# AI, AGI and Superintelligence, 2026-2031: Expert Consensus After the U.S. SI Rebranding

> A structured review of 24 AI experts and 12 university publications separates the U.S. SI rebranding from technical superintelligence and maps the five-year consensus.

- Canonical page: https://mtfinstitute.com/insights/ai-agi-superintelligence-2026-2031-expert-consensus-us-si-rebranding/
- Content type: Article
- Editorial category: Research & Reports
- Publisher: MTF Institute of Management, Technology and Finance
- Author: MTF Institute Editorial Team- Published: 2026-09-24
- Updated: 2026-09-24
- Language: English
- Topics: Artificial Intelligence, AGI, Superintelligence, AI Safety, Technology Forecasting

## Executive answer

On 22 September 2026, U.S. President Donald Trump told the United Nations that artificial intelligence would be “hereinafter officially called Super Intelligence.” The statement is real and consequential as political language, but it does not establish that current AI has crossed the technical threshold usually meant by superintelligence. Contemporary systems remain uneven: they can exceed people on selected tests while failing ordinary perception, long-horizon execution, robust truthfulness and unfamiliar physical tasks.

This report reviews 24 named experts and 12 publications from leading universities to ask what is likely to change by September 2031. The strongest consensus is not that one acronym will win. It is that AI capabilities will keep advancing, especially in reasoning, software, research assistance, multimodal interaction, agents and robotics; that reliability, evaluation, security and human control will become more important as systems gain autonomy; and that institutions will adopt capability unevenly. There is no defensible expert consensus that broad AGI will certainly arrive within five years, and still less consensus that broadly superhuman, controllable superintelligence will do so.

The practical conclusion is to retain three distinct terms. **AI** is the broad field and present technology class. **AGI** is a disputed threshold for general, adaptable human-level capability. **Superintelligence**, often shortened to SI or ASI, should remain a higher threshold: systems that substantially surpass humans across most relevant cognitive domains. Calling today’s AI “SI” may change government messaging. It cannot, by itself, change measured capability.

## 1. What happened in the United States

The starting point is an official statement, not internet rumor. In a [White House release](https://www.whitehouse.gov/releases/2026/09/president-trump-at-the-united-nations-while-others-have-talked-i-have-acted/) dated 22 September 2026, President Trump said that the United States rejected a global control scheme for “Artificial Intelligence … hereinafter officially called ‘Super Intelligence.’” A [White House video](https://www.whitehouse.gov/videos/president-trump-delivers-remarks-sep-22-2026/) preserves the speech context. Contemporaneous reporting by the [BBC](https://www.bbc.com/news/articles/cqy4z9pv4w0po), [Associated Press](https://apnews.com/article/692e1e171a791be0c9b9c1f56718a0d0), [Axios](https://www.axios.com/2026/09/22/trump-ai-super-intelligence-rebrand) and [Forbes India](https://www.forbesindia.com/article/news/trumps-proposal-to-change-ai-to-si-tech-leaders-call-it-misleading/2998676/1) described the proposal and early administrative use.

The evidence supports a precise formulation: Trump announced a preferred U.S. nomenclature, and at least part of the State Department reportedly began replacing “AI” with “SI” in documents. As of the source cutoff, the review did not identify a binding technical standard that had redefined all current AI systems as superintelligent. The scope, legal status and durability of the terminology were still developing. A [CNN Portugal video](https://cnnportugal.iol.pt/videos/trump-continua-igual-a-si-proprio-e-a-dizer-que-resolveu-oito-conflitos/6ab2afad0cf22b9f8904c6c9) supplied broader political-speech context but did not contain enough technical detail to establish the meaning of SI; it is therefore cited as context, not as technical evidence.

The distinction matters because “superintelligence” already has a technical history. It normally describes intelligence that exceeds the best human capability across a broad range of domains, not merely a useful chatbot or a system that is superhuman on one benchmark. The U.S. government itself had recently used separate categories. The [2026 Economic Report of the President](https://www.whitehouse.gov/wp-content/uploads/2026/04/2026-Economic-Report-of-the-President.pdf) discussed specialized AI, AGI and superintelligence as different concepts. Political rebranding therefore collides with, rather than automatically replaces, existing technical usage.

