The superintelligence era: what the research shows

What does superintelligence mean?

In the researchers’ sense, a superintelligence is an intellect “much smarter than the best human brains in practically every field”, in Nick Bostrom’s 1998 definition. ² He excluded collective entities: a company or the scientific community does not count. ²

The Levels of AGI framework, an ICML 2024 position paper last revised in September 2025, puts the term on one scale. ¹ It rates a system on two axes, depth of performance and breadth of generality, with five levels above No AI. ¹ General AI at the top level, Superhuman, is what it calls artificial superintelligence. A system that is superhuman at one narrow task does not qualify. ¹

AGI is a lower bar, and it is defined separately from that scale. A 33-author paper from October 2025 defines it as “matching the cognitive versatility and proficiency of a well-educated adult”, scored across 10 cognitive domains. ³ Its authors find current systems uneven across those domains, with long-term memory storage among the weak points. ³

Since 29 September 2026 the phrase has a second meaning. ⁴ A US executive order signed that day directs federal departments and agencies, to the maximum extent permitted by law, to use ‘Super Intelligence’ in place of ‘Artificial Intelligence’ in their non-statutory documents. ⁴ For the order, the term covers the same technologies as the existing statutory definition of AI, 15 U.S.C. 9401(3). ⁴ The White House fact sheet presents it as a name that conveys the capabilities of the technologies being developed today. ⁵

A legal analysis published by the IAPP on 30 September 2026 calls the immediate change linguistic: existing legal definitions stay as they are for now. ⁶ It warns that organisations will manage two vocabularies that may not always refer to the same systems, and notes that proposed legislative language for a federal definition is due within 60 days. ⁶

This page uses the research meaning. In that sense, none of the sources it cites reports a superintelligence.

The idea is old. The label is new.

1. 1950 · The question. Alan Turing’s paper in Mind opens with “Can machines think?” and replaces it with a testable imitation game. ⁷
2. 1955 · The programme. A proposal dated 31 August 1955 asks for a two-month, ten-person study of artificial intelligence at Dartmouth College in summer 1956, on the conjecture that any feature of intelligence could be described precisely enough for a machine to simulate it. ⁸
3. 1965 · The idea. The mathematician I. J. Good describes an ultraintelligent machine, one able to exceed the intellectual work of any person however clever, and argues that machines designing better machines would set off an intelligence explosion (Advances in Computers, vol. 6). ⁹
4. 1993 · The window. At a symposium sponsored by NASA Lewis Research Center and the Ohio Aerospace Institute, Vernor Vinge predicts the means to create superhuman intelligence within thirty years, calls the event the Singularity, and says he would be surprised by a date before 2005 or after 2030. ¹⁰
5. 2014 · The modern definition. Nick Bostrom’s book Superintelligence: Paths, Dangers, Strategies describes an intellect far beyond human thinking in virtually every domain that matters, and argues that such a system would be hard to control. ¹¹
6. 29 September 2026 · The label. A US executive order directs federal agencies to use ‘Super Intelligence’ as their word for the technologies already covered by the statutory definition of AI; see the two meanings. ⁴

Every entry but the last is about an idea. The last is about a word.

What has actually been measured?

METR (Model Evaluation & Threat Research) measures AI agents by the length of software task they can finish, timed by how long the same task takes a human professional. ¹² The length that frontier agents finish with 50% reliability doubled roughly every 7 months over the six years to early 2025. ¹²

The International AI Safety Report, published in February 2026, adopts the same figure: a doubling about every seven months for some software engineering tasks. ¹³ It adds that performance remains jagged: frontier systems still fail at some seemingly simple tasks. ¹³

The inputs are growing too. Epoch AI estimates that the compute used to train frontier language models has grown about 5x a year since 2020, with a 90% interval of 4x to 6x, which is a doubling about every 5.2 months. ¹⁴

The money follows the curve. The International Energy Agency reported in April 2026 that capital expenditure by the largest technology companies passed USD 400 billion in 2025 and was expected to rise by another 75% in 2026. ¹⁵ By the agency’s count, the capital expenditure of five technology companies is now larger than global investment in oil and natural gas production. ¹⁵

METR states those limits itself. Time horizons differ between domains by orders of magnitude, and are 40 to 100 times lower on visual computer-use tasks. ¹⁶ A 50% horizon does not mean a task can be handed over: some reliability-critical work needs 98% success or more before it is worth automating. ¹⁶ Its error bars have historically spanned roughly a factor of two in each direction. ¹⁶

Measured progress is real, and it is narrow.

