Capability
How capable frontier AI is: breadth of expertise, how long agents work on their own, how much AI research AI does itself.
Measure
CARI: the Civilizational AI Risk Index.
A single number for where AI risk stands, built from 70 anchored questions in six dimensions. Every answer has written anchors (what a 10, a 50 or a 90 means) and the story it is modelled on, from Neuromancer's Turing Police to The Machine Stops. Scored since 2019; daily updates are launching soon.
CARI44.5/ 100Beyond appetiteas of 1 October 2026
Six dimensions
Every dimension is a risk: higher is worse.
CARI measures the AI systems themselves, not their economic effects. Turning AI risk into what happens to jobs, trust and institutions is MoralWorld's job.
Diffusion
How widely and cheaply capability spreads: users, cost, devices, open weights, autonomous agents.
Accident
How deeply AI sits in critical decision loops (grids, finance, health, defence) and how severe failures have been.
Control
Whether humans can still oversee, verify and correct AI: deception, shutdown resistance, readable reasoning.
Governance gap
How weak the brakes are: binding coordination, independent evaluation, liability, limits on military AI.
Misuse
How much AI empowers harmful use: biological and cyber uplift, fraud, manipulation of shared information.
Reading the number
Five bands.
The bands follow how institutions handle risk: appetite (what was planned for), tolerance, capacity, and the point where correction is no longer possible.
| Level | Band | What it means |
|---|---|---|
| 0-40 | Within appetite | Risks institutions planned for. |
| 40-60 | Beyond appetite | Risk grows faster than safeguards. Where the world is now. |
| 60-80 | Beyond tolerance | Failures and misuse arrive faster than institutions can respond. |
| 80-90 | Beyond capacity | Institutions can no longer absorb the failures. |
| 90-100 | Beyond correction | Humans can no longer correct critical AI systems. |
History
From 9 to 44 in seven years.
Each question carries its answer history since 2019, so the index can be backtested and its movement explained question by question.
Method
Anchored judgement, not a black box.
Anchored answers
Each question is scored against written anchors at 10, 30, 50, 70 and 90, so two readers can disagree about the evidence but not about what a score means.
Critical floors
A dimension cannot look safe while one of its critical questions is near the top: the dimension score is floored at the highest critical answer minus 10.
Daily updates
The index will be recomputed daily as evidence changes. Large moves are reviewed before they are published, so one bad data point cannot move a public index.