The BBC and Forbes India captured that collision. Simon Coghlan argued that applying superintelligence to current systems exaggerates their ability. Kokil Jaidka noted that the name borrows prestige from a much larger technical leap. Ben Leong expected professionals to continue reserving SI for a capability level beyond present platforms. Those reactions do not prove which word governments will use. They show why researchers, businesses and journalists need explicit definitions.

## 2. Definitions used in this report

**Artificial intelligence (AI)** is the umbrella term for machine systems that perform tasks associated with perception, prediction, language, reasoning, planning or action. It includes narrow classifiers, recommendation systems, language models, agents and robots. A system can be AI without being general or superhuman.

**Artificial general intelligence (AGI)** has no universally accepted test. Some laboratories use an economic definition: performance equal to or better than people across most economically valuable cognitive work. Others require efficient learning of unfamiliar tasks, continual adaptation, causal understanding or the ability to generate genuinely new scientific explanations. These definitions can produce different arrival dates from the same evidence.

**Superintelligence (SI or ASI)** is used here for broad capability substantially beyond humans, not merely beyond an average worker on selected tasks. Domain-specific superhuman performance already exists in games, protein structure prediction and some mathematical or coding evaluations. That is important, but it is not identical to a general system that can learn, plan and act more effectively than human organizations across domains.

The report therefore treats the U.S. usage as a nomenclature event and the technical concept as a capability claim. Every later conclusion follows from evidence about capability, not from the new label.

## 3. Research question and method

The research question is: **Across current expert statements and university evidence, what development path for AI, AGI and superintelligence is most defensible for the five years from September 2026 to September 2031?**

The review uses a purposive, viewpoint-diverse sample of 24 experts. It includes frontier-laboratory leaders, foundational researchers, benchmark designers, AI-safety researchers, institutional-diffusion scholars and explicit forecasters. A direct essay, institutional publication, paper, testimony or official interview was preferred. Each record was coded on five dimensions: capability path through 2031; AGI outlook; superintelligence outlook; principal bottleneck; and safety or governance priority. A missing view was recorded as “not assessed,” never converted into agreement.

The second layer consists of 12 publications from Stanford, Princeton, Harvard, MIT, UC Berkeley, Carnegie Mellon, Oxford and Cambridge. They examine capability trends, labor effects, transparency, agent autonomy, benchmark integrity, workplace performance and preparedness. University evidence was not treated as a personal prediction unless the author made one explicitly.

Consensus thresholds were fixed before synthesis. At least 75% agreement among relevant explicit views is called strong consensus; 50% to 74% is moderate consensus; below 50% is divided. The counts describe this source set, not the worldwide AI community. Full coding is preserved in the archival CSV files.

## 4. What 24 experts expect

### 4.1 The acceleration group

Several laboratory leaders expect a rapid transition from chat systems to systems that reason, use tools, conduct research and act over longer horizons. [Sam Altman](https://blog.samaltman.com/the-gentle-singularity) writes that humanity is close to digital superintelligence and emphasizes scientific discovery, software and robots. [Dario Amodei](https://darioamodei.com/essay/machines-of-loving-grace) defines “powerful AI” as systems more capable than Nobel-level experts across fields and able to perform virtual work through digital interfaces; he has argued that such systems could arrive early, while acknowledging large uncertainty.

[Demis Hassabis and Google DeepMind](https://deepmind.google/blog/taking-a-responsible-path-to-agi/) state that AGI could arrive in coming years and pair that claim with a safety framework. Hassabis’s stricter conception includes discovery, world models and continual learning, which pushes his public timing toward roughly five to ten years rather than the shortest laboratory forecasts. DeepMind’s 2026 paper [From AGI to ASI](https://deepmind.google/research/publications/239142/) identifies multiple post-AGI pathways: scaling, architectural shifts, recursive improvement and multi-agent collectives. Shane Legg similarly uses graded levels of AGI rather than a single theatrical moment.

