Scenarios 2027 / Scenarios
Rated at Standard Severe. Likelihood type: systemic. Source of scores: ginc-desk-v0.3. Upside scenario: impact levels measure the scale of change, not loss.
The announcement is three paragraphs long. The laboratory's research agents have been improving their own training methods for eleven months, and the newest system has not been evaluated by any human who fully understands its results. Two rivals say the same within a quarter. By then the evidence is in the journals: a room-temperature catalyst, a malaria vaccine designed and trialled in simulation before a single dose is made, a reactor design that passes review in weeks. A year of this produces more patents than the previous twenty. Growth in the two economies that host the data centres runs at 9 per cent and their electricity demand at twice that. Elsewhere the picture is stranger. Finance ministries watch corporate profits migrate to three jurisdictions. A mid-sized country's entire professional-services export sector is repriced to the cost of inference. Defence staffs brief that the side with the better models now designs, tests and fields a weapon in months, and that nobody can verify what the other side has. Governments that asked for a seat at the table are offered an API key. The leaders' summit agrees a notification protocol and a compute registry, and neither can be inspected by the people who signed it. Hospitals are the first to feel unambiguous benefit, and the first to ask who is liable. The countries that do well are not the ones with the most engineers. They are the ones with power to spare, a state that can adopt as fast as its firms, and something the owners of the models need.
The dateline is illustrative, not a forecast. The narrative is hypothetical; the historical anchors below are real events.
AI systems automate AI research itself. Each generation designs a better successor and, within about two years of the loop closing, systems outperform the best human experts in every field of science and engineering. A century of research progress is compressed into a decade or less: new drugs, materials, energy designs and weapons arrive faster than institutions can absorb them. Output grows at rates with no precedent. The gains go first to whoever owns the models, compute and energy. Many researchers class this as the largest risk of all; the Library carries it as an upside scenario because the central case is abundance, with control, concentration and distribution as the open problems.
A shock most would count as progress. It is not upside for everyone: each record names who gains and who loses. Impact levels measure the scale of change, in either direction, and loads mark the capabilities a nation needs in order to capture the gain or absorb the loss.
| Event | Date | What happened | Calibrates |
|---|---|---|---|
| I. J. Good, 'Speculations Concerning the First Ultraintelligent Machine' | 1965 | First statement of the intelligence explosion: a machine that can design better machines leaves human intelligence far behind | the core mechanism |
| Bostrom, Superintelligence | 2014 | Set out slow, moderate and fast takeoff and the idea of a decisive strategic advantage for the first project to cross the threshold | takeoff speed as the key parameter |
| Aschenbrenner, Situational Awareness | June 2024 | Argued that automated AI research could compress a decade of algorithmic progress into about a year, and framed the race as a national security project | speed; state involvement |
| Amodei, Machines of Loving Grace | October 2024 | Described 'a country of geniuses in a datacenter' compressing 50 to 100 years of biological progress into five to ten | the upside case |
| Forethought, Preparing for the Intelligence Explosion | March 2025 | MacAskill and Moorhouse argued that AI could drive a century of technological progress in a decade, raising many grand challenges at once | breadth of consequences |
| AI 2027 | April 2025 | Kokotajlo and co-authors published a month-by-month scenario in which a superhuman coder leads to superintelligence within about a year, with race and slowdown endings | a concrete takeoff path |
| METR time-horizon measurements | March 2025 to April 2026 | The length of tasks AI agents complete doubled about every seven months, then faster; in April 2026 laboratories reported no dramatic speed-up of their own research | leading indicator and counter-evidence |
| Narayanan and Kapoor, AI as Normal Technology | April 2025 | Argued that diffusion is limited by institutions and that transformation will take decades | counter-anchor on speed |
Shown at their preset values. Parameters are not adjustable in this release and nothing on this page is computed from them. Custom settings run (Phase B) but are labelled 'non-standard run' and excluded from comparisons.
