“Do unto others as you would have them do unto you.”
The Golden Rule. This may have been the first Sunday School lesson I learned (thanks, Mrs. Monroe).
The Golden Rule is a moral principle that states that you should treat others as you would like others to treat you. It is a foundational principle across many cultures and religions. As such, the Golden Rule should play an essential role in developing AI systems that deliver meaningful, relevant, responsible, and ethical outcomes.
Let’s use this blog to explore how we might integrate the Golden Rule into the AI Utility Function, the mathematical formula governing the AI model’s decision-making process.
Encoding the Golden Rule into the AI utility function would involve specifying a set of rules – and the variables and metrics against which would measure the effectiveness of those rules – that govern the behavior of the AI system. This would include rules such as:
- The AI system should treat humans with respect and dignity.
- The AI system should not harm or allow humans to come to harm.
- The AI system should be transparent in its actions and explain its decisions when humans request it.
- The AI system should treat humans fairly and impartially without discriminating based on race, gender, or other protected characteristics and traits.
- The AI system should use data responsibly and respect the privacy and consent of data subjects.
- The AI system should be inclusive and work equally well across all spectra of society, avoiding bias and discrimination.
- The AI system should have a positive purpose and contribute to the well-being and flourishing of human beings and the environment.
- The AI system should be explainable and provide understandable and meaningful reasons for its actions and decisions.
- The AI system should be trustworthy and act reliably, consistently, and honestly.
This is a great start and starts to make actionable the aspirations of the White House Office of Science and TechnologyPolicy AI Bill of Rights (Figure 1).
Figure 1: The AI Bill of Rights
These rules become a mandatory checklist for any organization that seeks to design, develop, deploy, and monitor AI models that deliver meaningful, relevant, responsible, and ethical outcomes. And to make these rules actionable, we must integrate these rules, and their associated measures, into the AI Utility Function.
The AI utility function is a mathematical function that defines the goal or goals that the AI system is programmed to optimize.
The AI Utility Function assigns values to certain actions that the AI system can take. It captures the AI system’s preferences over possible alternatives. The higher the value, the more desirable the action or outcome is for the AI system. The AI utility function guides the decision-making process of the AI system by helping it to choose the action that maximizes its expected utility.
Figure 2: Defining the AI Utility Function
Integrating the rules associated Golden Rule into the AI utility function can guide the AI system to behave ethically. To integrate the Golden Rule into the AI utility function, one would want to assign higher utility values to actions or outcomes that are consistent with the Golden Rule and lower utility values to actions or outcomes that violate the Golden Rule. For example, if the AI system is faced with a choice between helping a human in need or ignoring them, it could assign a higher utility value to help them because that is what it would want others to do for it if it were in need.
We could use the following process to integrate the Golden Rule into the AI Utility Function:
- Define context. The Golden Rule can be interpreted in different ways depending on the context and scope of the AI system. For example, an AI system that interacts with customers in a retail store would have different rules than an AI system that monitors traffic flow in a city. Defining the context and scope helps narrow the relevant rules and metrics.
- Align rules. Identifying and aligning the rules that define the Golden Rules to the specific and relevant context and scope of the AI system is critical to the delivery of meaningful, relevant, responsible, and ethical outcomes. For example, in the case of a retail store AI system, rules could include treating customers with respect, providing accurate information, and protecting their privacy.
- Codify rules. Translating and codifying the rules into quantifiable measures is the heart of the process. Once the rules are aligned with the AI system context, the next step is to translate those rules into metrics that measure the effectiveness of those specific rules. For example, in the retail example, if the rule that we were seeking to integrate into the AI Utility Function was to “treat humans with respect and dignity,” then the metrics could include the number of customer complaints, the percentage of correct information provided, and the degree of transparency in data collection.
- Assign weights. Assigning weights to the metrics is necessary to determine the importance of each metric and the level of adherence required. This step requires the involvement and guidance of domain experts and stakeholders who are at the front lines of customer engagement and operational execution. In the retail example, our experts might decide that the weight assigned to customer privacy should be 50% higher than the weight assigned to the accuracy of product recommendations.
- Incorporate into the AI utility function. Once the metrics and their weights are defined, they can then be incorporated into the AI utility function. The AI utility function then maps the inputs to the desired outputs to facilitate the trade-off decisions necessary to deliver meaningful, relevant, responsible, and ethical outcomes.
By identifying and integrating the metrics associated with the Golden Rule into the AI Utility Function, the AI system can be designed, deployed, managed, and monitored to ensure that the behaviors exhibited by the AI system align with the principles of the Golden Rule (Figure 3).
Figure 3: Integrating Golden Rule into AI Utility Function
In Part 1 of the 2-part series on integrating the Golden Rule into the AI Utility Function, we first reviewed the Golden Rule (hey, it’s been a few years since Bible School for some of us). We then decomposed the Golden Rule into a series of rules that could provide actionable and measurable teeth to the AI Bill of Rights.
Then after a quick review of the AI Utility Function, we reviewed a simple process that any Citizen of Data Science could leverage to ensure that AI models are integrating the concepts of the Golden Rule to design, define, develop, and manage AI models in the delivery of meaningful, relevant, responsible, and ethical outcomes.
In part 2, we’ll dive into the specific variables and metrics one could use to integrate the Golden Rule into the AI Utility Function, but not before another (hehehe) lesson in economics.
