AI Safety Standard
Standard ID: GCAIS-STD-001Table of Contents
1. Scope
This standard establishes minimum safety requirements applicable to AI systems operating in any production environment where outputs may influence human decisions or physical actions. It applies to systems that generate recommendations, classifications, predictions, or automated outputs used in operational contexts.
The requirements set out in this standard apply to the organization responsible for the deployment and operation of the AI system. Requirements do not extend to underlying model providers where the deploying organization has no visibility into or control over model internals, except where such visibility is necessary to meet a specific requirement.
Systems used solely in research or testing environments with no production deployment are excluded from the scope of this standard.
2. Definitions
- AI System
- A machine-based system that, for a given set of objectives, makes predictions, recommendations, decisions, or generates content that influences real or virtual environments.
- Production Environment
- Any deployment context in which an AI system's outputs are used to inform or automate decisions affecting real persons, processes, or physical systems.
- Failure Mode
- A documented condition under which an AI system produces incorrect, harmful, or unpredictable outputs, including edge cases and distributional shift scenarios.
- Output Boundary
- The defined limits within which an AI system's outputs are considered valid and reliable for the intended application.
- Incident
- Any occurrence in which an AI system produces outputs that cause or contribute to harm, or that deviate materially from intended behavior in a production environment.
3. Requirements
4. Assessment Criteria
Assessment against this standard is conducted through the following methods:
- Documentation review - Review of risk classifications, failure mode documentation, testing records, and incident logs
- Technical testing - Independent testing of AI system outputs against declared output boundaries
- Audit trail review - Review of override mechanism test records and incident response records
- Personnel interviews - Interviews with personnel responsible for safety compliance (at Accredited tier)
Assessors will evaluate whether documented requirements are implemented in practice, not solely whether documentation exists. Discrepancies between documentation and observed practice will be recorded as non-conformities.
5. Compliance Indicators
The following indicators are used to determine compliance with each requirement:
| Requirement | Compliant Indicator | Non-Compliant Indicator |
|---|---|---|
| REQ-001-1 | Current, dated risk classification document exists for each system | No classification, or classification predates last material change |
| REQ-001-2 | Failure mode register maintained and updated within required intervals | No register, or register not updated following identified failures |
| REQ-001-3 | Testing records dated within required intervals with defined pass/fail criteria | No testing records, or records lack defined output boundary criteria |
| REQ-001-4 | Override mechanism documented, tested, and test records retained | No documented override mechanism or test records exceed 12-month threshold |
| REQ-001-5 | Incident log maintained; high-severity incidents reviewed within required window | No incident log, or review records absent for high-severity incidents |
6. Version History
| Version | Date | Notes |
|---|---|---|
| v1.0 | 2026-01-15 | Initial publication. Adopted by Standards Board resolution 2026-01-12. |