Specify the positive destination
Preventing Artificial superintelligence: capability substantially beyond humans across a broad range of important tasks. would still leave an enormous range of possible technological progress. People could continue developing medicines, energy systems, transportation, accessibility tools, scientific instruments, and ordinary software. The policy question is which forms of AI could support that work while remaining meaningfully controllable.
W360's proposal should preserve useful technology below an enforceable A proposed continuing limit on what systems or combinations of systems are permitted to do.. That requires more than calling a tool narrow or beneficial. Its real capabilities, operational permissions, dependencies, and potential combinations must be examined. A medical label, for example, does not by itself determine whether an underlying model has dangerous general abilities.
Judge systems by what they can do
Candidates for continued use could include statistical forecasting, bounded translation, assistive interfaces, diagnostic support, industrial optimization, and scientific analysis. Their acceptability would depend on the implemented system and the chosen ceiling, rather than on a permanent exemption for an entire sector.
Useful questions include: Can the task be specified clearly? Can the output be checked by an appropriately qualified person or independent process? What actions can the system take? Can it acquire additional resources, change its own permissions, or train successors? Can operators recover when it fails?
Design for limited authority and recovery
A tool could be designed with restricted access, bounded execution, human authorization for consequential actions, and independently maintained alternatives. Those measures would need evaluation in the actual setting. A nominal approval button is weak protection if the reviewer cannot understand the recommendation or has no practical alternative.
Preserving skills, documentation, and replacement systems matters alongside limiting software permissions. A service is easier to govern when people can operate it without depending on a single inscrutable provider. That may mean accepting some loss of convenience or efficiency in exchange for genuine recoverability.
Acknowledge the tradeoff
Restricting general capabilities could delay discoveries or make some services more expensive. The prevention case should state those costs openly and invest in alternative ways to obtain the benefits. It cannot promise that every benefit of unrestricted development will survive unchanged.
The intended destination is a technologically capable society whose members retain meaningful authority over its direction. Which tools fit that destination is an ongoing technical and democratic question, subject to evidence and public review.