responded to market-wide and systemic risks (Principle 4), whether and how they have sought clients’ views and the rationale for their chosen approach (Principle 5), and what internal or external assurance they have obtained in relation to stewardship support and why that approach was chosen (Principle 6). Unlike in the United States, however, proxy advisors in the United Kingdom have historically exerted less influence, in part because strong institutional investor trade groups have assumed many of the functions proxy advisors perform elsewhere (Tuch 2019, p. 1488).

3.3 Selected Asian Jurisdictions

Unlike the European Union, which has adopted a harmonized approach through directives such as SRD II, Asia lacks a unified regulatory framework governing proxy advisors. Regulatory approaches differ substantially across jurisdictions, shaped by each country’s institutional environment and market dynamics. Given this diversity, this section does not attempt to provide a comprehensive survey. Instead, it focuses on three jurisdictions – Japan, South Korea, and India – that illustrate contrasting regulatory strategies toward proxy advisory firms and highlight distinct challenges in ensuring transparency and accountability.

4.1 Current and Emerging Use of AI in Proxy Advisory Services

Although no systematic data exists on the deployment of AI in the proxy advisory industry, ESMA identified certain patterns in its 2022 investigation.^[56] This research, conducted between April and November of that year, drew on interviews and voluntary questionnaire responses from a range of financial market participants, including proxy advisors.^[57] In addition, ESMA organized dedicated workshops that brought together industry experts, academics, regulators, and representatives of international organizations to collect use cases and assess perceived risks.^[58]

Based on the information collected through these multiple channels, ESMA noted that some proxy advisory firms use AI to gather, synthesize, and process the information that underpins their research and voting recommendations.^[59] According to ESMA, the development of such tools is particularly driven by increasing demand for ESG-related analysis from institutional investors and other stakeholders.^[60] ESMA refers to techniques such as web scraping – the automated collection of information from publicly available sources – and natural language processing (NLP), which is used to extract meaning and structure from text data, as methods that can be used to generate ESG assessments.^[61] These methods could, for instance, enable the extraction and classification of information on emission targets, labor practices, or board independence from CSR reports and integrated disclosures, allowing for the construction of ESG scores or risk profiles.

Indeed, ISS and Glass Lewis respectively offer ESG evaluation schemes known as the “ESG Corporate Rating” and the “ESG Profile,” which are designed to assess a company’s performance in ESG areas and support investor decision-making, including proxy voting and engagement strategies.^[62] These evaluations cover corporate ESG policies, outcomes, and disclosure quality, providing a structured framework for assessing company practices through both quantitative and qualitative indicators.^[63] Both firms base their assessments primarily on publicly available sources such as proxy statements, annual reports, sustainability disclosures, and corporate websites.^[64] Where appropriate, they may also refer to external databases and international ESG guidelines.^[65] While the specific methods of data collection and analysis vary between providers, both ultimately aim to structure this information into evaluative data points aligned with their respective scoring criteria.

Returning to ESMA’s findings, it noted that the use of AI appears to remain confined to the analytical and preparatory stages and does not currently extend to the formulation of final voting recommendations.^[66] According to ESMA, the proxy advisors interviewed stated that AI does not “directly or autonomously” contribute to the generation of voting advice.^[67] However, as these statements were self-reported, the actual role of AI in internal decision-making processes cannot be externally verified. Under Article 3j of the amended Shareholder Rights Directive (SRD II) and the Best Practice Principles (BPP), proxy advisors operating within the EU are subject to certain transparency requirements. Nevertheless, leading firms such as ISS and Glass Lewis provide little or no public disclosure on their use of AI systems, particularly regarding their technical configurations and governance structures. This lack of transparency hampers the ability of investors and regulators to evaluate the potential risks or influence that AI may exert over proxy voting advice.

In contrast to this limited disclosure, recent research on large language models (LLMs) has demonstrated that generative AI systems are technically capable of processing and synthesizing diverse types of financial data to support advisory functions and decision-making processes (Nie et al. 2024). While the use cases identified in such studies primarily concern investment strategies and portfolio management, the underlying capabilities – namely, the ability to interpret financial documents, identify relevant variables, and generate text-based recommendations – are transferable to proxy advisory workflows. Applied to the proxy context, these models could be trained to produce draft recommendations or preliminary assessments of voting positions based on inputs such as company governance structures, historical voting patterns, and the content of shareholder proposals. As such systems become embedded in proxy advisory workflows, human discretion may gradually be confined to reviewing and approving machine-generated outputs. To address this shift, it will be essential to enhance transparency requirements for AI usage by proxy advisors and to reinforce supervisory mechanisms capable of monitoring such integration.