[Mark Zuckerberg](https://about.fb.com/news/2025/07/personal-superintelligence-for-everyone/) argues that self-improving systems and personal superintelligence are on the horizon, with agents embedded in everyday devices. [Mustafa Suleyman](https://blogs.microsoft.com/blog/2026/03/17/announcing-copilot-leadership-update/) frames Microsoft’s next five years around frontier models and “Humanist Superintelligence”: advanced, bounded systems directed toward medical, scientific and productivity goals while preserving human control. [Ilya Sutskever’s SSI](https://ssi.inc/) says superintelligence is within reach but makes safety a coequal engineering objective, not a later compliance layer.

This group supplies the strongest case for AGI or superintelligence within the review window. Yet its members do not predict the same artifact. Altman describes broadly available intelligence and robots; Amodei emphasizes virtual expert labor and accelerated biology; Hassabis requires deeper generality and scientific invention; Zuckerberg prioritizes personal agents; Suleyman prefers domain-focused, controllable superintelligence. Agreement on speed does not equal agreement on the finish line.

### 4.2 The safety-first group

[Yoshua Bengio](https://yoshuabengio.org/en/publication/superintelligent-agents-pose-catastrophic-risks-can-scientist-ai-offer-safer-path) accepts that increasingly capable general agents are plausible but argues that unchecked agency creates deception, misuse and loss-of-control risks. His proposed Scientist AI is a non-agentic, uncertainty-aware system that models and explains the world rather than autonomously pursuing open-ended goals. The forecast is not “stop useful AI”; it is “change the architecture and objective before autonomy outruns control.”

[Geoffrey Hinton](https://time.com/7339628/geoffrey-hinton-ai/) expects rapid capability growth and assigns meaningful probability to systems becoming more intelligent than people, while warning about misuse and control. [Stuart Russell](https://humancompatible.ai/app/uploads/2023/10/managing_ai_risks-1.pdf) argues that increasingly general systems require objectives and governance designed around provable benefit rather than fixed proxies. [Max Tegmark](https://futureoflife.org/statement/agi-manhattan-project-max-tegmark/) distinguishes beneficial tool AI from an uncontrollable race to AGI. [Roman Yampolskiy](https://arxiv.org/abs/2008.04071) goes further, arguing that full control of advanced AI has not been established and may be impossible.

[Helen Toner](https://www.exponentialview.co/p/the-collapse-of-long-ai-timelines) observes that long timelines to human-level AI have contracted and that dismissing the subject as science fiction is no longer serious, but she emphasizes uncertainty, transparency and policy preparedness. Eric Schmidt has similarly highlighted short AGI timelines, multi-agent systems, long context and action-taking while warning about geopolitical and control implications. These experts disagree on probability and remedy, but converge on a decision rule: the more autonomous and strategically capable the system, the less acceptable it is to postpone evaluation, security and governance.

### 4.3 The architecture-and-benchmark skeptics

Another group expects strong AI progress without accepting that current language-model scaling automatically produces AGI. [Yann LeCun](https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/) argues that advanced machine intelligence needs learned world models, memory, planning and grounded understanding. [Fei-Fei Li](https://hai.stanford.edu/news/why-ai-needs-spatial-intelligence) similarly makes spatial intelligence central: systems must reason about three-dimensional environments, objects and consequences rather than only language correlations.

[François Chollet](https://arcprize.org/blog/arc-prize-2025-results-analysis) defines intelligence as efficient acquisition of new skills. ARC-AGI results show major progress, but the 2025 grand prize remained unclaimed and Chollet concluded that AGI had not arrived. His test is deliberately resistant to memorization and benchmark preparation. [Gary Marcus](https://garymarcus.substack.com/p/six-or-seven-predictions-for-ai-2026) expects reliability, hallucination and agent failures to persist unless the field moves toward hybrid or neurosymbolic systems.

