GENEKI
Start a project
  1. Home
  2. Blog
  3. When will superintelligence arrive?

AI

When could superintelligence arrive? It depends who you ask.

Some lab leaders talk about a few years. Most surveyed researchers say decades. Forecasts have also moved. A fair look at the numbers, the doubts and a sensible way to plan when nobody knows.

6 min read Geneki team

On this page
  1. What lab leaders say
  2. What researchers say
  3. Scenarios and measurements
  4. How forecasts have moved
  5. How to plan
  6. Questions people ask
Two bright rocket launch trails arcing across a dark night sky
Photo: SpaceX, Wikimedia Commons (CC0)

The short version

  • Lab leaders talk in single-digit years. Surveyed researchers talk in decades, with a lot of spread.
  • Forecasts have been moving, and the loudest ones come from people with a stake in the answer.
  • Measured progress is fast on some tasks and uneven on others. Nobody has a reliable clock.
  • The sensible plan is scenarios and signals, not a date.

What lab leaders say

The people building frontier AI are the most optimistic about speed. Here is what they wrote, in their own words.

  • Sam Altman (OpenAI), in his essay "The Gentle Singularity", wrote "Humanity is close to building digital superintelligence". He expected that "2026 will likely see the arrival of systems that can figure out novel insights" and that "2027 may see the arrival of robots that can do tasks in the real world". He also wrote that "we do not know how far beyond human-level intelligence we can go, but we are about to find out".
  • OpenAI, in a November 2025 post, said "In 2026, we expect AI to be capable of making very small discoveries. In 2028 and beyond, we are pretty confident we will have systems that can make more significant discoveries".
  • Dario Amodei (Anthropic), in January 2026, wrote that powerful AI "could be as little as 1 to 2 years away, although it could also be considerably further out", and pictured a "country of geniuses in a datacenter" appearing around 2027.
  • Mark Zuckerberg (Meta) wrote in July 2025 that "Developing superintelligence is now in sight".

These are statements of expectation from companies that are also spending heavily on the effort. Meta told investors in July 2026 that it expects 2026 capital expenditure of $130 to $145 billion. OpenAI announced in January 2025 that the Stargate project intends to invest $500 billion over four years in AI infrastructure. Spending shows commitment. It does not show that the goal can be reached on schedule.

Two bright rocket launch trails arcing across a dark night sky
Launch dates slip, even for projects that succeed. Photo: SpaceX, Wikimedia Commons (CC0)

What researchers say

Surveys of researchers give a more cautious picture, with wide spread.

  • AI Impacts survey. Grace and colleagues asked 2,778 AI researchers. They gave a 10% chance that unaided machines outperform humans at every possible task by 2027, and a 50% chance by 2047. For all human occupations being fully automatable, they gave 10% by 2037 and 50% by 2116. The survey ran in 2023. We found no more recent run of it.
  • AAAI panel survey, 2025. In a survey in the AAAI presidential panel's report, 76% of the 475 respondents said that "scaling up current AI approaches" is "unlikely" or "very unlikely" to yield AGI. About a fifth of respondents were students.

Reading these together: many researchers expect big capability gains, but a majority in that survey doubt that simply making today's systems bigger is enough. Those two views are compatible, and they point to different timelines.

Scenarios and measurements

The AI 2027 scenario

In April 2025 a group led by Daniel Kokotajlo published "AI 2027", a detailed scenario in which superhuman coding arrives in March 2027 and superintelligence, defined as "much better than the best human at every cognitive task", in December 2027. It is a story built to make the debate concrete. The authors have since said that 2027 was their "modal" (most likely) year at publication, not their median, and their site records updates that pushed medians later, with one author's median at about 2030.

METR's measurements

The research group METR measures how long a task, in human working time, AI agents can complete about half the time. In March 2025 it reported that this length had been doubling about every seven months. It noted a separate analysis showing an even faster doubling time on one benchmark. METR's later published data suggests a faster rate since 2023 but warns that horizons above 16 hours are unreliable, and its page says it is no longer actively updated. It is a useful measure of one kind of progress: long, well-defined software tasks. It is not a measure of general intelligence.

