Introduction

The rise of artificial intelligence (AI) in sectors such as autonomous vehicles and automated policing has caused concerns regarding the ethical implications of integrating AI into domains where it can cause harm if not properly managed (Joshi; Yong). One area of concern is the use of such tools for financial trading, particularly for managed funds such as pension funds. Who will be held accountable for any large losses caused by the tool? Can people trust these models with their data? How can we be sure that AI trading tools will not cause the next big banking crisis? To address such concerns and to ensure responsible AI development, this paper proposes the Andre Graham AI Ethical Framework. This framework consists of five key principles: Privacy, Transparency, Fairness, Accountability, and Autonomy. The following sections will elaborate on each principle and apply them to the context of AI tools trading securities for pension funds.

Privacy: Safeguarding Data and Its Usage

The privacy principle emphasizes the importance of securing data and using it only for intended purposes. For an AI system to respect privacy, it must handle data securely and only use it for its stated purpose. This involves implementing strong data protection measures, collecting only necessary data, and giving individuals control over their information. It is imperative that all developers comply with the data protection laws that are relevant to their projects. This will ensure the public maintains trust in their AI tools. The Graham AI Ethical framework mandates continuous reevaluation of how the privacy and security of an AI tool can be improved.

The UK Data Ethics Framework supports this principle, highlighting the need for data protection and minimal data collection ("Data Ethics Framework"). The framework emphasizes the importance of complying with privacy guidelines such as the GDPR. Similarly, PWC's Responsible AI Toolkit recognizes privacy as a crucial component of ethical AI development ("Responsible AI"). Their framework includes assessing and validating models to ensure they adhere to privacy standards throughout the development lifecycle.

In the context of AI tools trading securities for pension funds, maintaining the security of sensitive data such as the financial and personal information of pension fund members is vital for preserving public trust. Data in this category includes but is not limited to market data, trading strategies, and individual pension account details. The most cutting edge encryption methods should be used and a data governance model should be followed. This will not only keep the member’s data safe from cyber attacks, but will also reduce the probability of internal data leaks. Best practice security systems such as multi-factor authentication and detailed access logs are required within this framework. Access to the trading tool and its data should be strictly controlled by using these measures. Regular third party audits should be conducted to ensure ongoing compliance with data protection laws and best practices. Finally, pension fund members should be clearly informed about how their data is being used and given the opportunity to opt out.

Transparency: Ensuring Clarity and Explainability

Transparency in AI means the decision-making process must be explainable. Organizations need to provide clear reasons for the AI's actions and decisions. This includes making the criteria and data used in decisions accessible and understandable. Regular audits and reports on the system's performance should be conducted and made available.

The AI4People framework emphasizes "explicability," which includes both intelligibility and accountability (Floridi et al. 696). This aligns closely with the transparency principle. For AI systems to be truly transparent, they must provide explanations for their decisions and be open about their overall functioning and limitations. This approach to transparency is particularly relevant for AI systems trading securities for pension funds, where trust and regulatory compliance are crucial. The AI's trading strategies and decision-making processes should be explainable in clear, non-technical language. This could involve using interpretable machine learning models or developing robust explanation methods for more complex models.

Funds using these tools should provide regular reports to members, explaining in understandable terms how the AI system is performing, what major trades it has made, and why. These reports should also clearly communicate the risks and limitations of the AI system. An example of this would be if the AI struggles to perform in certain market conditions, this should be openly disclosed. Furthermore, the fund should maintain a detailed log of all trades made by the AI, including the rationale behind each decision. This log should be available for audit by regulators and, in appropriate forms, to fund members.

Fairness: Promoting Equal Treatment and Ensuring Equity

For an AI system to be fair, it must avoid bias and ensure equal treatment for all pension fund members. This involves careful selection of training data, regular testing for bias, and monitoring outcomes across different protected groups. When disparities are found, they must be investigated and addressed. The goal is to make decisions based on relevant financial factors, not on sensitive personal data like age, gender, or race.

The UK Data Ethics Framework highlights the need to assess and mitigate potential biases in data and models ("Data Ethics Framework"). This aligns with the fairness principle, emphasizing the importance of proactive measures to ensure equitable outcomes. PWC's Responsible AI Toolkit also addresses fairness, arguing in favor of developing AI systems that are "bias-aware" and promote just outcomes (“Responsible AI”).

This can be applied to the use case of pension fund trading, where fairness is critical in ensuring that AI systems do not inadvertently favor groups of fund members over others. For example, the system should never make trading decisions that disproportionately benefit older members at the expense of younger ones, or vice versa. To promote fairness, AI developers should carefully examine historical trading data for potential biases before using it to train the system. They should employ regularization methods that explicitly optimize for fairness alongside financial performance during model training.

Once in use, the tool should be regularly audited to check for unfair outcomes across different demographic groups of fund members. If the tool’s trading strategies are found to favor certain groups, this should trigger an immediate investigation. Fairness checks should be conducted regularly to ensure that the system remains equitable as variables change over time.