[Andrew Ng](https://www.deeplearning.ai/the-batch/agentic-design-patterns-part-1/) focuses on the practical leverage of agentic workflows: reflection, tool use, planning and multi-agent collaboration. That is a strong five-year deployment thesis, not an assertion that AGI has already been reached. [Jensen Huang’s](https://www.nvidia.com/en-us/on-demand/session/gtctaipei26-stw61044/) roadmap similarly centers on agents, simulation and physical AI, with data and world interaction as the hard problems. Both perspectives predict major economic change before philosophical agreement about AGI.

### 4.4 The normal-technology counterview

[Arvind Narayanan](https://www.cs.princeton.edu/~arvindn/) and [Sayash Kapoor](https://www.cs.princeton.edu/~sayashk/public.html) offer the clearest alternative to sudden-takeoff narratives. Their “AI as normal technology” framework argues that capability diffuses through firms, laws, skills, infrastructure and institutions. Even very capable models do not instantly replace the complementary systems required for real-world impact. They also question AGI as a clean milestone because performance, adoption and social transformation do not arrive together.

This does not imply stagnation. It predicts a different causal chain: better models enable products; products require redesign; redesign meets organizational and regulatory constraints; impact accumulates unevenly. In this view, a five-year forecast should track deployment quality, complementary investment and institutional adaptation rather than only benchmark scores.

### 4.5 Explicit rapid scenarios

[Ben Goertzel](https://goertzel.org/TenYearsToTheSingularity.pdf) expects AGI on a short horizon and sees recursive improvement as a route from general intelligence to superintelligence. [Daniel Kokotajlo and collaborators](https://ai-2027.com/) describe a detailed rapid-progress scenario in which coding and research agents accelerate AI development. The latter is explicitly a scenario, not an observed future and not a consensus estimate. Its value is stress-testing decisions under a fast path; its weakness is that small assumptions about automation, compute, algorithmic progress and organizational response compound dramatically.

## 5. What university evidence adds

Expert forecasts are strongest when tested against observed performance. The [Stanford 2026 AI Index](https://hai.stanford.edu/ai-index/2026-ai-index-report) records accelerating performance in science, mathematics, coding and agents. It also reports a jagged frontier: agents improved sharply on computer-use benchmarks but still failed roughly one in three structured tasks; leading models could excel at competition mathematics yet struggle with basic clock reading; household robots succeeded on only a small minority of tasks. Capability is advancing, but breadth and reliability remain different variables.

The Stanford Index also reports that responsible-AI measurement is not keeping pace with capability. Incidents increased, transparency declined and adversarial attacks weakened safeguards. The [Stanford Foundation Model Transparency Index](https://crfm.stanford.edu/fmti/) supports the same governance concern: outside observers often lack the data needed to compare training, evaluation and deployment risks.

The [Harvard Business School and BCG experiment](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) provides causal workplace evidence. In 758 consultants, AI improved speed and quality on tasks inside its capability frontier, but users became less likely to reach the correct answer on a task outside that frontier. This finding explains why “better than people” is too coarse. The same model can be an accelerator, a plausible-error generator and a coordination burden in adjacent tasks.

Agent research reinforces the point. The [MIT 2025 AI Agent Index](https://aiagentindex.mit.edu/) documents rapidly rising autonomy but weak agent-specific safety disclosure. A [Cambridge-led analysis](https://www.cam.ac.uk/stories/ai-agent-index-safety) found that only four of 30 prominent agents published agent-specific system cards and most disclosed neither internal nor independent safety results. [UC Berkeley’s BenchJack audit](https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/) showed that major agent benchmarks could be gamed to achieve near-perfect scores without solving the intended tasks. Benchmark progress is useful only when the benchmark still measures the capability claimed.

[WebArena](https://webarena.dev/), developed by researchers from UC Berkeley and Carnegie Mellon, and [TheAgentCompany](https://arxiv.org/abs/2412.14161) test agents in realistic websites and workplace environments. Their existence is itself evidence of a field moving from conversation toward action. Their results also reveal long-horizon weaknesses: error accumulation, fragile navigation, poor recovery and limited completion of consequential tasks.