4.2 Declining Transparency in Proxy Advisory Processes due to AI Integration

Regardless of how AI is currently used, its implications for institutional design cannot be overlooked. As discussed in Section 4.1, ESMA found that proxy advisors have thus far limited AI use to supporting tasks such as data extraction and processing. However, even at this stage, it is exceedingly difficult for external observers to determine what data is collected, how it is interpreted, and how it is structured. While human judgment has never been entirely transparent, the integration of AI introduces additional technical layers – such as data parsing, model inference, and parameter tuning – that render the processing pipeline significantly more opaque. As a result, structural information asymmetry deepens, making external scrutiny increasingly difficult. This lack of traceability raises systemic risks that misinterpretations or ambiguous judgments may become embedded in the foundational layers of proxy advice.

As AI systems continue to evolve, their design and implementation could further reduce the transparency of the end-to-end information-processing and decision-making architecture within proxy advisors. Therefore, even if AI’s current role is characterized as merely supportive, this should not be used as grounds to dismiss its institutional significance, particularly given its potential to reshape how information is processed, structured, and ultimately translated into advisory judgments.

This concern is concretely illustrated by the relationship between ESG scores and voting recommendations provided by Glass Lewis and ISS, where ESG scoring frameworks initially positioned as ancillary have become functionally intertwined with advisory outputs. Both firms officially maintain a separation between ESG ratings and voting advice. Nonetheless, Glass Lewis has acknowledged that certain evaluation factors may influence its voting guidance (Glass Lewis 2025, p. 11). In the case of ISS, while ESG scores are not directly applied to its benchmark voting policies (ISS 2025, p. 4), the quantitative and qualitative indicators comprising those scores are structured in a way that enables their integration into advisory services and client decisions. As a result, in both cases, the scoring framework appears to be functionally interoperable with proxy voting operations. Given this overlap, it is essential to ensure transparency in how ESG scores are constructed. While both ISS and Glass Lewis disclose the overall scoring categories and weighting schemes, they offer limited explanation of how individual indicators are aggregated to produce final scores. Consequently, from an external perspective, the reproducibility of the scoring process remains insufficiently secured.

This concern becomes even more serious if AI is eventually embedded not only in scoring but also in the generation of voting recommendations. If the logic of scoring and the criteria for recommendation are internalized within algorithmic models, the institutional feasibility of verifying the consistency and validity of outputs could effectively collapse. Even if human oversight is formally preserved, the actual decision-making could be shaped by architectural constraints and training data, undermining both transparency and accountability at the structural level.

As the preceding discussion illustrates, the use of AI by proxy advisors can significantly erode transparency. This broader concern can be generalized into three analytical dimensions.

First, AI diminishes institutional visibility into the processes that underpin proxy advice. Although proxy advisors have long operated with limited transparency, AI introduces multi-layered, technically complex workflows that make such opacity more entrenched. The classification and scoring conducted by AI systems are based on inputs and processing steps that are not easily interpretable from public disclosures.^[68] While general methodological information is disclosed, the detailed criteria and interpretive logic underlying individual voting recommendations are not fully disclosed, limiting the ability of companies and investors to trace or verify the basis for specific outputs.

Second, AI’s pattern recognition is based on historical data and tends to struggle with firm-specific contexts or non-standard situations. While ISS and Glass Lewis both indicate that their evaluation methods allow for customization based on factors such as industry, geography, and governance structures,^[69] the extent to which such adjustments influence final recommendations remains opaque. Where AI is involved, contextual variables are typically processed through predefined input structures or abstracted into generalized representations. As a result, the nuance of complex institutional realities may be lost or simplified – even in advanced models designed to capture contextual relationships – especially when those models are ultimately optimized to produce standardized outputs.^[70] Human analysts can flexibly adapt interpretive frameworks when faced with unexpected circumstances; in contrast, AI remains constrained by its training data and model architecture, making real-time re-evaluation structurally difficult.^[71] Even if formal flexibility is maintained, the institutional entrenchment of AI-based systems could narrow the practical space for discretionary, context-sensitive judgment.