A task a system finishes half the time is a task someone still has to check. ¹⁶ The curve moves the checking; it does not remove it.

When could it arrive?

Most forecasts here are about AGI-level milestones, not superintelligence: machines outperforming people at every task, or a first general AI system. ^17,18 Only the scenario narrates superintelligence itself, and it is a story, not a forecast. ¹⁹ Each is given as a range, with the year it was made and its method.

In a survey fielded in October 2023 and published in the Journal of Artificial Intelligence Research, 2,778 researchers who had published at six major AI venues gave a 10% chance by 2027 and a 50% chance by 2047 that unaided machines outperform humans at every possible task. ¹⁷ That 50% year is 13 years earlier than in the same team’s 2022 survey. ¹⁷

The same survey gives a much later date for a related question. A 50% chance that all human occupations become fully automatable came out at 2116, more than 60 years after the every-task estimate. ¹⁷ The authors say they do not know what causes the gap: it would seem to be a framing effect if both questions describe the same event, but the question sets also differ in other ways. ¹⁷

On the forecasting platform Metaculus, the community median for when the first general AI system will be devised, tested and publicly announced was 20 December 2030, with a range of April 2028 to March 2037, from about 1,900 forecasters, when read on 1 October 2026. ¹⁸ It is a live aggregate, so the date can move after that reading. ¹⁸

AI 2027, published on 3 April 2025 by five authors at the AI Futures Project, is a scenario rather than a forecast. ¹⁹ It narrates superhuman coding in early 2027 and superintelligence by the end of that year, and its authors later noted that 2027 was their most likely year for AGI at publication, with their medians somewhat longer. ¹⁹

Plan for a range, not a date. A decision that only works if one forecast is right is a bet.

What could slow it down?

Each serious objection below is set against a measured number, so the scepticism carries the same weight as the curve.

Princeton’s Arvind Narayanan and Sayash Kapoor argue that AI is a normal technology, whose economic and social effects will arrive slowly, over decades. ²³ In a paper published in April 2025, they describe superintelligent AI “as incoherent as usually conceptualized”. ²³ It is a named, serious position, and this page treats it as one.

The case for caution is measured too.

How are governments responding, including India?

Law and policy are stated here as the sources state them, with their dates. India comes first, because this page is written for a business there.

India’s position has two parts. Its AI Governance Guidelines, published in November 2025, conclude that existing laws on IT, data protection, consumer protection and the civil and criminal codes can govern many AI risks, so a separate AI law is not needed at this stage. ²⁴ The binding rule is narrower: since 20 February 2026, synthetically generated audio-visual content made with an intermediary’s tools must be prominently labelled, with a prefixed disclosure for audio, and carry permanent metadata or another provenance mechanism where technically feasible. ²⁵ That is law now, not a proposal.

The International AI Safety Report, published on 3 February 2026, concludes that current risk-management measures do not reliably prevent harm in every setting. ¹³ It describes defence-in-depth: evaluations, technical safeguards, monitoring and incident response, used together. ¹³

The response aimed at superintelligence itself is a call, not a law. The Future of Life Institute’s Statement on Superintelligence, released on 22 October 2025, calls for a prohibition on developing it until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in. ²⁶

The rule now in force in India concerns labelling synthetic media, not superintelligence.

What should a business in India decide this year?

Adoption in India is already high. In a 2025 survey cited by the 2026 AI Index, more than 80% of employees in India said they used AI at work regularly or semi-regularly, against 58% worldwide. ²⁸ The open question is value, not adoption.

Each decision below leads with the number behind it. The suggested owner on each is editorial guidance, not a finding.