Princeton’s [AI as Normal Technology](https://www.cs.princeton.edu/~arvindn/) framework and CRUX work on open-ended evaluation add an institutional and measurement warning. Impressive demonstrations are not the same as durable production capability, and production capability is not the same as economy-wide replacement. Oxford’s [AGI preparedness report](https://aigi.ox.ac.uk/publications/europe-and-the-geopolitics-of-agi-the-need-for-a-preparedness-plan/) takes a complementary position: uncertainty about near-term AGI should not be used as a reason to avoid preparation. Together, these publications support preparedness without pretending the arrival date is known.

## 6. Consensus results

### Strong consensus: AI capability and deployment will advance

All 24 expert records expect materially stronger or more consequential AI during the review window, although four avoid an AGI timeline. The shared path includes better reasoning, coding, multimodal systems, agents, scientific tools and some forms of robotics. This is the report’s strongest conclusion. It does not depend on accepting AGI or SI terminology.

### Strong consensus: safety, reliability and governance become more important

Twenty-one of 24 records explicitly make reliability, security, human control, evaluation, transparency or governance a material condition. The remaining three focus mainly on capability architecture rather than arguing that safety is irrelevant. University evidence strengthens this consensus: agents are becoming more autonomous while disclosure and benchmark integrity remain inadequate.

### Moderate consensus: agents are the main near-term interface

Eighteen of 24 experts explicitly emphasize agents, tool use, planning, research automation, workflow execution or embodied action. This is the most practical five-year forecast. Users will increasingly delegate sequences of work, not merely request text. That shifts the critical metric from answer quality to controlled task completion: permissions, provenance, monitoring, recovery, cost and accountability.

### Divided: AGI by September 2031

Among 20 experts with a codable AGI position, ten place AGI or an equivalent powerful system within the five-year window or treat it as their central rapid scenario. Six consider it possible but uncertain. Four reject a discrete milestone or argue that current approaches are insufficient. The largest category is exactly 50%, which meets only moderate agreement under the prespecified rule and is highly sensitive to definition. The honest conclusion is **plausible but not a consensus forecast**.

### Divided: broad superintelligence by September 2031

Only five expert records clearly expect broad or named superintelligence within the five-year window. Nine treat it as possible, a scenario or a risk requiring preparation. Ten either do not forecast it, reject the near-term framing or focus on domain-specific superhuman systems. No category reaches 50%. The evidence supports preparation for high-impact advanced systems; it does not support declaring that broad SI already exists or will certainly exist by 2031.

## 7. The five-year outlook

### 2026-2027: from copilots to bounded agents

More products will plan and execute multi-step work across software, documents, browsers and enterprise systems. Reliability will improve in bounded environments with explicit tools and verification. Failures will remain common when objectives are ambiguous, environments change or an agent encounters hostile content. Organizations will create permission tiers, audit trails and human approval gates.

### 2027-2029: research and domain superhumanity expand

AI will become stronger in coding, mathematical assistance, drug and materials discovery, design and simulation. Some systems will be described as “superintelligent” within a domain. That usage can be defensible if the domain and test are explicit. It should not be generalized to all cognition. Physical AI will advance fastest in structured factories, vehicles and laboratories, more slowly in open homes and unpredictable public environments.

### 2029-2031: the uncertainty widens

If continual learning, world models, long-horizon reliability and automated AI research improve together, systems could cross some AGI definitions. If benchmark gains continue without robust adaptation and deployment, AI may instead become an extremely capable but institutionally bounded technology. Both paths imply large economic and governance effects. Neither logically guarantees broad superintelligence.

## 8. A practical terminology and decision framework

Organizations should label a system by evidence, not ambition.

