AI Safety
Measures and engineering practices ensuring AI systems do not cause unreasonable harm to people, property, or the environment during intended use or foreseeable misuse.
How the laws define this
Both standards agree that AI safety centers on system robustness and ensuring that systems operate as intended without causing unintended harm or risks. The Canadian Artificial Intelligence Safety Institute (CAISI) and 人工智能应用伦理安全指南 1.0 both mandate that artificial intelligence function reliably to mitigate potential dangers. They diverge, however, in how they define the scope of harms and operational methods. The Canadian Artificial Intelligence Safety Institute (CAISI) frames safety as research and operational practices that reduce specific harmful outcomes, namely misinformation, privacy leakage, cyber-exploitation, and hazardous autonomous behaviour, through detection, mitigation, alignment, and safe deployment. In contrast, 人工智能应用伦理安全指南 1.0 defines safety through technical system attributes including reliability, security, and controllability, while explicitly extending its scope to guard against risks to national security alongside harms to individuals and society.
Synthesised from the 2 statutory definitions below. Each one is quoted in full, with its source.
How jurisdictions differ
What is distinctive in each definition, binding law in force first.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| China | Ethics-Safety Guidelines for Artificial Intelligence Applications 1.0 | In Force | Adds emphasis on security, controllability, and protection of national security alongside robustness and reliability. |
| Canada | Canadian Artificial Intelligence Safety Institute (CAISI) | In Force | Widens scope to specific harmful outcomes like misinformation, privacy leakage, cyber-exploitation, and hazardous autonomous behaviour, framing research areas such as detection, mitigation, alignment, and safe deployment. |
What it triggers
- China's Ethics-Safety Guidelines for Artificial Intelligence Applications 1.0 sets out ethics and safety guidelines emphasizing human-centered development and risk mitigation across the AI lifecycle.source
- The Canadian Artificial Intelligence Safety Institute (CAISI) describes its work on detection, mitigation, robustness, alignment, and safe deployment.source
Definitions (2)
For CAISI, AI safety refers to research and operational practices that reduce the probability of harmful outcomes from AI systems, including but not limited to misinformation, privacy leakage, cyber‑exploitation, and hazardous autonomous behaviour. It frames the institute's work on detection, mitigation, robustness, alignment, and safe deployment.
Encompasses the robustness, reliability, security, and controllability of AI systems, ensuring they operate as intended without causing unintended harm or risks to individuals, society, or national security.
Not to be confused with
Robustness is specifically the ability of an AI system to maintain performance under failures, attacks, or unexpected conditions, narrower than the broader safety measures in these definitions.
Risk management is the systematic process of managing AI-related risks, while AI safety here focuses on specific practices and attributes to prevent harm.
Trustworthy AI encompasses safety plus fairness, transparency, accountability, and privacy, whereas these definitions center on harm prevention without those additional qualities.
Questions and answers
What does AI safety include under China's guidelines?
It encompasses robustness, reliability, security, and controllability to ensure AI operates as intended without unintended harm or risks to individuals, society, or national security. The guidelines emphasize human-centered development and risk mitigation across the AI lifecycle.
How does CAISI define AI safety?
AI safety refers to research and operational practices reducing harmful outcomes like misinformation, privacy leakage, cyber-exploitation, and hazardous autonomous behaviour. It frames work on detection, mitigation, robustness, alignment, and safe deployment.
What are the key focuses of these AI safety definitions?
China's definition highlights security, controllability, and national security protection. CAISI specifies harmful outcomes and research areas like alignment and safe deployment.
Related Terms
Safety Component
A part of a product or AI system that performs a safety function and whose failure could endanger people or property....
Risk Management
Systematic process of managing AI-related risks....
Robustness
The ability of an AI system to maintain acceptable performance and resist failures, attacks, or unexpected conditions across its lifecycle....
Trustworthy AI
AI that is safe, fair, transparent, accountable, and privacy-preserving....