Superintelligence is a label before it is a technology

The term “superintelligence” is gaining official use in the United States, but the technology it is supposed to describe does not exist. The White House has directed federal agencies to use “super intelligence” instead of AI in documents and communications, while researchers and industry leaders use the word for a hypothetical system that would exceed human abilities and improve itself. That gap matters: the label is already entering public life, even as there is no shared definition, timeline or clear path from today’s AI to such a system.

A destination without a map

Superintelligence is often framed as a stage beyond artificial general intelligence, or AGI: a system able to work with knowledge across many fields, then extend human capabilities and improve itself without people. A Brookings analysis cited in the article describes the central concern: systems that improve themselves could learn faster than humans, then surpass them through recursive self-improvement.

But there is no agreed definition of the destination, let alone a route to it. Goldman Sachs says at least two major advances beyond current technology would be needed, and considers “superautomation” more likely than superintelligence.

The systems available today can recognize patterns, solve mathematical problems, write code and process large amounts of context. Meta senior engineer Akanksha Sharan says those abilities do not make them superintelligent. Intradiem president and co-CEO Jennifer Lee also doubts that progress will follow a neat sequence from today’s AI to AGI and then superintelligence.

The distinction matters for companies already making plans around AI. The near-term case for changing jobs, processes and skills does not depend on superintelligence arriving. The larger claims about autonomous discovery and self-improvement do.

What the promise contains

Experts cited in the article imagine superintelligence connecting AI agents, finding missing information and combining insights across systems. It could help with research, medicine, city planning, transport and climate science, where analysis requires large datasets and long time horizons.

The possible applications range from better economic forecasting and supply-chain monitoring to drug development, genome mapping and safer transport. The article also describes a more physical prospect: UPS senior engineer Ravindra Bhargava suggests humanoid robots working around the clock in factories could cut production costs and increase output.

Those are possibilities, not demonstrated capabilities. My read is that the forecasts are most useful as a map of what people hope to delegate, not evidence that a system can do it. The more important question is who decides when an AI’s recommendation becomes an action.

That boundary already matters in less speculative settings. SAS vice president of applied AI and modeling Udo Sglavo distinguishes between a system that flags a suspicious transaction and one allowed to block it. A faster decision is not automatically a better one, especially when people may not have time to check the reasoning or assess its consequences.

The control problem comes first

The risks described are not limited to a future machine exceeding human intelligence. AI dependence could weaken human expertise and skills; people may stop checking outputs precisely because a system appears capable. Lee argues that organizations need to decide in advance which decisions can be delegated and where responsibility must remain with people.

There is also a mismatch between the pace of AI improvement and the infrastructure around it. Ravel co-founder and CEO Denise Muyco says models, agents and reasoning mechanisms are advancing daily, while power grids, data centers, security systems, workplace practices and institutions are not built to adapt at the same speed.

Sharan warns that a small error in a large system can cascade. Reliable limits, monitoring and ways to stop or reverse actions are needed before speed is treated as the goal. The article points to trading algorithms as a reminder that automated failures can damage markets in minutes.

The infrastructure and oversight questions are less glamorous than the prospect of machines making discoveries. They are also more immediate. A system that can act faster than people can review it shifts power toward whoever sets its rules—and away from the people expected to live with the results.

A Cambridge University article prepared by 22 leading AI experts, including Geoffrey Hinton and Anton Korinek, warns that years of progress could be compressed into months or less. The authors say AI systems could automate much or all AI research and development within several years, bringing potential benefits alongside extreme risks: capabilities could grow faster than society can adapt, control could be lost, and checks on power inside governments and companies could weaken.

That is the tension hidden by the word itself. Superintelligence may remain hypothetical, but decisions about delegation, oversight and accountability cannot wait for a settled definition.