Technical

Safe AI

AI designed to prevent unintended harm across its lifecycle.

How the laws define this

International guidance consistently establishes that safe AI requires minimizing risks and preventing harms throughout design, development, and deployment using risk assessments and mitigation measures. The primary divergence among these standards concerns the specific categories of harm and protected targets. The Belgrade Ministerial Declaration on Artificial Intelligence focuses on preventing unintended physical, social, or economic harm. In contrast, the Declaration: Towards the Safe, Secure, and Trustworthy Development and Deployment of Artificial Intelligence in the Americas: The Importance of Governance, Regulatory, and Institutional Frameworks explicitly incorporates environmental harms alongside risks to individuals and society. Addressing physical assets, the UN General Assembly Resolution on Artificial Intelligence (A/78/L.49) specifies protecting property and critical infrastructure using fail-safes and robustness measures. Finally, the International Strategy for Safe and Responsible AI Transition sets itself apart by addressing deliberate security threats, emphasizing resilience against malicious attacks and prevention of system misuse.

Synthesised from the 4 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. Showing 3 of the 4 instruments that define this term; all are listed under Definitions below.

Jurisdiction	Instrument	Status	What is distinctive
OAS	Declaration: Towards the Safe, Secure, and Trustworthy Development and Deployment of Artificial Intelligence in the Americas: The Importance of Governance, Regulatory, and Institutional Frameworks	In Force	Adds explicit mechanisms to identify, understand, and mitigate risks and harms to individuals, society, and the environment.
Netherlands	International Strategy for Safe and Responsible AI Transition	In Force	Adds minimization of risks to security and potential for misuse, plus emphasis on system robustness, reliability, and resilience against malicious attacks.
Serbia	Belgrade Ministerial Declaration on Artificial Intelligence	Adopted	Adds emphasis on lifecycle risk assessment, mitigation measures, and ongoing monitoring across design, deployment, operation, and retirement to prevent unintended physical, social, or economic harm.

Definitions (4)

Systems engineered and managed to minimize and prevent unintended physical, social, or economic harm throughout design, deployment, operation, and retirement. Emphasis is placed on lifecycle risk assessment, mitigation measures, and ongoing monitoring to ensure continued safety.

Belgrade Ministerial Declaration on Artificial IntelligenceDefinition 1 of 4

AI systems designed and operated to minimize risks to people, property, and critical infrastructure throughout their lifecycle, incorporating risk assessments, robustness measures, fail-safes, and mitigation strategies to prevent harm.

UN General Assembly Resolution on Artificial Intelligence (A/78/L.49)Definition 2 of 4

AI systems developed and deployed with mechanisms to identify, understand, and mitigate potential risks and harms to individuals, society, and the environment.

Declaration: Towards the Safe, Secure, and Trustworthy Development and Deployment of Artificial Intelligence in the Americas: The Importance of Governance, Regulatory, and Institutional FrameworksDefinition 3 of 4

'Safe AI' implies the development and deployment of AI systems that minimize unintended harm, risks to security, and potential for misuse. This includes addressing issues such as system robustness, reliability, and resilience against malicious attacks.

International Strategy for Safe and Responsible AI TransitionDefinition 4 of 4

Not to be confused with

Safety (AI)

Safety focuses on AI operating without causing unacceptable harm, while safe AI emphasizes development and deployment processes to prevent such harm.

Reliable AI (Trustworthy AI)

Reliable or trustworthy AI meets broader ethical, legal, and technical requirements, whereas safe AI specifically targets minimizing unintended harm.

Trusted/Trustworthy AI

Trusted or trustworthy AI includes respect for human rights, fairness, transparency, and safety, but safe AI narrows to harm prevention.

Questions and answers

What makes an AI system safe under the OAS Declaration?

Under the OAS Declaration, safe AI systems are developed and deployed with mechanisms to identify, understand, and mitigate potential risks and harms to individuals, society, and the environment. It sets a regional framework balancing innovation with human rights.

How does the Netherlands define safe AI in its strategy?

The Netherlands' strategy defines safe AI as systems that minimize unintended harm, security risks, and misuse potential, addressing robustness, reliability, and resilience against attacks. It outlines commitments to global collaboration for responsible AI.

What is distinctive about safe AI in the Belgrade Declaration?

The Belgrade Declaration describes safe AI as systems engineered to minimize unintended physical, social, or economic harm through lifecycle risk assessment, mitigation, and monitoring. It guides GPAI members and the EU toward human-centric AI.