The spread is the most important result
Predictions for highly capable or human-level AI range from a few years to many decades, with some researchers doubting that current methods can produce it at all. This disagreement is not a minor error bar around a shared scientific model. Forecasters disagree about what counts as AGI, what technical steps remain, how quickly inputs such as compute and data can grow, and whether progress will continue smoothly. Any timeline should therefore be presented as a dated forecast under a stated definition—not as an “AGI countdown.”
Demis Hassabis provides a useful example of why exact attribution matters. In a March 2025 CNBC interview, the Google DeepMind chief executive said human-level AI could arrive in five to ten years. He did not say meaningful evidence of AGI would arrive “within the year” (CNBC interview). Even the correctly quoted forecast leaves questions: human-level at which tasks, with what reliability, and with how much autonomy? The quote records one expert’s judgment, not a measured probability.
Four kinds of forecast that should not be blended
Lab-leader statements are visible and informed by direct exposure to frontier systems, but they are not neutral measurements. Executives possess valuable private information about experiments, yet also lead organizations competing for capital, talent, customers, and political attention. Their statements may use company-specific definitions and rarely expose a complete forecasting model.
Researcher surveys aggregate many views and reveal the distribution of expert opinion. The 2023 Expert Survey on Progress in AI, published in 2024, gathered 2,778 researchers and reported a wide range of estimates for machines outperforming humans in every possible task; the authors also stressed substantial disagreement and sensitivity to question wording (Grace et al., 2024). Surveys describe beliefs at a moment in time. They do not validate those beliefs or guarantee that the respondents share a definition.
Prediction platforms such as Metaculus continuously aggregate forecasts. Their medians can update faster than surveys, but the result depends on the exact resolution criteria and participant pool. “Weakly general AI,” “transformative AI,” and “AGI” are different questions. A market or community median should be displayed with its retrieval date, question wording, range, and resolution conditions rather than turned into a site-wide deadline.
Model-based projections extrapolate trends in compute, benchmark performance, algorithmic efficiency, or the length of tasks agents can complete. These make assumptions visible, which is valuable, but extrapolation can fail when a benchmark saturates, a resource constraint binds, or a new paradigm changes the trend. A curve is evidence about past measurements; its extension is a conditional forecast.
What current capability trends do show
There is real evidence of rapid improvement. The Stanford AI Index documents falling inference costs and rising scores on demanding benchmarks, while also showing that difficult evaluations remain unsolved and that standardized responsible-AI evaluation is uneven (Stanford AI Index 2025). METR reports that the human-expert duration of software-oriented tasks frontier agents complete at a given success probability has increased quickly (METR time horizons). These are observations about selected systems and tests, not direct observations of an approaching AGI date.
METR’s metric is especially easy to misstate. A 50-percent time horizon of several hours means that, on METR’s task suite, the fitted model predicts a 50-percent chance of completing tasks that take a human expert that long. It does not mean the AI works autonomously for that number of hours. The tasks are primarily software engineering, machine learning, and cybersecurity; they are unusually self-contained and objectively gradable; estimates above sixteen hours are currently described as unreliable; and real jobs are much messier. METR researchers have warned that error bars can span roughly a factor of two and that domain differences can span orders of magnitude (METR limitations note).
These results support a narrower conclusion: some forms of digital task completion are improving rapidly. They do not, by themselves, establish when a system will become a dependable scientist, manager, caregiver, political decision-maker, or general substitute for human labor.
Why reasonable forecasts diverge
Forecasters who expect an early arrival often emphasize the persistence of scaling, better reasoning-time computation, synthetic data, tool use, and AI-assisted research. If those trends reinforce one another, progress might accelerate. They may also argue that economically transformative systems need not reproduce every feature of human intelligence; broad digital competence could be enough.
Longer-timeline forecasters emphasize reliability, grounding, continual learning, causal understanding, robotics, and the gap between benchmarks and open-ended work. They also point to practical constraints: energy, chips, data-center construction, high-quality data, evaluation, and organizations’ ability to deploy systems safely. Some expect architectural advances beyond current large-model methods. None of these positions can be settled by confidence or seniority alone.
Definitions create another source of apparent disagreement. A system that can automate much remote computer work could be economically transformative while failing embodied tasks. A model that scores above typical humans on a broad exam battery may still be too unreliable to operate independently. One forecaster may call either system AGI; another may reserve the word for robust human-equivalent learning in the real world.
How to maintain a responsible timeline tracker
A useful tracker should preserve the original wording, speaker, date, source, definition, and forecast range. It should distinguish a personal estimate from an organizational commitment, record later revisions, and never silently replace an old prediction. Aggregate forecasts should show uncertainty intervals rather than only a median. The page should also disclose potential incentives and whether the forecast relies on public or private evidence.
Readers can apply five questions:
- What outcome is being predicted? AGI, human-level AI, task automation, or superintelligence?
- What probability and date are attached? “Could” is not a median forecast.
- What evidence drives the estimate? Scaling, a survey, private experiments, or intuition?
- What would change the forecaster’s mind? A forecast without possible disconfirmation is difficult to evaluate.
- Has it been scored? Past predictions should remain visible after their deadline.
The practical conclusion
Planning should not depend on picking one date. Near-term harms and opportunities deserve action regardless of whether AGI arrives in 2028, 2040, or never. Governments and organizations can use scenario ranges: prepare for fast capability gains, build policies that remain useful under slower progress, and define measurable triggers for stronger safeguards.
The uncertainty is itself actionable. It argues for monitoring capabilities rather than marketing labels, investing in evaluation and resilience, and revisiting forecasts on a fixed schedule. A timeline page earns trust by showing how little the dates prove, not by making a moving median look like a clock counting down to destiny.