1. Redesign the work. Nearly three-quarters of AI high performers in McKinsey’s August 2026 survey had fundamentally redesigned workflows because of AI, against about one-quarter of other respondents, and they were twice as likely to have defined processes for measuring impact. ²⁹ Set the metric before the build. Suggested owner: the business owner.
2. Keep switching cheap. The cost of querying a model at a fixed benchmark score (64.8% on MMLU) fell from USD 20 to USD 0.07 per million tokens between November 2022 and October 2024, a more than 280-fold drop. ³⁰ Avoid long lock-ins and plan to re-price. Suggested owner: whoever signs contracts.
3. Use public compute. The IndiaAI Mission had expanded shared compute capacity to more than 45,000 GPUs as of June 2026, and by August 2026, 237 projects had used subsidised AI compute. ³¹ Price the public option before committing to private capacity. Suggested owner: the technology lead.
4. Keep a person checking. An independent database lists 2,097 court and tribunal decisions worldwide dealing with AI-hallucinated material such as fabricated citations, 16 of them from India, as of 30 September 2026. ³² Name a reviewer for every output that leaves the business. Suggested owner: a named reviewer.

Each decision rests on a number that can be re-read. None rests on a forecast.

Where Run AI stands.

Run AI does not build superintelligence. It builds films, voices, software and agents with today’s systems, behind gates, on a written plan the client approves.

That stance is a practice, written down in each division’s own pages. Each item below links to the division that runs it.

1. Six gates between a generated shot and your brand. ³³ See Run AI Studios.
2. Every key still is human-approved before any motion spend is committed. ³³ See Run AI Studios.
3. Every real likeness and voice runs on written, revocable consent with a documented chain. ³⁴ See Run AI Voice.
4. Every release — film, agent or app — passes a written quality bar before a client sees it. ³⁵ See Run AI Forge.
5. Irreversible actions require sign-off — payments, sends, deletions never run unattended. Every action, decision and hand-off recorded — replayable end to end. ³⁶ See Run AI Agents.
6. On an agent pilot, in the division’s own words: “You judge us on a number you chose before we started.” ³⁶ See Run AI Agents.

The International AI Safety Report describes defence-in-depth: evaluations, technical safeguards, monitoring and incident response, used together. ¹³ A written quality bar is one evaluation and agent logs support monitoring; neither makes the work risk-free, and the report says current measures do not reliably prevent harm in every setting. ¹³

None of this depends on when, or whether, superintelligence arrives.

Questions

Has superintelligence arrived?

Not in the sense researchers use the word. The Levels of AGI framework, last revised in September 2025, lists artificial superintelligence, general AI at its Superhuman level, as not yet achieved. ¹ The US executive order of 29 September 2026 applies the name ‘Super Intelligence’ to the technologies already covered by the statutory definition of AI. ⁴ A 33-author definition of AGI, published in October 2025, finds current systems uneven across cognitive domains, with weak long-term memory storage. ³

What did the US executive order of 29 September 2026 change?

It directs federal departments and agencies, to the maximum extent permitted by law, to use ‘Super Intelligence’ and ‘SI’ in place of ‘Artificial Intelligence’ and ‘AI’ in official communications and other non-statutory documents. ⁴ For the order, the term covers the same technologies as the existing statutory definition of AI (15 U.S.C. 9401(3)). ⁴ A legal analysis published by the IAPP on 30 September 2026 describes the immediate change as one of language, and notes that proposed legislative language for a federal definition is due within 60 days. ⁶

When do researchers expect machines to outperform people at every task?

Estimates differ by decades, so this page reports them as ranges. In a peer-reviewed survey fielded in October 2023, 2,778 researchers gave a 10% chance by 2027 and a 50% chance by 2047 that unaided machines outperform humans at every possible task. ¹⁷ That question is about matching or beating people at every task, not about superintelligence as researchers define it. The survey’s authors caution that AI researchers have no known special skill at forecasting. ¹⁷ Run AI adds no forecast of its own.

Do businesses in India have to label AI-generated content?