The International AI Safety Report

The 2026 report, chaired by Yoshua Bengio, says "general-purpose AI capabilities have continued to improve, especially in mathematics, coding, and autonomous operation", that capabilities remain "jagged", and that "the trajectory of AI progress through 2030 is uncertain, but current trends are consistent with continued improvement".

How forecasts have moved

  1. 2023OpenAI writes that within ten years AI systems may exceed expert skill in most domains. The AI Impacts survey runs.
  2. April 2025AI 2027 publishes a scenario with superintelligence in late 2027.
  3. Mid to late 2025The AI 2027 authors move their medians later, saying 2027 was their most likely year, not the median.
  4. November 2025OpenAI expects "very small discoveries" in 2026 and more significant ones from 2028.
  5. January 2026Amodei writes that powerful AI could be 1 to 2 years away, or considerably further out.

A fair caution: people who forecast fast progress are often also raising money, recruiting or campaigning for policy changes, and people who forecast slow progress may be defending a research view or a business. That does not make either wrong. It is a reason to weigh the track record and the evidence, not the confidence.

How to plan when nobody knows

ScenarioWhat it looks likeWhat to do now
Faster than expectedAgents do days of work unsupervised. New capabilities arrive monthly.Keep human oversight and access controls ready. Know which processes you could automate safely.
Steady progressTools keep improving. Jagged strengths and weaknesses remain.Build skills and workflows that get value from today's tools. Re-evaluate every quarter.
A plateauGains slow. Costs matter more than capabilities.Focus on cost control and on projects that pay back without new breakthroughs.
  • Do the useful things nowToday's AI is already worth using for documents, support, code and analysis. See our guide to AI tasks you can hand over today.
  • Track three signalsHow long a task AI can complete without help, how often it fails in your own work and how fast your costs per task are falling.
  • Avoid single-date betsDo not sign contracts or cut teams on the assumption that a specific capability arrives by a specific date.
  • Keep the option to moveChoose tools and vendors you can switch, and keep your data portable.
  • Know the rules moving tooRegulation is changing faster than most forecasts. See how governments see it.
Related guideIs superintelligence dangerous? A risk guide for business leadersThe arguments, the doubts and what to do about today's risks.

Questions people ask

When will superintelligence happen?

Nobody knows. Lab leaders have talked about the next few years for powerful AI, a large survey of researchers put a 50% chance of machines beating humans at every task around 2047, and the International AI Safety Report 2026 says the trajectory through 2030 is uncertain.

What did Dario Amodei say about the timeline?

In his January 2026 essay he wrote that powerful AI 'could be as little as 1 to 2 years away, although it could also be considerably further out', and imagined a 'country of geniuses in a datacenter' appearing somewhere in the world around 2027.

What do AI researchers think?

The largest survey we found, by AI Impacts, asked 2,778 researchers. They gave a 10% chance that machines outperform humans at every task by 2027 and 50% by 2047. A separate 2025 survey in an AAAI report found 76% of 475 respondents thought scaling up current approaches was unlikely or very unlikely to deliver AGI.

Has the AI 2027 forecast been revised?

Yes. The authors say 2027 was their most likely year at publication, not their median, and later posts moved their medians later, for example to around 2030 for one author. Their scenario remains a story told to make the discussion concrete, not a prediction they stand behind date by date.

Should my company plan for superintelligence?

Plan for what AI can do in the next one to three years, which is already large, and track a few signals that would tell you the pace is changing. Avoid bets that only pay off if one specific date is right.

Planning under AI uncertainty?

Geneki helps teams build a practical two-year AI plan: useful now, flexible later and honest about what is not known.

Talk to us

Sources

Forecasts in this article are quoted from the sources shown and are not our predictions. The date of Sam Altman's essay is not shown on the archived page we read and was widely reported as mid-2025. Survey figures come from 2023 data. METR's measurements are described in general terms because the organization says they are no longer actively updated.