Seoul Statement unites IEC, ISO and ITU to advance inclusive, safe and sustainable AI through international standards

Seoul Statement: IEC, ISO and ITU set a common path for sustainable, safe AI

December 2, 2025 - The International Electrotechnical Commission (IEC), International Organization for Standardization (ISO) and International Telecommunication Union (ITU) issued a joint "Seoul Statement" committing to push international standards that make AI inclusive, open, sustainable, fair, safe and secure.

For engineering and product teams, this is a clear signal: AI features will be expected to meet shared, testable norms across safety, interoperability and environmental impact. Standards are moving from optional to assumed.

Why this matters for IT and development teams

• Interoperability by default: Data schemas, APIs and model documentation will need to line up across vendors to reduce integration friction.
• Safety you can prove: Expect conformance tests for model behavior, red-team results and incident handling to become standard release artifacts.
• Lifecycle governance: From problem framing to decommissioning, teams will need traceability for data, models, prompts and decisions.
• Inclusion and fairness: Bias assessments and accessibility checks won't be "nice to have" anymore-they'll be gated checks.
• Sustainability metrics: Energy usage and emissions reporting for training and inference will be requested by buyers and regulators.

Where standards are likely headed

• AI management systems: Organization-wide controls, roles and audits for AI projects (policy, risk, monitoring, improvement).
• Risk management: Consistent methods to identify, analyze and treat AI risks, including harm scenarios and mitigations.
• System lifecycle: Requirements for data lineage, model versioning, deployment approvals, rollback plans and retirement.
• Transparency and evaluation: Model cards, dataset statements, evaluation suites, content provenance and watermarking guidance.
• Security: Threat modeling for AI-specific risks (prompt injection, data poisoning, model theft) and continuous monitoring.
• Interoperability: Common vocabularies and exchange formats for datasets, metrics and model outputs.
• Environmental reporting: Standardized methods to calculate and disclose energy use and emissions.

90-day action plan

• Inventory: List every AI feature, model and third-party service. Note purpose, data sources, model versions and owners.
• Risk register: Capture foreseeable harms, likelihood, severity and mitigations. Assign owners and review cadence.
• Documentation: Add MODEL_CARD.md and DATASET_CARD.md to repos with inputs, metrics, limits and safe-use notes.
• Eval pipeline: Automate behavioral tests (quality, safety, bias, prompt injection). Gate releases on thresholds.
• Security controls: Add input validation, output filtering, secrets isolation, rate limits and abuse monitoring.
• Incident response: Define triggers, on-call paths and rollback steps for AI-specific failures.
• Energy tracking: Log compute hours, hardware type and power source during training and inference.
• Vendor clauses: Require suppliers to disclose model lineage, datasets (where possible), eval results and energy metrics.

Developer checklist for each AI feature

• Clear problem statement, intended users and misuse cases.
• PII handling: minimization, masking, retention rules and access logs.
• Pre-deployment evals: quality metrics, out-of-distribution behavior and red-team findings.
• Bias testing across relevant cohorts with thresholds and fixes.
• Fail-safes: fallback responses, human escalation and rate-limit behavior.
• Observability: prompt/model/version IDs, user/session IDs and key metrics (quality, safety, latency, cost).
• Change control: pull request templates capturing data/model changes and rollback plans.

Procurement and vendor management

• Request evaluation reports, safety test suites and monitoring dashboards.
• Require security whitepapers for model and data protection.
• Add sustainability disclosures to RFPs: training/inference energy and emissions.
• Ensure rights to perform independent testing and audits.

Open-source projects

• Add SECURITY.md for AI threats and disclosure process.
• Ship example evals and a minimal safety harness in CI.
• Provide MODEL_CARD.md and DATASET_CARD.md with known limits.
• Tag releases with training data windows and notable behavior changes.

What to watch next

• New work items and calls for contribution from IEC, ISO and ITU.
• Conformance and certification programs built on shared test suites.
• Crosswalks between international standards and regional rules to simplify compliance.

Useful references

• ISO/IEC JTC 1/SC 42 - Artificial Intelligence
• ITU - AI standardization initiatives

Level up your team

If you're standing up governance, evaluation and MLOps tooling, structured training helps teams move fast without creating gaps. Browse hands-on tracks by role at Complete AI Training - courses by job.