What Superintelligence Means

Narrow AI, general AI and superintelligence

Stanford's Institute for Human-Centered AI describes narrow AI as intelligent systems for particular tasks, such as speech or facial recognition. Artificial general intelligence, or AGI, is the aim of broadly intelligent, context-aware machines, which the institute also calls human-level AI. Superintelligence goes further still. Philosopher Nick Bostrom's widely quoted definition is an intellect smarter than the best human brains in practically every field. No robot has shown general, human-level ability, and superintelligent systems remain hypothetical, so the three terms describe a ladder of ambition rather than three kinds of product.

Where the idea came from

In 1965 the statistician I. J. Good described an ultraintelligent machine as one that can far surpass all the intellectual activities of any person, however clever. Because designing machines is itself an intellectual activity, he argued, such a machine could design even better machines, setting off an intelligence explosion. His often-quoted line that the first such machine would be the last invention people need ever make came with a condition: provided the machine is docile enough to tell us how to keep it under control. Later writers called the moment when AI builds AI beyond human level the singularity.

Intelligent is not the same as capable

Stanford HAI makes a useful distinction with two robots. A fully pre-programmed factory robot is flexible, accurate and consistent, but not intelligent. A hospital delivery robot that must find its own way along busy corridors needs genuine autonomy to succeed. For robots, intelligence shows up as coping with change. A far smarter planner would still depend on its motors, sensors, batteries and hands, so any superintelligent robot would be limited by its body as well as helped by its mind.

Why the word needs care

Researchers disagree about whether superintelligence is possible, and about what it would mean. A peer-reviewed 2021 paper in the Journal of Artificial Intelligence Research argued that total containment of a superintelligence is impossible in principle, because deciding whether such a machine will cause harm cannot be computed. The same authors note that it poses a different problem from everyday robot ethics. For anyone assessing robots, treat a product described as superintelligent as a claim to question, and ask what it does on clearly defined tasks.

Sources and further reading

• Brief Definitions of Key Terms in AI — Stanford Institute for Human-Centered Artificial Intelligence (HAI).
  Stanford HAI's short glossary separating narrow AI from AGI, with robot examples of what intelligent really means.
• Ethics of Artificial Intelligence — Vincent C. Müller, Eindhoven University of Technology (arXiv).
  Handbook chapter that quotes I. J. Good's 1965 intelligence explosion passage in full and explains the singularity debate.
• Superintelligence Cannot be Contained: Lessons from Computability Theory — Journal of Artificial Intelligence Research (arXiv).
  Peer-reviewed paper defining superintelligence and arguing that complete containment is impossible in principle.

Common questions

Is any robot superintelligent today?

No. Superintelligence is a hypothetical level of ability that no system has demonstrated, and no robot has yet shown general, human-level ability.

Is AGI the same as superintelligence?

No. AGI usually means broadly human-level ability across many tasks, while superintelligence means far surpassing the best people in practically every field. AGI is often seen as a possible step towards it.