AGI guide · Intermediate · 10 min
When will AGI arrive? Published forecasts, their methods, and why they differ
No AGI timeline has won broad agreement. Compare published forecasts, their definitions, methods, and the uncertainties behind their widely differing estimates.
There is no consensus on when AGI, or artificial general intelligence, will arrive among the sources reviewed as of October 24, 2026. Forecasts differ because they target different capabilities and draw on surveys, extrapolations, scenarios, or research judgments. To compare them, read the predicted event, its probability, its conditions, and the estimate’s publication date together.
Define the event before comparing dates
A forecast becomes meaningful when it specifies what needs to happen. Surpassing humans on academic tests, learning an unfamiliar occupation, and conducting independent research are distinct goals. A date without a success criterion leaves readers to supply their own definition.
In particular, distinguish performance, generalization, and autonomy. Performance describes a result on a task. Generalization concerns the transfer of capabilities to unfamiliar situations; autonomy adds the ability to pursue a goal with limited supervision.
Economic usefulness introduces further requirements. A system may complete a task successfully while requiring too much computation, checking, or intervention to replace an existing operation. Conversely, specialized automation can transform an activity without meeting a demanding definition of general intelligence.
Before relying on a timeline, ask:
- Which tasks are included, and which are excluded?
- What level of human ability is the comparison based on?
- What prior learning and assistance are allowed?
- What costs, completion times, and error rates are acceptable?
- Who determines whether the threshold has been crossed?
These questions complement our explanation of how AGI is defined. Above all, they prevent events that share a name but represent different thresholds from being placed on the same timeline as though they were equivalent.
Published forecasts target different outcomes
The following table brings together identifiable publications. The dates belong to the original estimates; they are not a schedule endorsed by this site.
| Publication | Target event or capability | Published timeline | Method and main limitation |
|---|---|---|---|
| Survey by Grace and coauthors, initially published in January 2024 | Unaided machines outperforming humans in every possible task | Aggregate probability of 10% by 2027 and 50% by 2047 | Survey of researchers; aggregation of opinions |
| Dario Amodei, Machines of Loving Grace | AI far more capable across many fields, able to work autonomously | A possibility as early as 2026 | Reasoned judgment; no guaranteed deadline |
| AI 2027, technical forecasts from April 2025 | Internal development of a superhuman programmer | Initial all-things-considered medians: 2028, 2030, and 2033, depending on the forecaster | Models and judgmental adjustments; a target focused on coding |
| State of AI Report 2026 | AGI presented as an editorial prediction | 2027 | A forward-looking position; its operational definition needs examination |
The survey by Grace and coauthors, initially published in January 2024, includes 2,778 researchers. Assuming that science continues without interruption, its aggregate estimate reaches 50% in 2047 for the general capability under consideration. A separate question about the full automation of all occupations reaches that threshold in 2116.
In Machines of Loving Grace, dated October 2024, Dario Amodei considers powerful AI possible as early as 2026, while acknowledging that it could take much longer. He describes broad intellectual capabilities and autonomy on extended projects. This is a reasoned possibility, without a numerical probability comparable to the survey’s estimate.
The State of AI Report 2026, published in October 2026, includes a prediction of AGI in 2027. That prediction alone does not establish that its threshold is equivalent to those used in the other publications.
Surveys measure expectations, not a countdown
A survey describes the distribution of expectations within a community. Its value depends on how questions are worded, how respondents are selected, and how transparently the answers are combined. Surveys can also track changing expectations when their methods remain comparable.
But expertise in machine learning does not directly measure the time remaining until a discovery. A researcher may understand particular architectures in detail while remaining uncertain about industrial resources, robotics, or how work is organized. Several knowledgeable people can share the same blind spot.
The wording of a question matters as much as the number of responses. An intellectual task performed in a digital environment faces different constraints from the complete automation of an occupation. An occupation may involve physical presence, responsibility, and interactions that are difficult to formalize.
It is also necessary to distinguish an aggregate estimate from agreement among respondents. A collective probability curve can be calculated from widely scattered opinions. Its median does not mean that every researcher personally places the transition near that date.
When reading a survey, look for the original question and the conditions attached to the answers. Check whether the result being presented is a cumulative probability, an average of dates, or another statistic. These summaries are not interchangeable.
