Ask a dozen AI researchers when artificial general intelligence arrives and you’ll get dates spread across four decades. Ray Kurzweil has held to 2029 for years. Demis Hassabis has talked about five to ten years. Yann LeCun argues the current approach won’t get there at all. That spread isn’t evidence that nobody knows anything. It’s a map of what each person thinks intelligence is, how it gets built, and how you’d recognise it once it existed.
The definition problem nobody escapes
Every AGI prediction smuggles in a definition, and the definition usually does more work than the forecast. If AGI means “beats the median human at every economically valuable task,” you get one date. If it means “can do everything a person can do, including learn a new physical skill from one demonstration,” you get a much later one. If it means “matches human reasoning but not embodiment,” you get something earlier still.
That ambiguity is why the phrase stays so slippery in public discussion. A plain-language walkthrough of what AGI actually means and how close we really are is worth reading before you take any specific date seriously, because half the apparent disagreement between experts dissolves once you check whether they’re arguing about the same thing.
What the aggregated numbers actually say
The most-quoted data point comes from a survey of around 2,700 AI researchers, published in 2024. The aggregate forecast put a 50% chance of “high-level machine intelligence” by 2047, with a 10% chance by 2029. When the same survey ran five years earlier, the 50% mark sat around 2060. Forecasts moved toward the present by more than a decade in half a decade.
Prediction markets show the same drift with more noise. Metaculus community estimates for weak AGI have hovered in the early 2030s, jumped forward after each big model release, then drifted back again. That pattern is worth noticing: public timelines respond to press cycles, not only to evidence.
The short end of the range is where things get loud. “AGI by 2026” has become as much a marketing position as a forecast, and it pays to separate the people making a genuine bet from the people making a slide. There’s a solid breakdown of who’s saying what in the 2026 AGI predictions, and which of those specific calls have held up under scrutiny.
The three camps, roughly
- Fast (2027 to 2032). Kurzweil, Shane Legg, Leopold Aschenbrenner, Ben Goertzel, and the writers behind the AI 2027 scenario. Their shared premise: scaling plus reinforcement learning on verifiable tasks is enough, and the remaining gaps close quickly once models start doing AI research themselves.
- Middle (2035 to 2050). Where most survey respondents land. Hassabis has nudged from “a decade” toward five to ten years; Hinton’s public range has run from five to twenty. This camp tends to assume a few more architectural breakthroughs are needed, but expects them.
- Skeptics. LeCun, Gary Marcus, Melanie Mitchell, François Chollet. The argument isn’t that machines can’t be intelligent. It’s that next-token prediction over text is a dead end for world models, persistent memory, and causal reasoning. The strongest version of that case sits inside the real arguments for and against human-level AI.
What actually moves a serious forecast
Long-horizon task completion
The most useful single metric right now comes from METR, which measures how long a task a model can finish with 50% reliability. That time horizon has roughly doubled every seven months since 2019. Extrapolate the curve and you reach full workdays within a few years. The catch is that extrapolating through a doubling curve has burned plenty of forecasters before.
Benchmark saturation and what replaces it
Every benchmark gets solved faster than the last. MMLU went from frontier result to routine in about three years. GPQA and SWE-Bench are heading the same way. When a benchmark stops discriminating, forecasters quietly switch to a harder one, which makes year-over-year comparison harder than the charts suggest.
Compute, capital, and the data wall
Training runs have grown by orders of magnitude, funded by capex figures that would have looked absurd in 2020. Whether that continues depends on power, chips, and whether returns on the next 10x hold. Several forecasters have already pushed timelines later purely on energy constraints rather than on anything to do with model capability.
How to read the next prediction that lands
A new AGI timeline usually arrives with a keynote or a podcast appearance attached. Five questions sort the substance from the noise:
- What’s the resolution criterion? If nobody can say in advance what would count as AGI, the prediction can’t be wrong, which also means it can’t be useful.
- Is it a date or a probability? “AGI by 2030” is a coin flip dressed as a statement. “70% by 2032” is a forecast you can score.
- What would change their mind? Someone who can name the evidence that would falsify them is usually reasoning. Someone who can’t is usually performing.
- Who benefits from the belief? Fundraising, hiring, and share prices all respond to timelines.
- What’s their track record? A handful of people have made public, checkable calls for a decade. Most haven’t.
The same instincts apply to product launches that get framed as milestones. There’s a practical checklist for telling a real AGI milestone from a marketing beat, and most of it comes down to whether the claim can be tested by someone outside the company.
What’s more useful to track than a date
Dates are seductive because they’re easy to argue about. Capabilities are harder to spin. This roundup of AGI technology trends worth actually tracking goes deeper, but the short version is four signals:
- Whether agentic systems can hold a task together for hours without a human rescuing them.
- Whether models get better at learning new skills from small amounts of in-context data, rather than from more pretraining.
- Whether robotics shows the kind of curve language models hit after 2020.
- Whether AI doing AI research produces compounding returns, since that assumption sits underneath almost every fast timeline.
Plan for a range instead of a date. The organisations that get this right won’t be the ones that called 2029 or 2045 correctly. They’ll be the ones that built systems, teams, and decision processes that hold up whether the answer is five years away or thirty. A prediction you can’t act on is just a conversation. The useful version is knowing which specific capabilities would force a change in what you’re doing, and watching for those directly instead of waiting for an announcement.