Superintelligence is a claim about breadth. A calculator outperforms every human at arithmetic without being superintelligent, because the advantage is confined to one task. Nick Bostrom’s original framing, and the one most AI-safety researchers still use, requires the system to exceed top human performance across essentially all cognitively demanding domains at once, science, strategy, persuasion, and everything else, rather than in a single narrow specialty.
No system meeting that definition exists yet, which is what makes the governance debate around it unusual. Every other AI risk on a policy agenda concerns behaviour someone has already observed in a deployed system. Superintelligence policy is being written against a projected capability curve, not an incident report, which is exactly why serious researchers who agree on almost nothing else, from those who think it is decades away to those who think it is imminent, still converge on the same practical point: nobody has demonstrated a way to reliably control a system smarter than the people supervising it.
Treat “superintelligence” as a capability claim to be checked against actual benchmarks and control-problem research, not against a release-date prediction. Nobody, including the labs closest to the frontier, has a reliable method for calling that date, and the governance question does not wait for one.