Faculty of law blogs / UNIVERSITY OF OXFORD

Can ChatGPT Personalize Index Funds’ Voting Decisions?


Chen Wang
JSD student at UC Berkeley School of Law


Time to read

5 Minutes

The groundbreaking ChatGPT introduces large language models (LLMs) to the general public. LLMs are AI systems designed to comprehend and generate human-like text using advanced algorithms. Built on state-of-the-art LLMs, ChatGPT has already been employed to decipher Federal Reserve speeches and predict stock movements. My article explores the potential of ChatGPT to assist small passive funds in making self-informed proxy voting decisions that align with their investors’ interests.

Passive funds, particularly index funds that track indices instead of picking stocks, have rapidly expanded, and by the end of 2021, they represented 16% of US stock market capitalization, surpassing the 14% ownership by actively managed funds. These passive funds charge lower fees than active funds due to reduced active management and competition among providers. However, scholars express concern about index funds’ commitment to investment stewardship, citing under-spending on stewardship and excessive deference to corporate managers. The Big Three index fund providers (BlackRock, Vanguard, and State Street Global Advisors) allocate limited resources to stewardship, and their engagement with portfolio companies occurs infrequently. This trend is even more apparent for small index fund providers. Consequently, index funds’ dispersed ownership and low expense ratio may result in a misalignment of incentives, leading to apathy towards improving portfolio companies’ governance to maintain low costs, despite the potential benefits of increased corporate integrity.

Due to limited resources, expertise, and financial incentives, index funds often rely on proxy advisory firms like Institutional Shareholder Services (ISS) and Glass Lewis for voting recommendations. This dependence leads to ‘robovoting’, where funds automatically vote according to proxy advisors’ recommendations. Large players like Vanguard frequently follow these recommendations, but robovoting is more prevalent among small and mid-sized institutional investors.

Proxy advisory firms may struggle to provide tailored recommendations for all investors due to high demand, potentially resulting in varying quality, factual errors, and biases in their advice. This situation can create problems for small passive funds and their shareholders, as robovoting might not always serve their best interests. Consequently, these funds face a dilemma: outsource proxy voting decisions to save costs but risk uncertain service quality and potential economic consequences or conduct self-informed voting at higher costs to enhance decision quality and better serve their shareholders.

Recently imposed SEC regulations, which enhance reporting duties on institutional investors, may encourage more personalized voting as opposed to robovoting. Personalized voting involves institutional investors making self-informed decisions that serve their shareholders’ interests. This approach can signal attentiveness to shareholder interests, improve fund reputation, and help small funds stand out in a competitive market. Market forces, regulatory initiatives, and academic perspectives all support the notion that investment funds should make personalized voting decisions to better align with their shareholders’ interests and preferences.

In light of the growing influence of index funds and the advancements in AI, my article conducts an experiment to assess the capabilities of both the zero-shot ChatGPT model and a fine-tuned ChatGPT model in generating personalized proxy voting recommendations.

  1. Zero-Shot ChatGPT Model:

We employed the zero-shot ChatGPT’s GPT-4 model (default model without additional training), to generate comprehensive proxy voting guidelines. The zero-shot model was prompted to assume the role of an experienced corporate governance expert and to generate voting guidelines reflecting its standards and perspectives on various corporate governance issues.

The model provided recommendations on numerous factors, including board composition, executive compensation, shareholder rights, and ESG factors. The ChatGPT model also constructed regression models to assess a corporation’s overall corporate governance. However, the zero-shot model was unable to detect the negative implications of factors such as staggered boards or poison pills when constructing regression models, hindering its ability to generate an accurate governance score for a corporation.

The zero-shot model was also tested by providing voting recommendations on seven proposals in Bank of America’s (BOA) 2022 proxy statement. It detected a conflict-of-interest issue in the election of the board and recommended voting for three ESG proposals that the board recommended against. The result demonstrates GPT-4’s likely pro-ESG inclination.

