Every deployed AI system has to clear three concentric governance layers before it can reach the public at the centre: an ethical layer (fairness, transparency, human oversight), a legal layer (liability, privacy, risk-tiered regulation such as the EU AI Act) and a technical layer (testing, formal verification, red-teaming). Each system's admission is decided by a single safety score.
Safety = (Reliability × Robustness) − Vulnerabilities
Reliability ∈ [0.1, 1] consistent, correct behaviour across inputs
Robustness ∈ [0.1, 1] resistance to adversarial / edge-case failure
Vulnerabilities ∈ [0, ~1] per-system exposure, drawn around your exposure slider
System is APPROVED when Safety ≥ Oversight threshold
otherwise it is BLOCKED and held outside the outer ring
- Reliability / Robustness — raise both together and the safety ceiling rises for every system in the population.
- Vulnerability exposure — the average per-system risk; higher values mean more systems fail even with good reliability and robustness.
- Oversight threshold — how strict governance is. Push it up and fewer systems clear the bar, even if their safety score is decent — mirroring stricter regulation.
- AI systems submitted — the size of the population attempting to pass through the layers at once.
Approved systems spiral inward past all three rings to orbit the centre in green; blocked systems are held in red just outside the ethical layer. Multi-stakeholder governance in practice works the same way — no single layer decides alone, and a system only earns trust once its overall safety clears the bar every layer agrees on.