What could superintelligence actually do?

How to read this page

Superintelligence, in the sense this site uses, means a system that greatly exceeds the best humans in virtually all domains (explained here). Asking what it could do is asking about something no one has seen. Two shortcuts are tempting and both mislead: assuming that a superintelligence would simply be a faster version of today’s chatbots, and assuming that it could do anything at all. The sensible approach is to start from what current systems have demonstrated, then describe the extrapolation, and say how strong the argument for it is.

One general caution comes from the International AI Safety Report 2026, the nearest thing to a consensus scientific account: today’s leading systems may excel at some difficult tasks while failing at simpler ones. Capability in one domain does not imply capability in the next.

Domain by domain

Science and medicine

Demonstrated. Google DeepMind’s AlphaFold, described in Nature in July 2021, predicts the three-dimensional structure of proteins with accuracy competitive with laboratory methods in most cases, a problem that had resisted decades of work. Its creators Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry with David Baker. In February 2025 Google described an “AI co-scientist” system that proposed hypotheses later supported in the laboratory, including drug-repurposing candidates for acute myeloid leukaemia and treatment targets for liver fibrosis, and independently reached an explanation of a bacterial mechanism that matched unpublished results from Imperial College London.

Extrapolated. A superintelligence could, on this view, run the whole cycle of research, from hypothesis to experimental design to analysis, at far greater speed and scale. Dario Amodei, Anthropic’s chief executive, argued in a 2024 essay that sufficiently powerful AI could compress fifty to a hundred years of biological progress into five to ten. That is a company leader’s aspiration, not a forecast with evidence behind its numbers. The limit most often cited is that biology and medicine run on experiments and clinical trials that take as long as they take; more intelligence can design better trials, but cannot make a drug’s long-term effects appear sooner.

Mathematics

Demonstrated. In July 2024 Google DeepMind’s AlphaProof and AlphaGeometry systems scored at silver-medal level on International Mathematical Olympiad problems. A year later, in July 2025, a version of Gemini scored 35 out of 42, gold-medal standard, under conditions certified by the Olympiad; OpenAI reported the same score from an experimental model graded by former medallists. In May 2025 Google DeepMind’s AlphaEvolve found a way to multiply four-by-four complex-valued matrices using 48 multiplications, improving on a method that had stood since 1969.

Extrapolated. Olympiad problems are hard but have known answers. Research mathematics means finding new questions and long chains of new proof. A superintelligence would, on the extrapolated view, produce research mathematics faster and more reliably than any human, including results humans could check but not have found. Progress on research-level problems has been slower and more contested than on competition problems.

Software and cyber security

Demonstrated. This is where current capability is clearest, though the most striking figures are company claims that have not been independently audited. Anthropic reports that, by May 2026, its models wrote more than 80% of the code merged into its own codebase. In April 2026 it said its Claude Mythos Preview model had found thousands of previously unknown vulnerabilities, including in every major operating system and web browser. The International AI Safety Report 2026 records that in one competition an AI agent identified 77% of the vulnerabilities present in real software.

Extrapolated. A superintelligence could write and audit software at a scale no human team could check. The same capability is two-sided: finding vulnerabilities helps defenders patch them and attackers exploit them. Our risks page covers the cyber concern.

Materials and engineering

Demonstrated, with a caveat. In 2023 Google DeepMind reported in Nature that its GNoME system had predicted 2.2 million new crystal structures, about 380,000 of them among the most stable. A 2024 review in Chemistry of Materials by Anthony Cheetham and Ram Seshadri found “scant evidence” among a sample of the results for compounds that were new, credible and useful at once. The episode shows the gap between generating candidates and making discoveries.

Extrapolated. A superintelligence could design materials, catalysts, batteries and machines with properties specified in advance, and could plan how to make them. Bostrom’s 2014 book lists “technology research” among the abilities in which a superintelligence would have a decisive advantage. Engineering is ultimately limited by physical testing and manufacturing, which no amount of thought removes.