Accountability: Establishing Responsibility and Correction Mechanisms

Accountability in AI systems means that there needs to be defined processes for oversight, error correction, and handling disputes. It is imperative that the AI developers understand that they can be held responsible in the event that their tool harms an individual or a group. This includes a governance structure that clearly defines roles for each team member so they understand what aspects of the tool they are accountable for. Clear procedures for addressing concerns or complaints from pension fund members are necessary. Regular third party audits should be conducted to verify compliance with these ethical standards.

The AI4People framework recommends developing new oversight agencies and legal instruments for AI accountability (Floridi et al. 703). This aligns with the accountability principle, highlighting the need for robust governance structures and clear lines of responsibility. PWC's Responsible AI Toolkit emphasizes the importance of accountability, suggesting the implementation of governance frameworks and the development of clear roles and responsibilities for AI system management ("Responsible AI"). This approach to accountability is particularly relevant for pension fund AI systems, where the stakes are high and the potential for harm significant.

In the context of a trading tool for pension funds, accountability could be implemented by creating a dedicated AI oversight committee responsible for monitoring the lifecycle of the tool. This committee should contain experts from all relevant domains such as data science, finance, ethics, and law. There should also be a clear process for pension fund members to raise concerns or complaints about the AI tool's trading decisions. This could include a dedicated portal where members can submit queries or request reviews of specific trading activities. Finally, there should be clear policies in place to handle errors or unfair outcomes produced by the tool. This could include compensation mechanisms for fund members who have been adversely affected by system errors. These measures ensure that those responsible for creating AI systems are held responsible for any harm perpetrated by the system.

Autonomy: Maintaining Human Involvement

While AI tools can improve efficiency, it's crucial to preserve human judgment, especially for complex or high-stakes decisions. AI should augment human decision-making, not replace it entirely. This involves defining appropriate levels of human oversight, establishing intervention procedures, and ensuring fund managers understand the AI system's capabilities and limitations. The AI4People framework introduces the concept of "meta-autonomy" or a "decide-to-delegate" model, where humans retain the power to decide which decisions to take themselves and which to delegate to AI (Floridi et al. 698). This aligns closely with the autonomy principle, emphasizing the importance of maintaining human agency in AI-assisted processes.

In the context of pension fund trading, this could be implemented through a tiered decision-making process. For routine trades within predefined parameters, the AI system could make autonomous decisions. However, for larger trades, unusual market conditions, or decisions that could significantly impact the fund's risk profile, the system should flag these for human review. This ensures that critical decisions benefit from both AI analysis and human expertise. To implement this properly, it is imperative that pension funds hold training sessions where staff will be educated on not only how to use the tools, but also its limitations.

A clear protocol should be established for overriding AI decisions, allowing fund managers to document their reasoning when they disagree with the system's recommendations. Additionally, in extreme market conditions the tool should be built to automatically stop trading and request human intervention. Finally, it is important to consider the long-term implications of AI on workforce skills. As tools become more advanced, there is a risk that human decision-makers might become overly reliant on AI recommendations, leading to a degradation of their own skills and judgment. To counter this, organizations should implement ongoing training and skill development programs to ensure that human staff maintain and enhance their expertise alongside AI systems. This approach ensures that AI remains a tool to augment human decision-making, rather than replace it.

Conclusion

In conclusion, the Graham AI Ethical Framework provides a robust foundation for the responsible development and deployment of AI systems. By adhering to these principles of Privacy, Transparency, Fairness, Accountability, and Autonomy, developers can build trust with their users, satisfy regulatory requirements, and contribute to the creation of a more ethical and equitable AI-driven future. As AI continues to transform the investment landscape, frameworks like this ensure that technological progress aligns with our societal values and ethical standards. The implementation of these principles will require ongoing effort and vigilance, but the benefits in terms of trust, fairness, and long-term sustainability make this investment worthwhile. When applied to pension funds, this framework allows them to leverage the power of AI while maintaining their commitment to the financial wellbeing of their members.

Works Cited

• Floridi, Luciano, et al. "AI4People—An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations." Minds and Machines, vol. 28, no. 4, 2018, pp. 689-707.
• Joshi, Naveen. “5 Moral Dilemmas That Self-Driving Cars Face Today.” Forbes, Forbes Magazine, 3 June 2024, www.forbes.com/sites/naveenjoshi/2022/08/05/5-moral-dilemmas-that-self-driving-cars-face-today/. Accessed 14 Aug. 2014.
• "Responsible AI: Putting Trust at the Core of AI." PwC, www.pwc.com/gx/en/issues/data-and-analytics/artificial-intelligence/what-is-responsible-ai.html. Accessed 14 Aug. 2024.
• "UK Data Ethics Framework." Government Digital Service, 13 June 2018, www.gov.uk/government/publications/data-ethics-framework. Accessed 14 Aug. 2024.
• Yong, Ed. “A Popular Algorithm Is No Better at Predicting Crimes than Random People.” The Atlantic, Atlantic Media Company, 29 Jan. 2018, www.theatlantic.com/technology/archive/2018/01/equivant-compas-algorithm/550646/. Accessed 14 Aug. 2024.