Ask a simple question, ‘When will AGI arrive?’, and you get a strange answer from the people who study it for a living: somewhere between the next two years and the next century. That spread is not a sign that researchers are clueless. It is what happens when a fuzzy idea, several definitions, and a handful of forecasting methods get pushed into one number. AGI timelines from researchers are best read as a distribution of bets, not a single countdown clock.
That distribution still tells you something. It shows where the centre of gravity sits, which estimates are outliers, and why a lab CEO can say ‘2027’ while a cognitive scientist says ‘2070’. The same pattern shows up in any careful AGI timeline built from expert forecasts: a wide range, a cluster in the middle, and a lot of argument about the tails.
What researchers actually mean by an AGI timeline
Most serious surveys do not ask, ‘Will AGI arrive in 2027?’ They ask for probabilities against an operational definition. That distinction matters because ‘AGI’ is not one technical milestone. It is a family of claims about generality, autonomy, and human-level performance.
Common definitions include:
- High-level machine intelligence (HLMI): a system that can outperform humans at nearly every economically valuable task.
- Transformative AI: a system powerful enough to reshape the economy or science, even if it is not human-like.
- Human-level AGI: a system that can learn and perform across cognitive domains at roughly human level.
Those definitions are not interchangeable. A model that can do remote knowledge work at a human level may still fail at driving a van in rain, running a lab, or caring for a child. A system that can do all of those is a much harder target.
The survey numbers: a 2047 median and very wide tails
In the most cited recent survey, Grace and colleagues collected responses from 2,778 AI researchers. The aggregate median put a 50% chance of HLMI around 2047. The same survey put a 5% chance by 2028 and a 90% chance by 2110. Those three numbers belong together.
The median is the headline. The tails are the story. A 5% chance by 2028 is small in probability terms, but it is not zero, and it is close enough to matter for anyone building safety cases or capital plans. A 90% chance by 2110 is a reminder that some serious researchers still see AGI as a multi-generational project.
Why the median moved earlier
A previous version of the survey had a median closer to 2060. The 2024 round moved it to 2047, a 13-year shift in one survey cycle. Some of that is genuine technical progress in language models, reasoning, and tool use. Some is definitional drift and recency bias. This is why AGI predictions keep moving, and why the date you remember depends on the survey you read.
Where the early dates come from
The 2027 to 2030 camp is small but loud. Dario Amodei has said powerful AI could arrive around 2026 or 2027. Sam Altman has framed AGI as a few thousand days away, a much nearer horizon than the survey median. Demis Hassabis has described AGI as a decade-scale project, with 50% odds by 2030 in some public remarks. These are not survey medians. They are lab leaders with capital, compute, and a reason to be quoted.
Scaling and the language-model shock
Their case is straightforward. Language models improved faster than many academics expected, and the recipe of more compute, more data, and reinforcement learning from human feedback produced systems that can write code, reason through problems, and use tools. If that curve continues, the gap between an impressive demo and a general worker could close faster than surveys assume. The case for AGI by 2040 draws much of its force from this camp, even when the lab leaders themselves name earlier years.
Where the later dates come from
The 2050-plus camp points to capabilities that do not obviously fall out of next-token prediction. Yann LeCun has argued that current LLMs lack the world models needed for robust general intelligence and that scaling alone will not get there. Geoffrey Hinton has given a range of roughly 5 to 20 years, which sounds early only if you anchor on 2027 rhetoric. Other researchers point to embodiment, causality, and the sheer difficulty of building systems that know when they are wrong.
The missing pieces
Three bottlenecks come up again and again in technical conversations:
- Continual learning: systems that can update their knowledge without retraining from scratch or forgetting earlier skills.
- Sample efficiency: learning from a handful of examples rather than a mountain of data.
- Long-horizon reliability: completing multi-step tasks without drifting, inventing facts, or needing a human to catch errors.
Robotics adds another layer. A system that can plan a meal and a system that can cook it in an unfamiliar kitchen are not the same achievement. The second requires perception, manipulation, and recovery from mistakes, all in real time.
Why expert surveys are not prophecy
Surveys are useful snapshots, not forecasts with error bars. A researcher answering a probability question in 2024 is not running a model. They are expressing a judgement with a number attached. The wording of the question can shift the median by years.
There are other traps:
- Selection bias: AI researchers are not a representative sample of scientists, and the people who answer surveys may differ from the people who do not.
- Incentives: lab leaders who raise money, hire talent, or sell a vision have reasons to sound early. Researchers who depend on public grants may have reasons to sound cautious.
- Definition drift: if ‘AGI’ means a chatbot with broad knowledge in one survey and a fully autonomous agent in another, the dates are not comparable.
- False precision: a median year looks exact, but it can hide a distribution that is wide enough to make planning on that year reckless.
A good AGI forecast separates the definition, the method, and the motive. That is more useful than memorising a single date.
The range is the signal
If you strip away the headlines, the researcher consensus is not ‘2027’ or ‘2100’. It is a distribution with a median in the 2040s, a meaningful left tail in the late 2020s, and a long right tail beyond 2100. That shape is useful for planning. It argues for investing in evaluations, redundancy, and model-agnostic infrastructure now, while not betting the firm on a 2027 arrival. It argues for building oversight before the capability is obvious. It argues for learning to work with these systems without assuming they will be fully autonomous next year.
How to read the next round of AGI timelines
When a new survey or lab blog post lands, ask four questions:
- Which definition is being used?
- Is the date a median, a threshold, or a vibe?
- Who was surveyed, and who was not?
- What would change the forecaster’s mind?
This is the discipline behind the AGI countdown: not watching a clock, but watching the capabilities that make the clock meaningful.
What to watch instead of a single date
AGI timelines from researchers will keep moving. The better move is to track indicators that are harder to fudge:
- Cost of frontier-level reasoning: how quickly the price of strong reasoning falls at fixed capability.
- Long-horizon task success: the length of tasks agents can complete without human rescue.
- Learning after deployment: whether systems improve from feedback without catastrophic forgetting.
- Evaluation quality: whether benchmarks still measure general capability rather than memorised test sets.
- Economic diffusion: how many real workflows change, not how many demos go viral.
None of these gives you a clean date. Together they tell you whether the 2027 camp is being vindicated or the 2050 camp is being patient. A team that can point to progress on long-horizon reliability, sample efficiency, and evaluation quality will learn more from the next two years than from another round of duelling date predictions.