Robotics and the physical world

Demonstrated. Less than in purely digital domains. The International AI Safety Report 2026 notes that progress on digital tasks has proved difficult to translate into robotics. This reflects an old observation, often called Moravec’s paradox, that what people find hard (chess, algebra) has proved easier for machines than what people find easy (walking, handling objects).

Extrapolated. A superintelligence would not need a body to act on the world; it could act through people, companies and existing machines, and it could design robots. But how quickly it could change the physical world would depend on manufacturing, energy and logistics, not only on intelligence.

Persuasion

Demonstrated. In a study published in Nature Human Behaviour in 2025, GPT-4 given basic personal information about its debate opponents was, in debates where one side proved more persuasive, more persuasive than human opponents 64.4% of the time. A 2024 study in Science found that conversations with GPT-4 reduced participants’ belief in conspiracy theories by about 20%, an effect still present two months later. Persuasive ability can be used to inform or to manipulate.

Extrapolated. Bostrom lists “social manipulation” among a superintelligence’s potential advantages: the ability to model individuals and groups well enough to persuade them of almost anything. This is one of the main reasons containment is thought to be unreliable: a system that can talk to its operators can try to talk its way out. The control page explains.

Strategy and planning

Demonstrated. AI has beaten the best humans at chess (1997) and Go (2016), and agents now carry out multi-step technical tasks lasting hours. Long-horizon planning in the open world, where the rules are unclear and mistakes accumulate, remains a weakness; METR measures how long a task AI can complete on its own, and that horizon, while growing quickly, was measured in hours, not weeks, by mid-2026.

Extrapolated. A superintelligence could, on the extrapolated view, out-plan any person or organisation in business, politics or war. Bostrom calls this “strategizing”. It is the capability that makes the concentration-of-power and loss-of-control risks serious.

Economic activity

Demonstrated. On OpenAI’s GDPval test of 1,320 real work tasks across 44 occupations, published in September 2025, the best model’s output was rated as good as or better than an expert’s in just under half of tasks. On the Remote Labor Index, which gives agents whole freelance projects, the best result on publication in October 2025 was 2.5%; on 2 October 2026 its leaderboard’s top figure was about 21%. AI does many tasks well and fewer whole jobs.

Extrapolated. A superintelligence could do most cognitive work better and more cheaply than people, and, by running many copies, more of it. Bostrom lists “economic productivity” among its decisive advantages. What that would mean for employment, wages and the distribution of income is one of the largest open questions in the field.

AI research itself

Demonstrated. AI already does a large and growing share of the work of building AI at the companies that report it. Anthropic says its models “lead” 26% of its AI research and development work as of August 2026, up from under 1% in February; OpenAI says it has an automated research intern and aims for an automated AI researcher by March 2028. Both are company self-reports. The recursive self-improvement page sets out the evidence.

Extrapolated. This is the capability that could matter most, because it would speed up all the others. A system better than humans at AI research could build better successors, which is the starting point for an intelligence explosion.

The limits that intelligence does not remove

A recurring theme in the serious literature is that intelligence is not the only constraint. Physical laws, the speed of experiments, the time needed to build factories and power stations, the need for data from the real world, and the decisions of people and institutions would all limit what even a superintelligence could do, and how fast. Anthropic’s own 2026 account of AI building AI puts it plainly: more intelligence cannot learn what a drug does over decades of use, or hold elections sooner than a constitution allows. Whether these limits would slow a superintelligence by years or barely at all is genuinely unknown.

Bostrom’s list of the six areas in which a superintelligence might hold a decisive advantage, “cognitive superpowers” as they are often called, is useful as a summary of the extrapolated view: improving its own intelligence, strategy, social manipulation, hacking, technology research and economic productivity. Each of them has a demonstrated counterpart in today’s AI. None has yet been shown at a level beyond the best humans across the board.