India’s IT Rules amendment, in force since 20 February 2026, requires synthetically generated audio-visual content made with an intermediary’s tools to be prominently labelled, with a prefixed disclosure for audio, and to carry provenance metadata where technically feasible. ²⁵ The rule sits in India’s intermediary guidelines, so how it applies to a particular business is a question for that business’s legal adviser. India’s AI Governance Guidelines, published in November 2025, see no need for a separate AI law at this stage. ²⁴

Does Run AI build superintelligence or AGI?

No. It builds films, voices, software and agents with today’s systems, behind gates, on a written plan the client approves. Every key still is human-approved before any motion spend is committed. ³³ Every release — film, agent or app — passes a written quality bar before a client sees it. ³⁵ Irreversible actions require sign-off — payments, sends, deletions never run unattended. ³⁶

What should a business in India do first?

Set the metric before the build. In McKinsey’s August 2026 survey, AI high performers were twice as likely as other respondents to report defined processes for measuring the impact of AI initiatives. ²⁹ Then keep a named person checking every output that leaves the business.

Sources

Numbered in the order the page first cites them. Quotes are the sources’ own words; everything else is paraphrased. Read 1 October 2026.

1. Levels of AGI for Operationalizing Progress on the Path to AGI (arXiv:2311.02462v5)arXiv / ICML 2024 position paper · 4 November 2023 (v1); 24 September 2025 (v5) · arxiv.org“a common language to compare models, assess risks, and measure progress”
2. How Long Before Superintelligence?Nick Bostrom (Oxford University); originally International Journal of Futures Studies, 1998, vol. 2 · 1997; revised 25 October 1998 · nickbostrom.com“much smarter than the best human brains in practically every field”
3. A Definition of AGI (arXiv:2510.18212v3)arXiv (Hendrycks, Song, Szegedy et al.) · 21 October 2025 (v1); 3 December 2025 (v3) · arxiv.org“matching the cognitive versatility and proficiency of a well-educated adult”
4. Inaugurating The Era Of Super IntelligenceThe White House · 29 September 2026 (checked 1 October 2026) · whitehouse.gov“The terminology used by the Federal Government should reflect the transformative capabilities”
5. Fact Sheet: President Donald J. Trump Inaugurates The Era of Super IntelligenceThe White House · 29 September 2026 (checked 1 October 2026) · whitehouse.gov“The name Super Intelligence conveys the true capabilities of the technologies being developed today”
6. AI by any other name? White House directs federal agencies to enter the era of 'super intelligence'IAPP (International Association of Privacy Professionals); author Jennifer Dickey, Dykema · 30 September 2026 (checked 1 October 2026) · iapp.org“The immediate change is linguistic, not a wholesale rewrite of U.S. laws”
7. Computing Machinery and Intelligence (Mind, Vol. LIX, No. 236, pp. 433–460)Mind (Oxford University Press) · October 1950 · academic.oup.com“Can machines think?”
8. A Proposal for the Dartmouth Summer Research Project on Artificial IntelligenceStanford University (John McCarthy archive) · 31 August 1955 · www-formal.stanford.edu“so precisely described that a machine can be made to simulate it”
9. Speculations Concerning the First Ultraintelligent MachineAcademic Press (Advances in Computers, vol. 6) · 1965 · doi.org“far surpass all the intellectual activities of any man however clever”
10. The Coming Technological Singularity: How to Survive in the Post-Human EraVISION-21 Symposium, NASA Lewis Research Center and Ohio Aerospace Institute (text hosted at San Diego State University) · 30–31 March 1993 · edoras.sdsu.edu“I’ll be surprised if this event occurs before 2005 or after 2030.”
11. Superintelligence: Paths, Dangers, StrategiesOxford University Press · 3 July 2014 · global.oup.com“greatly exceeds the cognitive performance of humans in virtually all domains of interest”
12. Measuring AI Ability to Complete Long TasksMETR (Model Evaluation & Threat Research) · 19 March 2025 · metr.org“has been doubling approximately every 7 months for the last 6 years”
