Compliance

Evaluation

Systematic assessment of AI products against set criteria.

How the laws define this

Frameworks uniformly understand evaluation as a systematic or methodical process of testing and examining AI systems to measure operational robustness, capabilities, and safety dimensions. They differ significantly, however, in their core focus, operational setting, and intended output. The AI Product and System Evaluation Centre (AIEC) focuses on testing products and systems against established criteria across specific dimensions like fairness and privacy to determine formal compliance and risk classification. In contrast, the AI Safety Institute (establishment following AI Safety Summit) centers evaluation on empirical benchmarks, stress tests, and taxonomies to analyze capabilities and failure modes in order to inform safety mitigations and policy. Differing from both, AI Agents in Action: Foundations for Evaluation and Governance restricts its scope specifically to AI agents, examining their performance, behavior over time, reliability, and user trust within dynamic, real-world operational environments.

Synthesised from the 3 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
United Kingdom	AI Safety Institute (establishment following AI Safety Summit)	In Force	Adds focus on empirical tests, analyses of failure modes and harms, capability- and harm-focused taxonomies, benchmarks, stress tests, and methodological outputs for safety mitigations and policy.
Taiwan	AI Product and System Evaluation Centre (AIEC)	In Force	Narrows to testing and measuring AI products and systems against established criteria across dimensions like safety, fairness, privacy, and robustness for compliance and risk classification.
World Economic Forum	AI Agents in Action: Foundations for Evaluation and Governance	In Force	Widens to examining AI agents specifically, emphasizing performance, behavior over time, reliability, user trust, and robustness in dynamic, real-world environments.

What it triggers

The UK AI Safety Institute conducts pre- and post-deployment evaluations using methodical empirical tests and analyses.source
Taiwan's AI Product and System Evaluation Centre evaluates AI products and systems deployed across public and private sectors against established criteria.source
The World Economic Forum framework sets out a structured foundation for organizations to evaluate, manage, and govern AI agents responsibly.source

Definitions (3)

The systematic assessment process by which AI products and systems are tested and measured against the AIEC's established criteria and guidelines across multiple dimensions (e.g., safety, fairness, privacy, robustness) to determine compliance and risk classification.

AI Product and System Evaluation Centre (AIEC)Definition 1 of 3

The process of examining an AI agent's performance, behavior over time, reliability, user trust, and operational robustness in dynamic, real-world environments.

AI Agents in Action: Foundations for Evaluation and GovernanceDefinition 2 of 3

Methodical empirical tests and analyses designed to assess AI system capabilities, failure modes, and potential harms across capability- and harm-focused taxonomies; this includes benchmarks, stress tests, and published methodological outputs used to inform safety mitigations and policy.

AI Safety Institute (establishment following AI Safety Summit)Definition 3 of 3

Not to be confused with

AI Evaluation

AI Evaluation is a broader term for assessing AI products for safety, fairness, and compliance, while these definitions specify evaluation processes or frameworks.

Risk assessment

Risk assessment identifies and mitigates risks, whereas these definitions center on systematic assessment or testing of AI capabilities, performance, or compliance.

AI Impact Assessment

AI Impact Assessment is a pre-deployment documented analysis of risks and harms, unlike these which involve testing, empirical analysis, or ongoing performance examination.

Questions and answers

What does the UK AI Safety Institute evaluate in AI systems?

It uses methodical empirical tests and analyses to assess capabilities, failure modes, and potential harms across capability- and harm-focused taxonomies, including benchmarks and stress tests.

How does Taiwan's AIEC classify AI products?

It systematically assesses them against criteria for safety, fairness, privacy, and robustness to determine compliance and risk classification.

What is distinctive about WEF's AI agent evaluation?

It examines performance, behavior over time, reliability, user trust, and operational robustness in dynamic, real-world environments to support governance.