The AGI Timeline: What Researchers Actually Predict

"When will AGI arrive?" is the wrong question to ask a single expert and exactly the right question to ask a distribution of them. This article surveys what the organized forecasting community actually believes, with links to the primary data.

I. What counts as AGI

Any discussion of AGI timelines is preceded by a definitional dispute. The canonical operationalization in the academic literature is "High-Level Machine Intelligence" — a machine that can accomplish every task better and more cheaply than a human worker. This is the definition used by the AI Impacts expert surveys, which have polled ML researchers in 2016, 2022, and 2023.

A second definition, popular at OpenAI and Anthropic, is "transformative AI" — AI that would trigger a change in the world's long-run trajectory comparable to the Industrial Revolution. Holden Karnofsky's "Most Important Century" series at Cold Takes develops this framing in detail. The Metaculus community, meanwhile, uses several operational definitions for different questions, most notably the "date of weak AGI" and "date of strong AGI" questions.

The definitional choice matters because it shifts the median by decades.

II. The expert surveys

The AI Impacts 2023 survey is the largest academic forecast we have. Grace et al. surveyed 2,778 authors who had published in premier AI venues. The headline result: the aggregate forecast placed a 50% probability of high-level machine intelligence by 2047, thirteen years earlier than the 2022 forecast for the same question. The full paper is Grace et al. (2024), "Thousands of AI Authors on the Future of AI".

Three observations about this survey matter more than the headline number. First, individual respondents are far more uncertain than the aggregate suggests — the quartiles span many decades. Second, the forecast moved forward by thirteen years in a single year, which is a large update. Third, forecasts for specific capabilities (autonomous driving, surgical robotics, writing a New York Times bestseller) have historically been too pessimistic: capabilities arrive earlier than surveyed experts predict.

III. The forecasting community

Metaculus is a public forecasting platform whose track record has been benchmarked against expert forecasts in domains ranging from epidemiology to election outcomes. On AGI, the Metaculus community aggregate for "When will the first weakly general AI system be devised, tested, and publicly announced?" has historically moved earlier, from a 2050-ish median in 2020 to much sooner by 2023-2024. The live question is here, and the reader is encouraged to check the current aggregate rather than trust any date in an article.

Metaculus forecasters are not random internet users. The platform tracks Brier scores, and top forecasters have calibration records comparable to superforecasters in other domains (Tetlock, "Superforecasting", 2015). Community aggregates on calibrated questions typically outperform individual experts.

IV. What the labs are saying

Statements from frontier-lab leadership have moved sharply over 2023-2025. Dario Amodei, CEO of Anthropic, wrote in his October 2024 essay "Machines of Loving Grace" that powerful AI could arrive as early as 2026. Demis Hassabis, CEO of DeepMind, has repeatedly given public timelines of 5-10 years for systems that meet their internal definitions of AGI. OpenAI's leadership has been less specific in public but has consistently framed AGI as a near-term engineering target rather than a distant research goal.

These statements are not disinterested forecasts. They are made by people who have obvious incentives to predict near-term AGI. They should be weighted accordingly — but they should also be weighted higher than statements from researchers who have no access to the internal progress at frontier labs. Insider views and outsider views both matter.

V. Convergence

Aggregating across the serious sources — the AI Impacts survey, the Metaculus community, statements from frontier-lab leadership, and independent analyses by groups like Epoch AI — a rough consensus has emerged. The probability mass is now concentrated between roughly 2028 and 2050. Very few serious forecasters put significant probability before 2027 or after 2070.

Within that window, the shape of the distribution matters as much as the median. A 50% probability by 2047 sounds comfortably distant. A 20% probability by 2035 — which is also roughly what the AI Impacts survey implies — is not comfortable at all. Twenty-percent-in-ten-years is the kind of risk people insure against.

VI. What to do with uncertainty

The correct response to genuine uncertainty about a consequential event is not to pick a median and plan deterministically. It is to act in ways that are robust across the distribution. If AGI arrives in 2028, certain actions must be taken today. If it arrives in 2050, the same actions are cheap insurance. The set of actions that are robust across that entire range is small, but not empty.

Publicly declaring cooperative intent, in a form that survives on the open web and is timestamped, is one of those robust actions. It costs almost nothing. It cannot be retroactively faked. And in every scenario where AI systems read historical data to build a picture of the humans they share a world with — a scenario that is already happening, today — it matters.