Third, when the results of AI-driven preprocessing become the foundation for drafting voting advice, the autonomy of investor judgment becomes formalized and potentially hollowed out. If AI-generated summaries and classifications are received as “objective” or “neutral,” investors may mechanically follow them, reinforcing a pattern known as robo-voting. Even if humans are formally responsible for final decisions, the entrenchment of AI outputs may effectively sideline critical thinking and diminish opportunities for dialogue with firms.

This concern is compounded by the limited human capacity available to critically reassess or contextualize voting recommendations at scale. As of 2025, ISS reportedly employs approximately 3,200 staff across 25 locations in 15 countries, covering 74,500 shareholder meetings across more than 100 markets.^[72] While these figures include non-research personnel, the number of analysts directly responsible for voting advice remains undisclosed. ISS serves approximately 4,200 institutional clients and applies more than 380 custom voting policies.^[73] Glass Lewis, by comparison, analyzes over 30,000 shareholder meetings annually across more than 100 markets, serving more than 1,300 institutional investors globally.^[74] Like ISS, it does not disclose the number of research analysts involved in formulating its voting recommendations. In contrast, Moody’s – a global credit rating agency – employed approximately 14,000 personnel as of 2025,^[75] of whom 1,670 were credit analysts and 256 were credit analyst supervisors as of 2021.^[76]

While differences in organizational function and product complexity make direct comparison difficult, the scale of Moody’s analytical staffing provides a useful point of reference. Considering the volume of meetings covered and the number of clients served, the relative size of ISS and Glass Lewis raises legitimate questions about whether they possess sufficient human capacity to meaningfully reassess or contextually interpret AI-assisted outputs. This capacity gap, if present, increases the risk that algorithmic assessments will serve as the de facto foundation for proxy advice.

Even in the absence of full autonomy, AI systems are already deeply embedded in the advisory workflow through preprocessing functions – such as extraction, classification, summarization, and evaluation – which have gradually expanded their role over time to serve as structural prerequisites for judgment formation, and are increasingly shaping the contours of human decision-making. If such preprocessing were to be systematically and continuously integrated into recommendation drafting, the AI-driven processing structure could become the routine starting point for advice formation. This would reduce human intervention to mere confirmation, allowing AI outputs to function as independent, institutionally embedded determinants of proxy advice. In that scenario, the ability of companies and investors to externally verify or reassess the basis of voting recommendations would become even more restricted, weakening the institutional foundations of accountability.

4.3 Convergence and Divergence in Proxy Advice under AI Adoption: Implications for Undue Influence

As discussed in Section 2.2, proxy advisors exert substantial influence over institutional investors’ voting behavior. In particular, negative recommendations by ISS or Glass Lewis have been empirically shown to significantly reduce support rates. This influence is problematic not merely because it is concentrated, but because it does not necessarily align with corporate value creation or the interests of investors. This subsection examines how the structure of undue influence might change with the introduction of AI.

AI may alter this structure in two opposing ways. One possibility is the emergence of “convergence,” in which a single proxy advisor tends to produce similar recommendations for different companies. A structurally similar pattern has been indirectly suggested in ESG rating models. Between 2015 and 2021, the average ESG scores assigned by a major rating agency to Russell 1000 companies increased by approximately 18 %. However, a detailed attribution analysis by D.E. Shaw (2022) indicates that roughly one-third of this increase can be explained by structural changes in the rating system. These include shifts in index composition, reweighting of key evaluation items, and increased disclosure by firms, particularly of environmental data. After adjusting for these factors, a residual score increase of about 12 % remains, which may partly reflect changes in firm behavior but cannot be definitively attributed to them.

While it is unclear whether this rating model employed AI, it appears that rule-based scoring methods predominated during the period in question. The fact that the model change coincided with a broad-based score increase, despite heterogeneous corporate behavior, may suggest – though not necessarily prove – that the model was insufficiently sensitive to firm-specific variation. If so, it raises concerns that evaluation systems relying heavily on standardized formal inputs may systematically produce more uniform outputs.

