Dimensions
Three dimensions, one question each
Every claim a company makes about AI falls under one of three questions: can it be stopped, who answers for it, and what does it do to people.
Dimension oneSafe by design
Pillar: AI SafetyCan it be stopped? Whether the people building a system can show that it holds up when someone pushes on it.
- Automated red-teaming with documented runs
- Containment for agentic tools: sandboxes, circuit-breakers
- Refusal-rate benchmarks across hazardous domains
- Human sign-off before state-changing actions
Dimension twoAccountable in practice
Pillar: Responsibility & GovernanceWho answers for it? Whether a named person is on the hook, and whether anyone outside the company has looked.
- Named executive oversight and officer-level sign-off
- Third-party audits before deployment
- Incident reporting with notification deadlines
- Red-team reports published by accredited evaluators
Dimension threeDownstream impact
Pillars: Impact Alignment · Fairness & BiasWhat does it do to people? What happens to communities, economies and ecosystems once the system is out there. Measured across the eight NEESSHEGI factors.
- Goal-drift prevention and interpretability
- Datasheets for datasets: where the data came from
- Disaggregated evaluation across marginal groups
- Impact accounted for across eight factors
Downstream impact is measured with NEESSHEGI: Natural Environment, Energy, Socio-culture, Socio-economics, Health care, Education, Government and Infrastructure. Each factor has its own research site.
The dimension names are working titles.