1. Use **AI** for present systems and the field as a whole.
2. Use **agentic AI** when a system can plan, call tools and act over multiple steps; state its permission and supervision boundary.
3. Use **domain-superhuman AI** only when a reproducible evaluation shows superior performance in a named domain.
4. Use **AGI** only with a declared definition and test for generalization, adaptation and breadth.
5. Use **superintelligence or ASI** only for broad, sustained capability substantially above humans, together with evidence about control and reliability.

For every capability claim, ask five questions: What exact tasks were tested? Were tasks unfamiliar? Does performance survive adversarial and real-world conditions? Can the system recover from errors over long horizons? Who remains accountable for consequential actions? This prevents a naming decision from becoming an unsupported procurement, policy or safety assumption.

## 9. Final consensus statement

The 2026 U.S. move from “AI” toward “SI” is best understood as political and strategic branding at the source cutoff, not a scientific finding that current systems are superintelligent. Across 24 experts and 12 university publications, the defensible consensus is narrower and more useful.

By 2031, AI will probably be more agentic, multimodal, scientifically capable and physically grounded. It will perform a growing number of bounded tasks at or above expert level. The economic impact will depend on institutions, complementary investment and trustworthy deployment, not on model capability alone. AGI within five years is plausible under several influential definitions, but experts remain divided because they disagree about required generality, adaptation and measurement. Broad superintelligence within five years is possible enough to justify safety work and preparedness, but not established enough to serve as a baseline forecast.

The most robust policy is therefore dual-track: accelerate verifiable beneficial applications while building evaluation, security, transparency, governance and control mechanisms for systems that may become much more autonomous. Names can change overnight. Capability, reliability and human control must be demonstrated.

## Limitations

This is a structured qualitative review, not a representative poll or meta-analysis. Public statements may be strategic, promotional or incomplete. Definitions of AGI and superintelligence differ. The review cannot observe proprietary model results. Some sources are essays or scenarios rather than peer-reviewed studies. University reports measure current systems and cannot prove a five-year future. Coding compresses nuanced positions, although the archive preserves direct sources and `not_assessed` values. The source cutoff is 24 September 2026; fast-moving claims should be rechecked before high-stakes decisions.

## Publication record

The version of record is archived on Zenodo under [DOI 10.5281/zenodo.22945817](https://doi.org/10.5281/zenodo.22945817). The record includes the [searchable final PDF](https://zenodo.org/records/22945817/files/MTF-RR-2026-09-24-02.pdf?download=1), the 24-expert matrix, the 12-publication university matrix and the complete reproducibility archive.

> ### Turn AI uncertainty into governed executive decisions
> Leaders need more than terminology: they need strategy, finance, operating-model, governance, risk and change capability. The [Advanced Executive Management & Business Administration programme](https://mtfinstitute.com/programs/advanced-executive-management-business-administration/#enroll) provides an applied path for integrating those decisions. Review the curriculum, evidence requirements and current terms before enrolling.

## References

The complete expert coding and university publication matrix are included in the Zenodo archival package. Core sources include the [White House statement](https://www.whitehouse.gov/releases/2026/09/president-trump-at-the-united-nations-while-others-have-talked-i-have-acted/), [BBC report](https://www.bbc.com/news/articles/cqy4z9pv4w0po), [Forbes India report](https://www.forbesindia.com/article/news/trumps-proposal-to-change-ai-to-si-tech-leaders-call-it-misleading/2998676/1), [CNN Portugal context](https://cnnportugal.iol.pt/videos/trump-continua-igual-a-si-proprio-e-a-dizer-que-resolveu-oito-conflitos/6ab2afad0cf22b9f8904c6c9), [Stanford AI Index 2026](https://hai.stanford.edu/ai-index/2026-ai-index-report), [International AI Safety Report 2026](https://internationalaisafetyreport.org/), [Harvard jagged-frontier study](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf), [MIT AI Agent Index](https://aiagentindex.mit.edu/) and [DeepMind’s AGI-to-ASI report](https://deepmind.google/research/publications/239142/).



## Citation

When citing or summarizing this material, link to the canonical HTML page: https://mtfinstitute.com/insights/ai-agi-superintelligence-2026-2031-expert-consensus-us-si-rebranding/