Ask five AI researchers for an AGI arrival date and you will get five answers that barely overlap. One says it is already here and we simply moved the goalposts. Another points at 2027. A third says 2045, a fourth thinks the question is malformed, and the fifth says 2029 with enough confidence that you start wondering if they know something you don’t.
That spread is not proof the field is clueless. It is what happens when a genuinely hard forecasting problem collides with a definition nobody can agree on. So before arguing about a year, it helps to separate the question into pieces that can actually be answered.
Three definitions, three very different dates
The label “AGI” gets applied to at least three distinct things, and they do not arrive together. The dates attached to them differ by fifteen years or more, which explains a lot of the shouting.
- Narrow AGI: a system that matches a median human on most cognitive tasks you could name, including novel ones. No single benchmark measures this, which is exactly why it is so contested.
- Economic AGI: a system that can perform most remote work at or below human cost. This is the version executives and labour economists care about, and it is measurable in wages and headcount.
- Transformative AGI: a system that outperforms top experts across fields and can meaningfully accelerate its own research. This is the version in the safety papers.
A lab can plausibly claim the first within a few years and still be two decades from the third. When someone tells you AGI is three years away, the useful follow-up is: three years from which of those?
The short end of the range: 2026 to 2030
The optimists have specifics on their side, which is why they get quoted so much. Ray Kurzweil has held 2029 for years and has not blinked. Demis Hassabis has put roughly a 50% chance on something he would call AGI by 2030, while noting the remaining research problems are the hard half. Elon Musk has repeatedly named years that have already passed, which tells you something about the reliability of confident public dates.
Dario Amodei has sketched “powerful AI” landing in 2026 or 2027, defined loosely as a country of geniuses in a datacentre. Shane Legg, who has been forecasting this longer than almost anyone, has pointed at 2028. The people making these calls are not cranks; they run the labs with the most compute and the most visibility into what is coming. You can see the full spread of near-term wagers in these 2026 timeline predictions and what would prove each one right.
What the short-timers are actually assuming
Nearly every 2027 forecast rests on the same bet: that scaling plus reinforcement learning on verifiable tasks plus agentic scaffolding keeps compounding. If that curve bends as it has for the last four years, human-level performance on most knowledge work shows up before the decade turns. If it flattens, the dates slide with it.
The middle: 2030 to 2045
This is where most working researchers sit, and it usually reflects respect for the messy parts. Data is getting scarce. Energy and chip supply are becoming physical constraints rather than budget line items. Robotic manipulation remains stubbornly hard, which means an AI that writes code brilliantly may still not stack a dishwasher. And reliability at 99.9% is a completely different engineering problem from impressiveness at 90%.
The middle camp tends to argue that the last stretch of human parity is the longest, because the remaining gaps are the ones we cannot articulate well enough to train against. There is a detailed version of this argument in the case for AGI by 2040 and the cracks in it, including the compute and cost arithmetic behind the bet.
The long end: 2050, or never
Yann LeCun spent years arguing that large language models are a dead end for real intelligence, and he has put his own timeline well beyond the current decade. Plenty of cognitive scientists go further and claim the whole framing is wrong, that intelligence is embodied, social and developmental in ways scale cannot shortcut.
The strongest version of this position is not “AI is overhyped.” It is that we have no validated path from next-token prediction to systems that build and test their own world models, and until one appears, extrapolating benchmark scores is astrology with better charts.
Why every forecast keeps sliding
Watch the goalposts. In 2021, beating human experts on MMLU was treated as a meaningful milestone. By 2024, frontier models were scoring above 88% and the benchmark was retired from serious discussion. GSM8K went the same way. Each saturation forced the field to invent a harder test, and each new test initially looked unbeatable.
There is a second, stranger pattern. Capability progress does not arrive in the order experts predict. Skills we assumed were hard, like passing bar exams or writing functional code, landed early. Skills we assumed were trivial, like reliably booking a flight without hallucinating a confirmation number, are still unreliable. In other words, what changed was the order in which abilities showed up, not the overall difficulty. That makes any forecast built on a ladder of difficulty shaky from the start.
Underneath all of it sits a physical trend line. Pretraining compute for frontier models has grown roughly four to five times per year for most of the past decade. Training runs that cost tens of millions of dollars in 2023 cost hundreds of millions by 2025. Growth at that rate hits power grids and fabrication capacity before it hits any theoretical limit, and those constraints have lead times measured in years.
What would actually count as evidence
If a lab announced AGI tomorrow, what would convince a skeptic? Not a demo, and not a benchmark score. The credible signals look more like this: a system that takes a novel research question, designs its own experiments, runs them and produces a result that survives expert review. Or one that holds a job for eighteen months with performance reviews indistinguishable from a competent human hire. Or a measurable jump in national productivity growth that cannot be explained by anything else.
None of those are quick to produce, which is precisely why they are hard to fake. A fuller breakdown of which evidence already exists and which does not is in what the evidence actually suggests about AGI happening soon.
Signals worth tracking instead of a single year
Dates are brittle. These indicators are not, and you can watch them in public:
- Cost per unit of capability. If the price of a fixed level of reasoning falls by an order of magnitude per year, deployment spreads fast regardless of what any lab announces.
- Agent task horizon. The length of task, measured in human working hours, that a model can complete unaided has been doubling roughly every seven months on published evaluations. Watch where that curve sits in 2027.
- Self-directed research output. Papers or results generated with minimal human steering that hold up under peer review.
- Hiring patterns in white-collar roles. Junior analyst and paralegal headcount is a more honest indicator than any keynote.
- Energy contracts. Multi-gigawatt power deals quietly signed by data centre operators tell you what the buyers expect to need.
How to plan when you cannot pin the year
Treat the AGI arrival date as a distribution, not a deadline. Almost every serious forecaster, including the aggressive ones, is describing a period somewhere between 2028 and 2040 for systems that would genuinely change how work gets done. If you are building a business, a career or a policy, that window is wide enough to act inside without betting everything on a single year.
The practical move is to identify what gets cheap and abundant first. Reasoning, drafting, coding and analysis are already close to free at the margin; physical manipulation, accountability and trust are not. Anyone whose work is mostly the first category should be redesigning it now, and anyone whose work is mostly the second has more runway than the headlines suggest.
And treat every confident date you read, from a lab CEO or a sceptical professor, as a claim about a definition rather than a prophecy. The year matters far less than which version of AGI the speaker has in mind.