Page 35 - AI Governance Day - From Principles to Implementation

AI Governance Day - From Principles to Implementation

• Inclusive and Collaborative Governance

– Stakeholder Involvement: Include government, private sector, NGOs, scientists, and
academia; create an international AI Governance Agency with an advisory board of
experts and citizens.
– Education and Legislation: Integrate AI education into schools; develop AI laws and
policies at national and international levels.

• Regulation and Enforcement
– Global Standards: Cooperate internationally to set common standards; establish
national institutes for ongoing regulatory discussions.
– Decentralized Decision-Making: Empower specific units for quicker regulatory
decisions.
• Data Governance

– Fairness: Develop and integrate standardized fairness metrics.
– Privacy: Implement differential privacy and federated learning; set regulatory standards
for privacy.
– Transparency: Mandate clear data usage policies.

• Compute Governance

– Verification: Create reporting standards for compute usage.
– Monitoring: Use compute providers for real-time monitoring and reporting.

• Model Governance
– Evaluation: Standardize safety, ethics, and reliability evaluations; establish third-party
certification bodies.
– Security: Implement encryption and secure access protocols for models.
• Deployment Governance

– Output Control: Develop systems to monitor and constrain harmful outputs.
– Detection: Invest in tools to detect AI-generated content.
• Interoperability and Common Definitions

– Technical Standards: Ensure systems are interoperable globally; establish common
definitions for software, hardware, data, and resources.
– Enforcement Models: Create audit models for balance between innovation and
control.
• Practical Implementation and Testing

– Testing Tools: Develop practical testing tools for AI systems; propose certification
frameworks.
– Continuous Testing: Emphasize ongoing testing for AI safety and efficacy.

• Monitoring and Enforcement
– Frameworks: Develop interoperable assessment frameworks considering diverse
regulatory capacities and application levels.

25