Introduction
Discussion of AI governance has shifted from “whether to regulate” to “how to regulate”. Publicly visible principle documents are now dense, with words such as safety, transparency, accountability and fairness recurring. Yet conflicts between principles, interpretive discretion and the absence of enforcement tools leave many commitments on paper. The “implementation problem” is becoming the core issue of AI governance.
Why principles do not take effect automatically
The common weakness of principle documents is abstraction. Take “transparency”: does it mean disclosing training data, model architecture, or the basis of decisions? The costs and boundaries differ entirely. Take “accountability”: when a system is deployed by multiple actors, does responsibility lie with developers, deployers or users? Lists of principles cannot answer these questions automatically — they need institutionalised interpretation and discretion.
Three levers from principles to mechanisms
- Specification: translate abstract principles into operational requirements by domain and risk level, such as ex ante assessment and record-keeping duties for high-risk applications.
- Tooling: develop enforceable technical and standards tools — model cards, audit logs, explainability methods — so compliance is not an empty promise.
- Institutionalisation: clarify which level of authority interprets and enforces, and establish cross-domain coordination instead of fragmented mandates.
A realistic path for cross-national coordination
Because AI research and deployment are highly globalised, any single country’s rules face spillover and circumvention. A pragmatic path is not one global law but gradual convergence through mutual recognition of standards, sharing of assessment results and multilateral initiatives. Regional pilot experiments can provide experience samples for global rules.
Conclusion
The maturity of governance is measured not by the number of principle documents but by the reliability of enforcement mechanisms. At this stage, the value of research institutions is to track the real distance from “declaration” to “implementation” and provide comparable evidence, rather than repeating the principles once more.
Sources and version: this is an editorial opinion piece; version v1.0 (2026-09).