When could superintelligence arrive? What the forecasts say

Forecasts for human-level or superhuman AI range from the late 2020s to the next century. AI lab leaders tend to give the earliest dates. The largest survey, of 2,778 AI researchers in 2023, put a 50% chance on machines outperforming humans at every task by 2047. Much of the disagreement comes from different definitions rather than different evidence.

Why is it so hard to say when superintelligence might arrive?

Superintelligence timelines are hard to pin down because forecasters do not agree on the finish line, and no one has crossed it.

A forecast is only as precise as the thing it forecasts. “Superintelligence” in Nick Bostrom’s sense means an intellect far beyond humans in virtually all domains, as the definition guide explains. Most forecasts are not about that. They are about AGI, human-level AI, or a research milestone, each defined differently. The AGI vs superintelligence guide sets out the definitions.

Before comparing two dates, compare what each date is for.

What have AI lab leaders said?

The dates lab leaders have given mostly fall between 2026 and the early 2030s. These are public statements, listed by date, from people who run companies building the systems.

• May 2023, OpenAI. “Governance of superintelligence” said it was conceivable that within ten years, AI systems would exceed expert skill level in most domains.
• July 2023, OpenAI. “Introducing Superalignment” said superintelligence “could arrive this decade.”
• October 2024, Anthropic. Dario Amodei wrote in “Machines of Loving Grace” that powerful AI, his term for a model smarter than a Nobel Prize winner across most fields, could come as early as 2026, though it could also take much longer.
• March 2025, Anthropic. In a submission to the White House Office of Science and Technology Policy, the company said it expected powerful AI systems to emerge in late 2026 or early 2027.
• April 2025, Google DeepMind. Its post on AGI safety said AGI “could be here within the coming years.” The accompanying paper says its authors are “highly uncertain” about timelines but find it “plausible” that powerful AI systems will be developed by 2030.
• June 2025, OpenAI. Sam Altman wrote in “The Gentle Singularity” that 2026 would likely see systems that can figure out novel insights.
• July 2025, Meta. Mark Zuckerberg wrote that “developing superintelligence is now in sight.” He gave no date.
• October 2025, OpenAI. In a livestream, Altman set targets of an intern-level AI research assistant by September 2026 and an automated AI researcher by March 2028.
• November 2025, OpenAI. “AI Progress and Recommendations” said it expected AI capable of very small discoveries in 2026 and more significant discoveries in 2028 and beyond.
• September 2026, OpenAI. The company said it had met the research-intern target. The milestone was defined and assessed by OpenAI itself.

Two cautions apply to every item on this list. First, these companies can benefit when investors and recruits believe the technology is close. Second, several of them also argue that the technology is dangerous, which cuts the other way. Neither motive makes a forecast right or wrong. Both are reasons to read the definition attached to it.

What have other prominent forecasters said?

Three other forecasts are widely cited: Ray Kurzweil’s long-standing dates, Leopold Aschenbrenner’s essay series and the “AI 2027” scenario.

Ray Kurzweil

The inventor and author has predicted AGI by 2029 since 1999 and the singularity in 2045. He kept both dates in The Singularity Is Nearer (June 2024).

“Situational Awareness” (June 2024)

Leopold Aschenbrenner, a former OpenAI researcher, argued in an essay series that AGI by 2027 is plausible and that automated AI research could then produce superintelligence quickly. He reasoned from trends in computing power and algorithmic efficiency.

“AI 2027” (April 2025)

Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean published a detailed scenario in which superhuman AI arrives around 2027. The authors later added notes on the site. They said 2027 had been their single most likely year, not their median, and that their medians were later. A December 2025 note gives Lifland’s median for the arrival of a “superhuman coder” as 2030.

The scenario is named for 2027. Its own authors later placed their medians after it.

What do surveys of AI researchers say?

Surveys of AI researchers give later dates than most lab leaders. The largest put a 50% chance on human-level machine intelligence by 2047.

AI Impacts, 2023

The largest survey to date was run in October 2023 by AI Impacts and published as “Thousands of AI Authors on the Future of AI” (Grace et al.). It drew 2,778 respondents who had published in top AI venues, a 15% response rate.

• High-level machine intelligence (unaided machines doing every task better and more cheaply than human workers): 10% chance by 2027 and 50% by 2047.
• Full automation of labor (every occupation fully automatable): 10% by 2037 and 50% by 2116.
• Superintelligence after HLMI: a median 10% chance within two years and 60% within thirty years.
• Change from the year before: the 2047 estimate was 13 years earlier than in the 2022 survey.

The two milestones describe nearly the same world, yet the 50% dates are 69 years apart. The main difference is how the question was worded, which makes the gap a clear sign of how much framing moves these numbers.

Müller and Bostrom, 2012–2013

Vincent Müller and Nick Bostrom surveyed four groups of experts. The median respondent gave even odds of high-level machine intelligence around 2040 to 2050 and a nine-in-ten chance by 2075. Respondents expected superintelligence less than 30 years after that. They put the chance that this development would turn out bad or extremely bad for humanity at about one in three.

What do forecasting platforms say?

Metaculus, a public forecasting site, runs long-standing questions on when the first general AI system will be built, tested and announced. The answers depend heavily on the resolution criteria.

• A question about a “weakly general” AI, tested against a set of specific benchmarks, showed a community median of mid-2028 in an archived copy from late June 2026.
• A stricter question about a general AI system showed a median of late 2032 in an archived copy from early June 2026.

These numbers change daily. Check the live questions before citing them.

Are there forecasts for superintelligence itself?

Only a few forecasts name superintelligence directly. Most target AGI or human-level AI and treat superintelligence as a later step.

The ones that name superintelligence directly are OpenAI’s 2023 posts, the follow-on questions in the Müller and Bostrom and AI Impacts surveys, and the “AI 2027” scenario. The Future of Life Institute’s Statement on Superintelligence also makes a timing claim in its context note: that many leading AI companies aim to build superintelligence within the coming decade.

So a headline that says “superintelligence by 2030” is usually reporting an AGI forecast plus an assumption about what comes next.

Why do the forecasts disagree so much?

1. Different targets. “Every task,” “most economically valuable work,” “smarter than a Nobel Prize winner” and “an automated AI researcher” are four different milestones.
2. Different methods. Lab forecasts often extrapolate recent trends in computing power and benchmark scores. Surveys aggregate the judgment of many researchers. Scenario writers build step-by-step stories. Each method has different blind spots.
3. Different positions. People inside frontier labs see unreleased systems. They may also have commercial reasons to sound confident. Academic researchers see less of the frontier and have fewer such incentives.
4. Different views on the last step. Some expect superintelligence soon after human-level AI, through an intelligence explosion. Others expect diminishing returns and slow adoption. Arvind Narayanan and Sayash Kapoor argue that AI’s effects will unfold over decades.
5. Earlier dates have passed. Vernor Vinge wrote in 1993 that superhuman intelligence would be technically possible within thirty years. That date passed in 2023 without a superhuman general intelligence being built.

The policy responses to these forecasts, including proposed bans, are in the debate guide.