Ask when machines will match human intelligence and you will get a date. Ask again a year later and you will usually get a different one. The movement is more informative than any single answer.
Three kinds of forecast
- 2028Metaculus community estimate for weakly general AI, about 1,700 forecasters (snapshot, February 2026)
- 2033Metaculus community estimate for a general AI system, about 1,800 forecasters (snapshot, February 2026)
- 2047even-odds date for human-level AI in the largest researcher survey, 13 years sooner than the year before (AI Impacts)
Forecasting communities, which are scored on their track record, put a weak form of general AI in early 2028 and a full general system in 2033. Researchers surveyed by AI Impacts put even odds on human-level AI by 2047. That is later, but the same survey moved thirteen years earlier in a single round.
The direction of travel has been toward sooner.
Not every revision goes one way
Honest forecasters also move the other way. The lead author of AI 2027, one of the most discussed scenarios of the last two years, moved his own estimate from around 2027 to around 2030 during 2026. That is not a failure. It is what taking your own forecast seriously looks like.
The idea itself is older than any of these numbers. In 1965 I. J. Good wrote that the first ultraintelligent machine would be the last invention humanity need ever make, provided it could be kept under control. In 1993 Vernor Vinge, who gave the singularity its name, put it at 2005 plus or minus twenty years, a window that ran to 2025. We are living just past its far edge.
“We are past the event horizon; the takeoff has started.”
Why we keep score
Dates without a record are just opinions. That is why one of the Foundation's four programmes is to forecast in public: dated, probabilistic forecasts with written resolution criteria, scored every year, including when we are wrong. Our optimism should be something you can check, not just something you can read.