  1. Fine-Tuned ChatGPT Model:

To enhance the model’s reliability in offering proxy voting advice, we fine-tuned it using domain-specific data from Bank of America’s 2022 proxy statement and proxy voting guidelines generated by the zero-shot GPT-4 model. This data was used to fine-tune the GPT-3.5 text-davinci-003 model. The fine-tuned model then assumed the role of an experienced corporate governance expert, providing recommendations for or against all proposals in Bank of America’s proxy statement, disregarding any suggestions from the board and relying solely on the generated voting guidelines.

  1. Comparing Recommendations:

 Table 1: Recommendations of BOA Board, Zero-Shot GPT-4 and Fine-Tuned Davinci-3 (The subject matter of each proposal was also summarized by the fine-tuned text-davinci-3.)


Subject Matter

BOA Board

Zero-Shot GPT-4

Fine-Tuned Davinci-3


Electing directors





Approving executive compensation





Ratifying appointment of independent registered public accounting





Ratifying Delaware Exclusive Forum Provision in Bylaws





Shareholder proposal requesting civil rights and nondiscrimination audit





Shareholder proposal requesting adoption of policy to cease financing new fossil fuel supplies





Shareholder proposal requesting report on charitable donations





Both models exhibited a preference for ESG-related proposals, but the fine-tuned model demonstrated some cognitive limitations, resulting in contradictory recommendations for Proposal 6: it recommended voting against Proposal 6, while supporting Proposal 6 when asked for the reason why it provided such a recommendation.

  1. Key Findings:
  • Potential for AI in Proxy Voting: The experiment demonstrated that AI language models like ChatGPT can offer well-informed recommendations on various proxy voting topics, suggesting that AI has the potential to support decision-making in corporate governance.
  • Limitations of the Zero-Shot Model: However, the zero-shot GPT-4 model encountered limitations in generating detailed recommendations. Token constraints, long-range dependencies and flawed mathematical capabilities made it challenging to provide in-depth analysis, leading to less precise or incomplete recommendations.
  • Improvement with Fine-Tuning: The fine-tuned GPT-3.5 text-davinci-003 model showed improvements in overcoming token limitations compared to the zero-shot GPT-4 model. Results could be further enhanced by fine-tuning the GPT-4 model with domain-specific data to generate more informed recommendations aligned with the context of corporate governance.
  • Cognitive Limitations and Inconsistencies: Despite the improvements, the fine-tuned GPT-3.5 model exhibited some cognitive limitations, resulting in contradictory recommendations for one proposal. This underlines the importance of addressing potential inconsistencies and biases that may arise when using AI-generated recommendations.
  • Importance of Human Oversight: The experiment highlighted the critical role of human oversight in utilizing AI-generated proxy voting recommendations. While AI models like ChatGPT can offer valuable insights, human judgment is essential in identifying and addressing potential biases, inconsistencies, and errors in the recommendations.

As a next step, small index funds can consider fine-tuning ChatGPT using ISS or their own proxy voting guidelines to create personalized voting models. ISS guidelines categorize proxy voting issues and offer a recommendation framework. While the proprietary ISS model is not publicly disclosed, their recommendations are available in the Voting Analytics database. To evaluate the effectiveness of the ISS model, market reactions, such as stock price changes in response to ISS recommendations, can be observed while considering factors like general market conditions, macroeconomic indicators, and company-specific news. After parsing the ISS guidelines and processing available data, ChatGPT can be fine-tuned to create a personalized model based on ISS or funds’ own principles and market reactions. The resulting model can be assessed and iterated upon to enhance its accuracy and usefulness.

In conclusion, funds can fine-tune ChatGPT to develop personalized corporate governance evaluation and proxy recommendation models that align with their priorities, using proprietary and reliable public data sources. By creating personalized voting models, small index funds can better align with their shareholders’ interests and preferences, ultimately leading to improved corporate governance outcomes.


The author’s full article referenced in this post can be found here.

Chen Wang is a JSD student at UC Berkeley School of Law.


With the support of