| Parameter | Default | Range or options | Note |
|---|---|---|---|
| Research speed-up at peak | ×10 | ×3 to ×100 | — |
| Time from automated researcher to superintelligence | 12 months | 3–60 | — |
| Leading projects | 3 | 1–10 | — |
| Control of the leading systems | adequate | options: contested / lost | Lost at Extreme |
| Access for nations that host no project | paid API | options: open weights / restricted | — |
| Binding constraint | compute and power | options: data / regulation / none | — |
High: capability band shifts expected under current plans. Medium: band shifts under 'none' policy response only. Low: strain without band shift. Loads are judgement-based until the Atlas connects. Domains link to the Atlas.
| Domain | Load | Channel |
|---|---|---|
| Hard | ||
| Defence and security | High | decisive advantage for the leading projects; verification gap |
| Strategic infrastructure | High | data centres, power and grid become the binding constraint |
| Critical technology | High | frontier models, chips and compute decide who hosts the takeoff |
| Soft | ||
| Government effectiveness | High | state adoption speed, liability and control rules, tax redesign |
| Human capital | High | knowledge work repriced; education and retraining rebuilt |
| Influence and cohesion | High | distributional conflict; influence over standards and treaties |
| Economic | ||
| Macro-financial | High | growth surge, profit migration, tax base and asset repricing |
| Industry, trade and supply | High | services exports repriced; research-led industries leap |
| Energy and resources | High | electricity demand surge; minerals for compute |
| Region | Exposure | Rationale |
|---|---|---|
| North America | High | Host of most leading projects; concentration of gains and of risk |
| Europe | High | Dependence without a frontier project; regulation as its main lever |
| China | High | The other host; domestic chips and power decide its pace |
| Indo-Pacific | High | Chip manufacturing leverage in Taiwan, Korea and Japan; services exporters repriced |
| South Asia | High | IT services and outsourcing repriced; large gains from cheap expertise |
| Gulf and Middle East | Medium | Energy and capital to host compute; a seat bought, not built |
| Africa | Medium | Largest gains from health and agricultural science; least bargaining power |
| Latin America and Caribbean | Medium | Minerals and power as bargaining chips; services exposed |
| Russia and Eurasia | Medium | Cut off from frontier chips; military implications dominate |
| Indicator | Source | Threshold |
|---|---|---|
| Length of tasks AI agents complete autonomously | METR | — |
| Share of frontier laboratory research and code produced by AI systems | company disclosures | — |
| Frontier compute and data-centre power under construction | Epoch AI, IEA | — |
| Price of inference at fixed capability | — | — |
| Forecasts for automated AI research | Metaculus | — |
| Export controls on chips and model weights | BIS | — |
| AI safety notification and verification agreements | — | — |
From the GINC 250: trends rated Very high or Critical for this scenario. All S15 trend scores.
| Rating | Band or level | Why |
|---|---|---|
| Likelihood, two years | 3Possible | Frontier laboratories have set dates for automated research: an 'intern' in 2026 and a full AI researcher by 2028. Measured task horizons have been doubling every few months. Yet as of April 2026 no laboratory reported a doubling of overall research progress, and sceptics put the feedback loop well short of self-sustaining. |
| Likelihood, ten years | 4Likely | Published forecasts cluster the start of an intelligence explosion between 2027 and the early 2030s; expert surveys put it far later. Band 4 splits the difference. |
| Systemic impact | 5Transformative | Level 5 on irreversible structural change rather than loss: output may rise sharply while the distribution of power and income is rewritten. |
| National impact | 5Transformative | For a nation with no frontier capability, the tax base, labour market and strategic position all shift at once. |
| Confidence | very low | The central mechanism has never been observed and informed forecasters disagree by decades. |
Source of scores: ginc-desk-v0.3. Confidence refers to the rating, not the scenario. Calibration sources are listed with the anchors above and on the methodology page.
Contested assumptions for the panel to resolve.
No signed commentary in this build.
0.3.0 · active0.3.0 · 3 October 2026 · Entered the Library at v0.3 as an upside scenario, with GINC desk scores.GINC (2027). Scenario S15 Superintelligence takeoff, Scenario Library v0.3. scenarios.ginc.org/library/superintelligence-takeoffContent and data are published under CC BY 4.0.