AI Surpassing Humans: Broad Cognitive AI by 2033, Full Supremacy in the 2040s
"When will AI surpass humans?" is three questions wearing one coat, and they have three different answers. The first is already settled. The second is the live question of this decade. The third is further away than most headlines admit.
The settled one is narrow tasks. Chess, Go, image recognition, protein structure prediction, code completion, and now even forecasting tournaments, where AI teams rival human superforecasters — in slice after slice, machines already outperform the human average, and in several cases the human best. Nobody noticed a threshold moment, because there wasn't one. Surpassing, in the narrow sense, arrived the way most thresholds do: one capability at a time, quietly, until the list of remaining human-only tasks got shorter.
The live question is broad cognitive work: a system that can do most knowledge work, learn new domains with little data, and plan over weeks rather than minutes. The evidence for soon is real. Frontier models keep compounding, agents now complete multi-hour software tasks, and lab leaders talk openly about the near term — DeepMind's Demis Hassabis puts AGI two to three years away, Anthropic's Dario Amodei calls 2027 plausible. The evidence against is also real: reliability is still brittle, long-horizon planning fails in ways that are hard to see coming, and benchmarks saturate without the gains always transferring to new domains.
Then there is the strict bar, and it is stricter than most people mean when they say AGI. The canonical forecasting question on Metaculus requires three legs at once: a two-hour adversarial Turing test, general robotic capability demonstrated by assembling a model from instructions, and expert-level breadth across fields. Robotics is the long pole here. Language models run on text; bodies run on physics, and physics does not scale just by adding GPUs. That is why the strict bar trails pure cognition by years, and why anyone quoting a date for "AGI" without saying which bar they mean is selling you a number without a definition.
The forecasts, once you sort them by definition, tell a coherent story. The Metaculus community's median for the strict question sits at January 2033, with roughly a 25% chance by 2029; its cognitive-only "weakly general" question has a much earlier median, June 2028. Aggregating across platforms, forecasters center near 2031, with an 80% interval spanning 2027 to 2044. The 2023 survey of 2,778 AI researchers put even odds for machines outperforming humans at every task in 2047 — a broader bar than the strict one, and a later date to match. Prediction markets add a narrower data point: Polymarket prices OpenAI announcing it has achieved AGI before 2027 at just 11%, a bet on one lab, self-announced, on a near date — consistent with near-term skepticism, but not a price on my question.
My own judgment splits the same way. For broad cognitive AI by the end of 2033, I put the chance at 55 to 60%. For the strict bar — Turing test plus robotics plus breadth — by the same date, it is a coin flip, and that is the prediction I am tracking here, at 50%. I set it no higher because the robotics leg has no public benchmark timeline and the broader expert survey points later; no lower because frontier pace and lab timelines point the other way.
What would move my number? Robotics benchmarks at scale — general assembly, dexterous manipulation, robots that learn new physical tasks from instructions — would pull it earlier. A scaling plateau, whether from data or energy walls, would push it later. And the best public signal to watch is the Metaculus median itself: it moved from around 2070 in 2020 to around 2033 today, and its next moves will tell us more than any single lab leader's claim.
There will be no single day when "AI surpasses humans." It will look like the narrow wins did: one capability at a time, then suddenly the list of things only humans can do is short. The honest position is a date range with a number attached, not a prophecy.