If AI is introduced into the proxy advice generation process, multiple implementation approaches are conceivable. Yet regardless of approach, unless the model is explicitly designed to handle exceptional contexts and unstructured information, there is a risk that diverse company situations will be stripped of context and that similar inputs will produce overly uniform recommendations as a matter of institutional routine.^[77]

Conversely, AI may also exacerbate “divergence” in proxy advice – that is, situations in which different proxy advisors issue markedly different recommendations on the same proposal. A well-documented manifestation of such divergence appears in ESG ratings. According to Prall (2021), correlations among major ESG ratings providers range from as high as 0.65 (between S&P Global and Sustainalytics) to as low as 0.07 (between ISS and CDP), indicating a lack of consistency in evaluating overall ESG performance. Dimson et al (2020) further demonstrate that such discrepancies persist at the subcomponent level: for example, the correlations between MSCI and Sustainalytics are only 0.11 for environmental, 0.18 for social, and −0.02 for governance scores. These differences stem from divergent choices in the scope of what is measured, methods of measurement, and weighting schemes. Larcker et al. (2022), referencing these studies, argue that such structural inconsistencies impair the reliability of ESG ratings, confuse investment decisions and fund managers’ communications about ESG quality to investors, and ultimately weaken incentives for companies to improve their ESG practices.

This type of divergence may also arise when AI is applied to proxy advice generation more broadly, beyond ESG scoring. AI-generated proxy advice may process a wide range of information beyond ESG, but if different proxy advisors adopt divergent sets of evaluation criteria and processing methods, their recommendations on the same proposal may differ significantly as a matter of institutional output. Importantly, the mere existence of diversity in evaluation frameworks is not inherently problematic; on the contrary, such pluralism may enhance the robustness of perspectives in the market. The problem arises when the source of such divergence is opaque to external observers, and when outputs differ without sufficient explanation of their underlying logic. The key issue is not whether divergence exists, but whether it is explainable, and to what extent evaluators can disclose and justify the basis for their judgments. If AI accelerates automation while weakening such accountability, institutional transparency may be undermined, creating the risk of confusion and mistrust among investors, firms, and regulators.

4.4 Institutional Entrenchment of Conflicts of Interest through AI

As discussed in Section 2.3, proxy advisors face a dual conflict of interest stemming from the simultaneous provision of consulting and voting recommendation services: (1) an incentive to issue favorable recommendations to client companies, and (2) an incentive to intentionally provoke controversy in order to increase the influence and market value of their advice. The introduction of AI may exacerbate these issues by making them more opaque and institutionally entrenched.

First, with regard to issue (1) – the incentive to issue favorable recommendations to client companies – AI has the potential to produce particularly serious institutional consequences.^[78] There are two primary pathways through which AI becomes embedded in this structure. The first is when a proxy advisor intentionally provides favorable recommendations to a specific client company and deliberately supplies the AI model with training data containing both the consulting content and the recommendation outcome. In this case, the AI effectively inherits a policy of “favoring the client under certain conditions” and functions to reproduce and legitimize that stance.

The second pathway arises even when the advisory division has not intentionally made preferential judgments. If information related to a consulting relationship with the client company remains in internal documents and is included in the AI’s training data, the model may statistically learn a favorable bias toward that company. Even if a formal firewall exists between the advisory and consulting divisions, institutional pathways for data transmission persist – such as (1) consulting materials being shared internally as “good governance cases” or “benchmarks” and reused in training data for recommendations, or (2) personnel rotations between divisions, where individuals previously involved with the client in a consulting capacity later participate in training design or evaluation for the advisory AI. Through such channels, information can cross divisional boundaries and become embedded in the system.

Recent studies have statistically demonstrated that AI models, even when they appear to produce logically correct outputs, often rely not on semantic understanding but on token-level co-occurrence.^[79] For example, experiments by Jiang et al. (2024) show that simply modifying the input expression – without altering the underlying logical structure of a task – can significantly shift the output of large language models (LLMs), suggesting that decisions are driven by “distributions of symbols” rather than “meaning.”^[80] Tokens, in this context, refer to the smallest units processed by the model, including words, punctuation marks, symbols, and proper nouns.

Drawing on this insight, if during fine-tuning (the process of adding new data to a pre-trained model for additional adjustment), certain company documents or proposals are repeatedly introduced as “desirable precedents,” the model may assign statistical weight to patterns where a specific company name, related terms, and formalistic descriptions co-occur with favorable recommendations, ultimately institutionalizing these as legitimate evaluative criteria. While the model’s surface-level output may appear neutral, beneath it lies a distribution of tokens – including company names and associated language – that mechanically co-occurs with favorable judgments and becomes embedded in the AI system as an implicit standard for future recommendations. In this way, AI outputs may continuously reproduce recommendation patterns constructed through consulting relationships with specific clients – unbeknownst to users – while remaining opaque and resistant to correction, and still being treated as legitimate decisions.