Extrapolation needs a justified link between the measure and the goal
Extrapolation takes an observed trend and extends it under specified assumptions. It makes the reasoning open to examination: readers can question the slope, the period studied, and the threshold to be reached. It does not demonstrate that a trend will continue beyond the area measured.
METR’s task time horizon provides a useful example. This horizon expresses how long a human would take to complete tasks on which an agent achieves a given success rate. It therefore does not simply measure how long a program remains active.
In its January 29, 2026 update, METR reports a historical doubling time of approximately seven months on its hybrid series, and 131 days for the period since 2023 using its revised method. The organization emphasizes that the results are sensitive to the mix of tasks and that the intervals remain wide.
These measurements provide evidence of progress under experimental conditions. To forecast AGI, the next step is to justify the connection between success on those tasks and the broader capabilities being predicted. This link between the measurement and the goal is often the most uncertain part of the reasoning.
Reliability also changes the interpretation. Success on half of all attempts may be useful for investigating a capability, yet insufficient for handing over an activity without oversight. An extrapolation needs to specify the success rate it extends, along with the cost of failures.
Scenarios make assumptions visible
A scenario describes a possible sequence of events. It can show how technical advances, corporate decisions, and political responses might interact. Its detail makes the proposed mechanisms easier to discuss; it does not automatically make the complete narrative highly probable.
The technical forecasts accompanying AI 2027 distinguish, among other approaches, an extrapolation of task time horizons and a method that starts from a research benchmark, followed by an estimate of the remaining distance. Their target is a superhuman programmer within a company developing AI. The initial all-things-considered estimates from April 2025 had different medians, despite sharing the scenario’s title.
This distinction helps avoid treating the year in a narrative as its authors’ single median forecast. Read the probability distributions and technical documents that accompany the story.
The AI Futures Project update of August 16, 2026 adds approaches based on programming productivity gains and revenue. The authors highlight uncertainty in the relationship between task time horizons and coding automation. They also specify that their timelines assume progress proceeds as quickly as technically possible.
To assess a scenario, separate its decisive transitions: coding assistance, automation of work, acceleration of research, and then broader capabilities. Each transition needs its own evidence. The plausibility of a step does not automatically validate the steps that follow.
Disagreements also concern the nature of intelligence
Some disagreements are about what is missing, rather than how quickly progress is happening. If the main obstacle involves engineering and resources, extending a trend may seem reasonable. If a conceptual problem must first be solved, that extrapolation becomes less informative.
In On the Measure of Intelligence, published in November 2019, François Chollet proposes characterizing intelligence through the efficiency of skill acquisition. His analysis distinguishes performance on a task from the ability to learn, taking prior knowledge and experience into account. This proposal directs attention to learning conditions rather than deriving a timeline from an isolated score.
Our page on François Chollet’s approach to generalization explores this distinction. The distinction alone does not provide an arrival date for AGI.
Yann LeCun sets out another proposal in A Path Towards Autonomous Machine Intelligence, dated June 27, 2022. He presents an architecture combining, among other elements, a predictive world model, self-supervised learning, and hierarchical planning. The paper describes a research direction without establishing a calendar deadline.
These approaches should be assessed through their specific proposals. They do not justify assigning their authors a date absent from the publication, or assuming that criticism of current methods proves AGI must be far away.
Why estimates differ so widely
Differences in definitions explain some of the gap. Assumptions about future progress explain the rest: whether improvements continue, how much additional resources deliver, and whether systems can contribute to research themselves.
A short timeline may assume that several obstacles will fall together. Another forecast may treat reliability, lasting learning, or planning as requiring additional stages. Identify those assumptions before comparing the final years.
| Assumption to examine | If it supports an early arrival | If it adds delay |
|---|---|---|
| Continuity of progress | Existing methods remain productive | Diminishing returns require a new approach |
| Contribution to research | Agents accelerate useful experiments | Verification and hardware limit the acceleration |
| Transfer of capabilities | Skills carry over to new fields | Each field requires substantial adaptation |
| Reliability | Errors become detectable and recoverable | Rare failures prevent delegation |
| Deployment | Systems are affordable and can be integrated | Costs and operational constraints delay use |
This table is a tool for examining reasoning; it does not assign probabilities to the different paths. It also shows why the same observation can support different interpretations, depending on the mechanism assumed.
