An AI governance framework is a set of rules, processes, and responsibilities that are used to ensure that AI is used in a responsible and ethical manner. It should cover the entire lifecycle of AI systems, from development to deployment to operation.
A good AI governance framework will address the following key areas:
- Ethics: The framework should ensure that AI systems are aligned with the organization’s ethical values and principles. This includes things like fairness, transparency, accountability, and safety.
- Risk management: The framework should identify and assess the risks associated with AI systems, and implement measures to mitigate these risks. This could include things like data security, privacy protection, and bias mitigation.
- Transparency: The framework should ensure that stakeholders have a good understanding of how AI systems work and how they are being used. This includes things like explainability, documentation, and auditing.
- Accountability: The framework should establish clear lines of accountability for the development, deployment, and operation of AI systems. This includes things like roles and responsibilities, decision-making processes, and oversight mechanisms.
AI governance pillars
The White House Office of Science and Technology Policy’s National Artificial Intelligence Initiative Office created an AI governance framework built on the following six pillars:
- Innovation. Facilitating efforts in business and science to harness and optimize AI’s benefits.
- Trustworthy AI. Ensuring AI doesn’t violate civil liberties, the rule of law, data privacy and transparency.
- Educating and training. Encouraging the use of AI to expand opportunities and access to new jobs, industries, innovation and education.
- Infrastructure. Focusing on expanding access to data, models, computational infrastructure and other infrastructure elements.
- Applications. Expanding the application of AI technology across the public and private sectors including transportation, education and healthcare.
- International cooperation. Promoting international collaboration and partnerships built on evidence-based approaches, analytical research and multistakeholder engagements.
Some other components of a strong AI governance framework include the following:
Decision-making and explainability. AI systems must be designed to make fair and unbiased decisions. Explainability, or the ability to understand the reasons behind AI outcomes, is important for building trust and accountability.
Regulatory compliance. Organizations must adhere to data privacy requirements, accuracy standards and storage restrictions to safeguard sensitive information. AI regulation helps protect user data and ensure responsible AI use.
Risk management. AI governance and responsible use ensures effective risk management strategies, such as selecting appropriate training data sets, implementing cybersecurity measures, and addressing potential biases or errors in AI models.
Stakeholder involvement. Engaging stakeholders such as CEOs, data privacy officers and users is vital for governing AI effectively. Stakeholders contribute to decision-making, provide oversight, and ensure AI technologies are developed and used responsibly over the course of their lifecycle.
There is no one-size-fits-all AI governance framework, as the specific needs of organizations will vary depending on their size, industry, and the types of AI systems they are using. However, there are a number of resources available to help organizations develop and implement an AI governance framework, such as the following:
The discourse surrounding AI ethics and governance has advanced in recent years, and governments and international organizations have begun issuing principles, frameworks and recommendations accordingly:
- Singapore issued the Model AI Governance Framework, a sector-, technology-, scale-, business model- and algorithm-agnostic framework that converts relevant ethical principles to practices that can be implemented throughout an AI deployment process. This enables organizations to operationalize the principles.
- The Australian government released the AI Ethics Framework that guides organizations and governments in responsibly designing, developing and implementing AI.
- The European Commission proposed what would be the first legal framework for AI, which addresses the risk of AI and aims to provide AI developers, deployers and users with a clear understanding of the requirements for specific uses of AI.
- The University of Turku (Finland), in coordination with a team of academic and industry partners, formed a consortium and created the Artificial Intelligence Governance and Auditing (AIGA) Framework, which illustrates a detailed and comprehensive AI governance life cycle that supports the responsible use of AI.
There are several elements that the mentioned AI governance frameworks and principles have in common. These components can be used by an organization to inform its own AI governance strategy.
AI Governance Implementation Pathway
The Hourglass Model
Citation: Mäntymäki, M., Minkkinen, M., Birkstedt, T., & Viljanen, M. (2022). Putting AI Ethics into Practice: The Hourglass Model of Organizational AI Governance (arXiv:2206.00335). arXiv. https://doi.org/10.48550/arXiv.2206.00335
AI Goverance Life Cycle
List of AI Governance Tasks
The tasks are distributed among task categories A-H.
- T1. AI system repository and ID
- T2. AI system pre-design
- T3. AI system use case
- T4. AI system user
- T5. AI system operating environment
- T6. AI system architecture
- T7. AI system deployment metrics
- T8. AI system operational metrics
- T9. AI system version control design
- T10. AI system performance monitoring design
- T11. AI system health check design
- T12. AI system verification and validation
- T13. AI system approval
- T14. AI system version control
- T15. AI system performance monitoring
- T16. AI system health checks
B. Algorithms
- T17. Algorithm ID
- T18. Algorithm pre-design
- T19. Algorithm use case design
- T20. Algorithm technical environment design
- T21. Algorithm deployment metrics design
- T22. Algorithm operational metrics design
- T23. Algorithm version control design
- T24. Algorithm performance monitoring design
- T25. Algorithm health checks design
- T26. Algorithm verification and validation
- T27. Algorithm approval
- T28. Algorithm version control
- T29. Algorithm performance monitoring
- T30. Algorithm health checks
C. Data operations
- T33. Data pre-processing
- T34. Data quality assurance
- T31. Data sourcing
- T32. Data ontologies, inferences, and proxies
- T35. Data quality metrics
- T36. Data quality monitoring design
- T37. Data health check design
- T38. Data quality monitoring
- T39. Data health checks
D. Risk and impacts
- T40. AI system harms and impacts pre-assessment
- T41. Algorithm risk assessment
- T42. AI system health, safety and fundamental rights impact assessment
- T43. AI system non-discrimination assurance
- T44. AI system impact minimization
- T45. AI system impact metrics design
- T46. AI system impact monitoring design
- T49. TEC expectation canvassing
- T50. TEC design
- T47. AI system impact monitoring
- T48. AI system impact health check
