Key principles of RAI
Figure 9.1 – Responsible AI principles
Microsoft has established a Responsible AI Standard, presenting a comprehensive framework that guides the development of AI systems. This framework is grounded in six key principles:fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability, as depicted in the preceding above. They follow two guiding principles: ethical and explainable. These principles form the bedrock of Microsoft’s commitment to a responsible and trustworthy approach to AI. This approach is increasingly vital as AI becomes more integrated into the products and services we use daily. In my opinion, this framework from Microsoft is exceptionally well-rounded for the design of generative AI solutions and should always be a primary consideration when architecting such solutions. A good mnemonic to remember these principles by is “Friendly Robots Safeguard Privacy, Inspire Trust, Assure Safety,” or FAST-PaIRS.
Let’s dive deep into each of these principles with the help of examples.
Ethical and explainable
From an ethical standpoint, AI ought to do the following:
- Ensure fairness and inclusiveness in its statements and tasks
- Hold responsibility/accountability for its choices
- Avoid discrimination against various races, disabilities, or backgrounds
Explainability in AI provides clarity on decision-making processes for data scientists, auditors, and business leaders, enabling them to understand and justify the system’s conclusions. It also ensures adherence to corporate policies, industry norms, and regulatory requirements.
Fairness and inclusiveness
This principle ensures that AI systems do not discriminate, are not biased against certain groups or individuals, and provide equal opportunities for all.
- For example, designing AI systems with features that accommodate users with disabilities, such as voice-activated assistants that can understand and respond to users with speech impairments or AI-driven web interfaces that are navigable by people with visual impairments.
- This article from The New York Times, titled Thousands of Dollars for Something I Didn’t Do discusses the case of an African American individual who was wrongfully charged and fined due to an erroneous facial recognition match. This incident highlights the limitations of AI-based facial recognition systems in accurately identifying individuals with darker skin tones. Such incidents necessitate the need for fairness and inclusiveness principles in AI systems.
Reliability and safety
This focuses onthe AI system being dependable and not posing any harm to users.
For example, an AI system used in a self-driving car must be reliable and safe. It should consistently make correct driving decisions, such as stopping at red lights and avoiding obstacles, to ensure the safety of passengers and pedestrians.