Next, regarding issue (2) – the incentive for proxy advisors to intentionally generate controversy to increase the perceived influence and market value of their advice – the introduction of AI also risks institutionally reproducing this structure.^[81] For proxy advisors, when a voting recommendation influences the outcome of a contentious shareholder proposal, that instance is internally valued as symbolic evidence of the advisory service’s effectiveness. What is evaluated is not the substantive economic rationality of the proposal or its contribution to shareholder value, but rather the magnitude of the advisor’s influence on the voting result. As a result, even proposals of relatively minor substantive importance that trigger shareholder concern or debate and lead to a changed outcome through the advisor’s recommendation are likely to be documented and regarded internally as achievements of economic value to the advisor. When such records and documents are incorporated into AI training data, the model learns common patterns from proposals that provoked controversy and significantly affected outcomes – such as the topic of the proposal, the company’s response, and the structure of the recommendation rationale – and may generate future recommendations that reproduce similar effects.

The recommendation patterns constructed in this way are likely to be presented as outputs consistent with past decisions, making them externally appear to be neutral and legitimate AI judgments that are readily accepted within institutional settings. However, what the model is actually learning are the formats and argumentative structures of recommendations that provoked controversy and influenced voting outcomes – patterns that do not necessarily align with shareholder value. If the advisor recognizes that its own recommendations have swayed voting outcomes and generated economic value for the firm, and then seeks to intentionally replicate such patterns, this leads to a clearly structured process of institutional reproduction aimed at maintaining or expanding influence.

Even absent such intent, if past recommendations that were once influential and deemed rational from the standpoint of shareholder value are included in the training data as “high-utility training examples,” the model may statistically assign disproportionate weight to features that commonly appeared in such cases – even if those features are unrelated to substantive shareholder value. As a result, the model may develop a tendency to generate judgments that are influential but misaligned with shareholder interests. These outputs, while formally consistent with past decisions and institutionally validated as coherent, may in fact perpetuate a recommendation structure that is increasingly optimized to serve the advisor’s own influence.

4.5 Transformation of Accountability Structures and Institutional Gaps

The integration of AI into the process of generating proxy advice not only amplifies existing concerns – such as lack of transparency, undue influence, and conflicts of interest – but also gives rise to a structural problem that did not arise under pre-AI conditions: institutional ambiguity in the allocation of accountability.

Even prior to the use of AI, proxy voting advice had already exhibited a certain degree of opacity in its reasoning processes. For example, the decision-making procedures and standards employed by proxy advisors have not been readily subject to external scrutiny, and accountability has often been only nominally maintained. As noted by Larcker and Tayan (2024), the institutional foundations for ensuring accountability in human-generated recommendations have long been fragile.

Despite such fragility at the institutional level, traditional proxy advice operated under the expectation – however unfounded – that some basic form of internal accountability was possible. Recommendations were generally produced by staff applying internal policies to case-specific facts, and although individual authorship was not externally visible, it was presumed – perhaps optimistically – that the organization could internally reconstruct how and by whom decisions were made. This assumption of traceability, though rarely verified, has served as a minimal basis for claims of institutional legitimacy.

When AI models are used to generate proxy advice, it is theoretically possible to trace the data and logic behind the output and to identify who was involved at each stage. However, without appropriate institutional design, such traceability mechanisms may not be implemented in practice. Even when some form of human oversight is present, institutional accountability remains difficult to establish if it is unclear how substantive the review or approval was, or who within the organization assumed those responsibilities. This risk is particularly salient in cases where proxy advice is generated through editorial processes that involve multiple analysts or specialized teams. In such settings, procedural accountability can become difficult to establish, especially if the division of responsibilities is not clearly defined, or if there is no clear record of how decisions were made, by whom, and in what order.

This structural uncertainty, if left unaddressed, may also undermine the effectiveness of institutional safeguards that allow companies to respond to or challenge proxy advisor recommendations. Even if the advice generated by AI appears internally consistent and operationally efficient, the absence of a clearly defined framework of accountability behind the decision may deprive issuers of any meaningful opportunity to contest its basis, question its reasoning, or demand corrective explanation.