13. 2026 Report: Extended Summary for PolicymakersInternational AI Safety Report · 3 February 2026 · internationalaisafetyreport.org“Current risk management measures do not reliably prevent harm in all settings”
14. Trends in AI (dashboard)Epoch AI · Updated 5 February 2026 · epoch.ai
15. Key Questions on Energy and AI: Executive summaryInternational Energy Agency · 16 April 2026 · iea.org“now larger than global investment in oil and natural gas production”
16. Clarifying limitations of time horizonMETR · 22 January 2026 · metr.substack.com“differs between domains by orders of magnitude”
17. Thousands of AI Authors on the Future of AIJournal of Artificial Intelligence Research (Grace et al.; AI Impacts with University of Bonn, University of Oxford and independent co-authors) · 6 October 2025 (JAIR vol. 84); survey fielded October 2023 · doi.org“unaided machines outperforming humans in every possible task was estimated at 10% by 2027”
18. When will the first general AI system be devised, tested, and publicly announced? (question 5121)Metaculus · Live aggregate, read 1 October 2026 (IST) · metaculus.com
19. AI 2027AI Futures Project · 3 April 2025 (note added 22 November 2025) · ai-2027.com“2027 was our modal (most likely) year at the time of publication”
20. AAAI 2025 Presidential Panel on the Future of AI ResearchAssociation for the Advancement of Artificial Intelligence (AAAI) · March 2025 · aaai.org
21. AI Index Report 2026, Chapter 2: Technical PerformanceStanford Institute for Human-Centered AI (HAI) · April 2026 (takeaways article dated 13 April 2026) · hai.stanford.edu
22. Will we run out of data? Limits of LLM scaling based on human-generated dataEpoch AI · 6 June 2024 · epoch.ai“language models will fully utilize this stock between 2026 and 2032”
23. AI as Normal Technology: An alternative to the vision of AI as a potential superintelligenceKnight First Amendment Institute at Columbia University · 15 April 2025 · knightcolumbia.org“‘superintelligent’ AI, which we view as incoherent as usually conceptualized”
24. India AI Governance Guidelines: Enabling Safe and Trusted AI InnovationMinistry of Electronics and Information Technology (IndiaAI), via PIB · November 2025 · static.pib.gov.in“a separate law to regulate AI is not needed”
25. Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026 — G.S.R. 120(E)Ministry of Electronics and Information Technology, Gazette of India (Extraordinary) · 10 February 2026 · egazette.gov.in
26. Statement on SuperintelligenceFuture of Life Institute · 22 October 2025 · superintelligence-statement.org“significantly outperform all humans on essentially all cognitive tasks”
27. AI ActEuropean Commission (Shaping Europe’s digital future) · Last updated 3 August 2026 · digital-strategy.ec.europa.eu
28. The 2026 AI Index Report (Chapter 9, Public Opinion, pp. 362, 370–371)Stanford Institute for Human-Centered AI (HAI), citing University of Melbourne and KPMG International (2025) · April 2026 · hai.stanford.edu
29. The state of AI in 2026: On the road to ROIMcKinsey & Company (QuantumBlack) · 25 August 2026 · mckinsey.com“fundamentally redesign workflows that are enabled by AI rather than insert AI”
30. AI Index 2025: State of AI in 10 ChartsStanford Institute for Human-Centered AI (HAI) · 7 April 2025 · hai.stanford.edu
31. Semiconductor and AI Revolution: Powering the Future, From Chips to AIPress Information Bureau (PIB Research factsheet) · 13 August 2026 · pib.gov.in
32. AI Hallucination Cases DatabaseDamien Charlotin (independent researcher; CC BY 4.0) · Last updated 30 September 2026 · damiencharlotin.com“2097 cases identified so far”
33. RUNAI STUDIOS · No AI tells; RUNAI STUDIOS · How it’s madeRun AI Intelligence · As published on /studios, 1 October 2026
34. RUNAI VOICE · Built for procurementRun AI Intelligence · As published on /voice, 1 October 2026
35. RUNAI FORGE · No surprisesRun AI Intelligence · As published on /forge, 1 October 2026
36. RUNAI AGENTS · Security architecture; RUNAI AGENTS · The experienceRun AI Intelligence · As published on /agents, 1 October 2026