How to Operationalise “Responsible AI”
The operationalisation of Responsible AI is a pressing question for many organizations as they scale up their use of artificial intelligence (AI). With the rapid development and implementation of AI technologies, there is an increasing need for consistent practices and approaches across enterprises.
Background
A recent inquiry from a client in the financial services sector raised an essential question: “How do you, and should you, operationalise responsible AI across the enterprise?” This client, recognized for their leadership in data analytics and data science, has been striving to provide enterprise-level guidance for responsible data use. However, the swift evolution of AI necessitates that AI becomes everyone’s concern, highlighting the need for a unified operational strategy.
Understanding Responsible AI
It is crucial to differentiate Responsible AI from Responsible Data Use and Ethical AI. While Responsible Data Use focuses on the methods of data collection, storage, utilization, and sharing—prioritizing privacy, security, and fairness—Ethical AI emphasizes the development and deployment of AI systems that uphold fairness, transparency, accountability, and respect for human values. Responsible AI integrates both concepts while adding elements of system robustness.
The principles of Responsible AI, as summarized from industry standards, include:
- Explainability — Ensuring accurate predictions, traceability, and clarity in decision-making processes of AI.
- Fairness — Utilizing diverse and representative data, being aware of biases, and implementing bias-mitigation strategies.
- Robustness — Ensuring the system is resilient against both intentional and unintentional anomalous data while maintaining strong cybersecurity measures.
- Transparency — Allowing end-users to see operational mechanisms and evaluate strengths and weaknesses.
- Privacy — Protecting personal data from leakage or misuse.
Although these principles are commendable, the challenge lies in translating them into practical Standard Operating Procedures (SOPs) that can be systematically enforced across organizations.
Operational Frameworks
While principles provide conceptual guidance, operational frameworks offer structured approaches that incorporate SOPs, policies, roles, and responsibilities. Developing an operational framework that embodies the principles of Responsible AI is a critical step.
Some principles lend themselves more readily to policy translation than others. For instance:
- The Fairness principle is primarily addressed through existing AI/ML model governance.
- The Privacy principle is typically covered by IT oversight.
- However, translating the Transparency principle poses challenges, necessitating a comprehensive approach.
Key Components of the Responsible AI Operational Framework
The operational framework comprises the following elements:
1. AI Model Governance
Established for over a decade, AI/ML model governance requires updates to accommodate the unpredictable nature of Generative AI. Organizations must stay informed and participate in global efforts to refine governance methodologies.
2. AI Experience Design
Designing AI solutions demands policies that ensure psychological and digital safety. Users may feel discomfort or distrust towards AI; thus, explicit checklists detailing necessary information and warnings should be established. This includes testing for potential digital safety issues, such as privacy concerns and biases.
3. AI Diagnostic
Beyond standard governance, clear architectural design instructions must be in place for troubleshooting AI solutions. Specific procedural steps should guide data scientists and engineers when addressing defined failure and complaint issues.
4. AI Integration
Architectural designs should minimize critical co-dependencies in integrating AI solutions with existing workflows. Solutions must be designed with potential failures in mind, ensuring users can disengage from AI systems without significant disruption.
Example Scenarios
Integrating AI tools like co-pilots into employee workflows illustrates the need for a robust operational framework. Considerations must include:
- Identifying potential loops in AI processing and user steps to address them.
- Detecting and alerting teams to instances of offensive or triggering outputs.
- Establishing SOPs for immediate notification and troubleshooting in case of persistent issues.
- Implementing contingency plans for seamless transitions in the event of AI tool unavailability.
Conclusion
Translating guiding principles into practical operational frameworks is a complex task. Responsible AI extends beyond model development; it encompasses user experience, architecture, and integration design. To effectively operationalise Responsible AI, organizations must commit to comprehensive frameworks that address the multifaceted nature of AI implementation and its implications.
