When Will Superintelligence Arrive? What the Forecasts Say

Nobody knows when superintelligence will arrive, or whether it will. Estimates range from a few years to many decades, and some people think it will never be built. The largest recent survey of AI researchers, published in 2024 by Katja Grace and colleagues with 2,778 participants, produced an aggregate estimate of a 50% chance that machines outperform humans at every task by 2047, with huge variation across answers. That is about human level AI, not superintelligence, and it is a collective opinion, not a measurement. For a company, the date matters less than being ready to use each improvement with control.

Key takeaways

• There is no reliable date for superintelligence. Any announced year is a bet.
• The Grace et al. (2024) survey estimated, in aggregate, a 50% chance of AI outperforming humans at every task by 2047, and 10% by 2027. Individual answers varied widely.
• Lab leaders tend to give much shorter timelines than the average researcher. There are commercial and fundraising incentives behind those statements.
• AI forecasts have been wrong in both directions: overpromises in the 1960s and recent surprises with language models.
• The useful decision in a company does not depend on the date: a mapped process, limited permissions, evaluation on real cases and a named owner.

What do surveys of AI researchers say?

The most cited reference is the AI Impacts survey series. The edition published in 2024, led by Katja Grace and colleagues, surveyed 2,778 authors of papers at major AI conferences. Asked when unaided machines could perform every task better and more cheaply than humans, the aggregate answers gave a 10% chance by 2027 and 50% by 2047. In the previous 2022 edition, the 50% estimate was thirteen years later, which shows how expectations shortened after language models.

Three cautions when reading these numbers. First, the question is about human level AI at every task, not superintelligence, which would come later. Second, small changes in question wording shifted the answers a lot, a sign of high uncertainty. Third, experts at building models are not necessarily good at forecasting dates.

Why do lab leaders talk about such short timelines?

Executives at some AI labs have spoken publicly about systems far more capable than humans within a few years. These statements deserve attention, because these people see internal results the public does not. They also deserve a discount, because short timelines help attract investment, talent and attention.

AI labs pursuing superintelligence: how to read their claims offers questions to separate evidence from rhetoric: which capability was demonstrated, on which test, at what cost and with what error rate.

Why do forecasts differ so much?

Because they rest on assumptions nobody can confirm today:

• Is scale enough? Some think more data, compute and reasoning time will lead to general systems. Others think new ideas are missing.
• Will there be self-improvement? If AI starts speeding up its own research, timelines shorten. If not, the pace follows industry.
• Physical limits. Chips, energy and data centers take years to build.
• Regulation and society. Laws, accidents or shifts in public opinion can speed things up or slow them down.
• Definition. With no accepted test for superintelligence, each person may be forecasting something different.

History also calls for humility. In 1965, Herbert Simon wrote that within twenty years machines would be capable of doing any work a man can do. It did not happen. On the other hand, few people in 2015 predicted the leap in language models. Does superintelligence already exist? How to check who says yes helps you evaluate anyone who claims the point has already been reached.

What should a company do without knowing the date?

Treat the topic as uncertainty to manage, not a bet. In practice:

• Build two or three pace scenarios (slow, fast, uneven) and see which decisions make sense in all of them. Superintelligence in business scenario planning without dates shows how.
• Start with concrete workflows that current AI already handles with human review, and measure.
• Keep real cases with approved answers to test each new model before switching.
• Review permissions and owners every quarter, because model capability changes at that pace.
• Track verifiable signals, such as results on independent evaluations, instead of announced dates.

These steps hold whether superintelligence arrives in ten years, in fifty or never. How to prepare your company for superintelligence organizes this plan into stages.

Illustrative scenario: the five year plan

Illustrative scenario: a chain of language schools is preparing its five year strategic plan. One board member suggests postponing AI investment "because by 2030 everything will change". Another wants to restructure the whole operation now "before it is too late".

The director proposes another path. The chain picks two workflows with volume and clear rules, enrollment and class rescheduling, and runs pilots with bounded agents and human review. It sets time and error metrics, keeps the test cases and schedules a twice yearly review of the AI plan with three questions: what models do today that they did not do before, what our tests show and which permissions change. The plan no longer depends on a date nobody knows.

How does the map, limit, pilot and hand over method apply here?

The method I use was designed to work without a date forecast. Mapping means describing the process and measuring the baseline. Limiting means defining the agent's data and actions, with human approval where an error is costly. Piloting means running a real workflow with logs and an agreed metric. Handing over means leaving the operation, documentation and evaluation criteria with the team, so the cycle can be repeated with each new model.

I do not sell superintelligence or AGI, and nobody does today. The work is to prepare the company to benefit safely from each real improvement. If you want to talk about one of your company's workflows, see the AI consulting for business page.

Frequently asked questions

In what year will superintelligence arrive? There is no reliable year. The largest recent survey of researchers estimated a 50% chance of human level AI at every task by 2047, and superintelligence would come after that, if it comes.

Could superintelligence arrive before 2030? Some lab leaders suggest short timelines, and the 2024 survey gave about a 10% chance of human level AI by 2027. Most estimates fall later, with great uncertainty.

Is superintelligence inevitable? No. Some think it will never be built, because of technical or physical limits or a decision by society. It is a possible scenario, not a destiny.

Should my company wait before investing in AI? No. Bounded pilots with current AI produce learning about process, data and evaluation that carries over at any pace of progress.

Sources

• Grace et al.: Thousands of AI Authors on the Future of AI (2024)
• AI Impacts: 2023 Expert Survey on Progress in AI
• Nick Bostrom: Superintelligence (Oxford University Press, 2014)
• NIST AI Risk Management Framework