AI Safety
Automated red-teaming, containment protocols, refusal-rate benchmarks and agentic fail-safes.
Theory
Four pillars, three anchors
The pillars turn "responsible" into thresholds you can test. The anchors are the frameworks those thresholds are checked against.
Responsibility & Governance
Legal liability architecture, third-party auditing and incident reporting loops.
Impact Alignment
Goal-drift prevention and mechanistic interpretability, so that what a model does matches what it was meant to do.
Fairness & Bias
Data transparency through datasheets, and equity standards that report results group by group.
EU
EU AI Act
Prohibited, high-risk, limited-risk and systemic-risk tiers, each with its own obligations.
Open NISTAI RMF 1.0 and AI 600-1
Govern, Map, Measure and Manage, with a profile for generative AI.
Open ISOISO/IEC 42001
A management system for AI that an organisation can be certified against.
OpenThe patchwork shield
How much of the shield is actually stitched?
We picture global AI governance as a shield made of patches. Where it holds, where it frays and where it is open depends on what can be enforced.
Binding and enforceable
Statutory regulation with penalties behind it, such as high-risk enforcement under the EU AI Act.
Promised, not required
Soft-law commitments, voluntary pledges and self-reported responsible scaling policies.
Nobody is watching
Unmonitored open-weights distribution, unregulated agentic execution and regulatory arbitrage across borders.