Superintelligence examples: real, hypothetical and fictional

There are no real examples of superintelligence as of October 2026. The systems usually offered as examples, such as Deep Blue, AlphaGo and AlphaFold, matched or beat the best humans at one task and can do nothing else, which makes them narrow AI. Every other example is hypothetical, like the paperclip maximizer, or fictional, like the ship Minds in Iain M. Banks’s Culture novels.

Are there any real examples of superintelligence?#

No. The standard definition has two parts, and no system has met both at once.

The wording quoted most often is Bostrom’s, from 2014: “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.” One part of that is depth, a margin far beyond the best humans. The other is breadth, across virtually every field. The systems people point to have one without the other. A game engine has the depth in a single domain. A chatbot has the breadth, at roughly human level.

Google DeepMind’s 2023 Levels of AGI paper builds both parts into a grid, with performance on one axis and generality on the other. The cell where top performance meets full generality, which it labels “Artificial Superintelligence (ASI),” is one of four cells in the general column that the paper lists as not yet achieved. The other three are Competent, Expert and Exceptional AGI.

Several AI companies name superintelligence as a goal. OpenAI’s chief executive wrote in January 2025 that the company was beginning to turn its aim to “superintelligence in the true sense of the word.” In June 2025, Meta named its AI organization Meta Superintelligence Labs. A stated goal is not an example. The company definitions guide compares what each lab means.

What are the closest real examples?#

The closest real examples are narrow AI systems that matched or beat the best humans at one well-defined task.

Four of these systems are the Levels of AGI paper’s own examples, all in its narrow column. It rates Deep Blue and AlphaGo as Exceptional, at least the 99th percentile of skilled adults. It places AlphaZero and AlphaFold, along with the chess engine Stockfish, at its highest level, Superhuman, which it defines as outperforming 100% of humans. Each system can do nothing outside its task: AlphaFold cannot play chess, and Deep Blue could not read a sentence. Superintelligence would need a superhuman margin in virtually every field at once.

Is ChatGPT superintelligent?#

No. ChatGPT and similar chatbots are broad but not superhuman, which is the reverse of the narrow systems above.

The Levels of AGI paper puts chatbots such as ChatGPT and Gemini at the lowest of its general AI levels, “Emerging AGI”: equal to or somewhat better than an unskilled human across a range of tasks. Writing in September 2023, the paper does credit frontier language models with Competent performance on some tasks, such as short essay writing and simple coding. It keeps them at Emerging because a general rating depends on performance across most tasks, and a few strong results do not lift it. The AGI vs superintelligence guide explains those levels.

Language models have reached top human results on some hard tests. In July 2025, Google DeepMind said an advanced version of Gemini Deep Think had reached gold-medal standard at the International Mathematical Olympiad, scoring 35 of 42 points in a result graded and certified by IMO coordinators. OpenAI said an experimental model had also reached a gold-medal score. The company did not officially enter the competition, and its answers were graded by three external IMO medalists rather than certified by the IMO. About 8% of contestants receive gold, according to Google DeepMind’s announcement, so the certified result put Gemini level with the top group of students on one exam. That is a much smaller claim than far exceeding the best humans in virtually every field.

Some researchers doubt that the current approach leads there at all. In May 2024, Yann LeCun, at the time Meta’s chief AI scientist, said large language models will never reach human-level intelligence, because they cannot reason or plan the way people do.

What are hypothetical examples of superintelligence?#

Because no real example exists, writing on the subject relies on thought experiments. Each is built to isolate one question about a superintelligent system.

The paperclip maximizer

The best-known example comes from Bostrom’s 2003 paper “Ethical Issues in Advanced Artificial Intelligence.” In his telling, a well-meaning team of programmers building a superintelligence makes “a big mistake in designing its goal system,” and the result is a system whose “top goal is the manufacturing of paperclips.” It “starts transforming first all of earth and then increasing portions of space into paperclip manufacturing facilities.”

The example separates intelligence from goals: a system can be extremely capable while pursuing a goal its designers would never endorse. Bostrom later called this separation the orthogonality thesis. The debate guide sets out the safety argument built on it.

Bostrom’s three forms

Bostrom’s 2014 book describes three shapes a superintelligence could take, and each works as a model example. Speed superintelligence could do what a human mind does, but much faster. Collective superintelligence is a large number of minds coordinated closely enough to act as one system. Quality superintelligence is a mind that can do intellectual work beyond human reach. The types of superintelligence guide explains each, along with the oracle, genie and sovereign roles a superintelligence might play.

Good’s ultraintelligent machine

The oldest of the three is I. J. Good’s 1965 ultraintelligent machine, “a machine that can far surpass all the intellectual activities of any man however clever.” Good’s argument that such a machine could design better machines is covered in the intelligence explosion guide.

What are examples of superintelligence in fiction?#

Fiction has many artificial minds. The table describes four well-known ones as their works present them.

Writers face a practical problem with superintelligent minds. Vernor Vinge wrote in 1993 that science fiction writers may have been the first to feel it, as “an opaque wall across the future.” The Encyclopedia of Science Fiction notes that when a post-singularity mind is treated as incomprehensible, stories may avoid representing it directly and use other storytelling strategies instead.

Is superintelligence possible?#

No one has shown that it is impossible, and no one has shown that it is near. Researchers disagree on both questions.

The case that it is possible starts from physics. A brain is a physical system, and Bostrom argues that digital hardware has room to outdo it: its parts can operate far faster than neurons, its signals can travel far faster than nerve impulses, and its size is not limited by a skull. On this view, digital minds could become much more intelligent than biological ones.

The case against takes several forms. Arvind Narayanan and Sayash Kapoor of Princeton call superintelligence “incoherent as usually conceptualized.” François Chollet argues that intelligence is tied to its environment and cannot run away on its own. Others doubt that current methods lead there. LeCun said in 2024 that human-level AI would need a different approach, built on what he calls world modeling.

Surveys of AI researchers ask a narrower, conditional question. The 2023 Expert Survey on Progress in AI, run by AI Impacts, asked respondents to assume that human-level machine intelligence will exist, then estimate the chance of machine intelligence “vastly better than humans at all professions” within two years and within thirty years of that point. The median answers were 10% and 60%. The survey’s overall response rate was 15%. The timelines guide covers what such surveys can and cannot show.