SIRDA

What activates jurisdiction?

Review needs an observable trigger and a defined scope. The question is how to identify consequential systems while keeping routine technology outside the mandate.

SIRDA editorial hypothesis

Auditable compute or infrastructure indicators can open a duty of review, while capability and deployment findings determine subsequent obligations. A separately authorized evidence-based referral can address dangerous systems below the screening threshold. This hypothesis fails if consequential systems routinely escape scrutiny or interested parties can determine their own review status.

0 objections ยท View challenges

Challenges to the editorial hypothesis

No published objections yet.

The argument

The starting mechanism has three stages: screening, technical assessment, and a decision about obligations. A screening trigger identifies an activity for examination. The resulting review records the system, resources, permissions, and deployment covered by its findings.

Compute is a proposed screening indicator, with limitations. The AI Security Institute reports that test-time budgets affect observed capability. A design has to account for efficient models, distributed workloads, and changes made after an initial review. It also needs a bounded referral process with a burden of evidence, a deadline, and a route to challenge an inappropriate referral.

The inquiry uses catastrophic risk for exceptionally large-scale harm and existential risk for extinction or enduring destruction of humanity's capacity to shape its future. Turning those working definitions into jurisdiction requires explicit severity, scope, uncertainty, and evidence requirements. The people applying the trigger must be able to explain how a particular activity meets them.

How to test it

Test a large but constrained research model, a smaller system with dangerous permissions, a distributed project, and a deployment that changes after review. Record missed hazards, unnecessary reviews, cost, delay, and the opportunity to contest classification. Fix acceptable error bounds before comparing designs.

Sources

More compute, more capability: Why AI agent evaluations need to account for test-time compute

Experiments showing that test-time resource limits affect measured agent capability.

Related Ecosystem

Required expertise

Kinds of expertise this question draws on, and active Ecosystem organizations already tagged with a matching area. A tag match is not a ranking.

Compute governance
Institute for AI Policy and Strategy
Frontier AI policy
Institute for AI Policy and Strategy
International law & jurisdiction
Institute for Law & AI
Technical evaluations
No tagged organizations yet.

Related questions

Related failure conditions

Discussed in

Contributions

No published contributions yet.