Short timelines should not be dismissed simply because they are short. Long timelines do not become more credible merely by appearing cautious. The quality of a forecast depends on how precisely it states its conditions and whether its reasoning can be revised.
Reading probabilities and intervals
A median date divides the estimated probability distribution into equal halves. It does not predict that the event will probably happen during that particular year. A forecast can have a near-term median while still allowing a substantial possibility of very long delays.
The most likely year and the median date also answer different questions. The former identifies the peak of the annual probability distribution; the latter depends on all the probability accumulated before a date. Confusing them can create a misleading impression of precision.
Intervals require similar care. Some represent a forecaster’s uncertainty about the future event. Others describe only the statistical uncertainty of a fitted model, assuming that the chosen model is correct.
A narrow interval around a slope does not necessarily account for a technological break or a change in testing procedures. Precision in the calculation does not remove uncertainty in the assumptions.
Finally, a conditional forecast must retain its condition when summarized. Arrival under the fastest technically possible progress is not equivalent to arrival after accounting for every possible interruption. Removing that qualification changes the meaning of the published result.
What evidence should lead to a forecast revision?
A useful forecast states what could strengthen or weaken it. Without that rule, every result risks becoming a confirmation after the fact. Revisions should address the proposed mechanisms, rather than simply reflect a new overall impression.
For a forecast based on autonomy, look for long, unfamiliar tasks evaluated independently. The evidence needs to include human interventions, abandoned attempts, and resources consumed. A successful final result does not describe the entire process.
For a forecast focused on learning, examine whether systems acquire unfamiliar rules and reuse them in another context. For an economic forecast, examine net gains after checking and corrections.
The main signals to track are:
- Success on tasks kept out of training.
- Recovery after an error or a change in circumstances.
- Reductions in the supervision required.
- Transfer of learned skills between environments.
- The total cost of a usable result.
Our page on AGI benchmarks explains the limitations of these measurements. The article on how to read model rankings helps distinguish measured progress from a comparison altered by its testing procedure.
Use forecasts without depending on a single date
For a business, a public organization, or an individual, the practical task is to prepare decisions that remain useful across several possible paths. Training people to verify outputs and documenting processes can be worthwhile even if AGI takes longer to arrive.
This preparation can be organized around three qualitative paths: rapid progress in general capabilities, extensive but uneven automation, or a slowdown caused by particular obstacles. These paths help test whether a decision will hold up, without assigning them arbitrary probabilities.
Then tie each investment to an observable trigger. Broader delegation might depend on demonstrated reliability on your tasks and an acceptable cost. The guide to evaluating AI agents provides a method for designing those checks.
Keep a record of the forecasts you follow: the publication, target event, conditions, probability, and revisions. This prevents an earlier estimate from being silently replaced by a corrected version. It also makes it possible to understand what the evidence has actually changed.
The AGI and research hub brings together the concepts needed to continue following this work. The lasting questions concern which capabilities have been demonstrated, under what conditions, and what they reasonably allow us to anticipate.
Key takeaways
- AGI dates can only be compared after examining their success criteria.
- Surveys, extrapolations, scenarios, and research proposals provide different kinds of information.
- A median, an early possibility, and the year in a narrative are different claims.
- Revisions should be dated and linked to the observations that prompted them.
- Base decisions on verified capabilities and several possible paths, rather than a single deadline.
Related pages
Frequently asked questions
When is AGI expected to arrive?
Every date needs to be tied to a specific definition and publication. A timeline for automating programming answers a different question from a forecast about learning entirely new tasks.
Does a 50% forecast mean AGI will arrive that year?
No. It marks a cumulative probability threshold: under that estimate, the event is equally likely to happen before or after that point. A much later arrival therefore remains possible.
Why do researchers' AGI forecasts differ so much?
Researchers disagree about both the obstacles ahead and the criteria for success. Technical expertise also cannot give them certain knowledge of future discoveries, industry decisions, or deployment constraints.
Can benchmarks tell us when AGI will arrive?
A benchmark provides evidence about a defined set of tasks under a particular testing procedure. Turning that evidence into an AGI timeline requires additional assumptions about capabilities the test does not cover and whether progress will continue.
How should a business use AGI forecasts?
Prepare for several possible paths and tie investments to observable results in your own operations. A decision is more robust when it remains worthwhile even if progress is slower than expected.