Safe AI

Safe AI systems do not, under defined conditions, lead to states where human life, health, property, or the environment is endangered. Safety is a critical characteristic of trustworthy AI that requires proactive design and ongoing management.

Safety Principles:

Responsible Design and Development: Safety considerations should be incorporated from the earliest planning stages through deployment, including:

• Rigorous simulation and in-domain testing
• Real-time monitoring capabilities
• Ability to shut down, modify, or enable human intervention
• Clear documentation of safety risks based on empirical evidence

Risk-Based Prioritization: Different safety risks require tailored approaches:

• Highest Priority: Risks of serious injury or death require urgent attention and thorough risk management
• Context-Dependent: Safety requirements vary based on application domain and potential consequences
• Human-Facing vs. Non-Human-Facing: Systems directly interacting with humans may require higher initial prioritization

Information and Training: Safe operation requires:

• Clear information to deployers on responsible system use
• Responsible decision-making by deployers and end users
• Proper training and competency development

Sector Alignment: AI safety approaches should align with existing safety guidelines and standards in relevant fields such as transportation, healthcare, and industrial systems.

Ongoing Monitoring: Safety is not a one-time achievement but requires continuous assessment, especially as systems operate in real-world conditions that may differ from development environments.

Safety considerations must be balanced with other trustworthy AI characteristics while maintaining focus on preventing harm to people and the environment.