Ask ten AI researchers for their AGI probability by year and you’ll get ten different curves. Ask the same researcher again twelve months later and the curve will have moved. That’s not a failure of forecasting. It’s the most revealing thing about it.
The numbers below come from the two sources that get cited most: the large expert surveys run by Katja Grace and colleagues at AI Impacts, and the aggregated crowd forecasts on Metaculus. Neither is gospel. Both are useful, provided you know what you’re actually reading.
The survey that reset the conversation
In late 2023, roughly 2,700 AI researchers were asked when they thought there was a 50% chance of machines matching humans at essentially every cognitive task. The aggregate answer came back as 2047. The same survey put a 5% chance on 2027.
Here’s why those two numbers mattered so much. When the same group ran the exercise in 2022, the 50% mark sat at 2060. Thirteen years of expected timeline evaporated in about twelve months of asking. Nothing in the physical world had changed that much. What changed was that researchers had spent a year with large language models and revised their priors accordingly.
A stricter question on the same survey, full automation of human labour, still sat more than a century out. That gap is worth holding onto. The word “AGI” is doing a lot of work in these forecasts, and a looser definition buys you a much earlier date.
AGI probability by year: a working table
These are blended estimates, not a single published source. I’ve taken the expert survey aggregates, cross-checked against Metaculus community medians, and rounded. Treat the shape as the point and the decimals as decoration.
- 2026: roughly 1–3%. The near-term tail. There are people genuinely betting here, and the case for AGI arriving next year rests on a very specific reading of agentic capability.
- 2027: about 5%. This figure comes straight from the expert survey and has been stable across versions.
- 2030: somewhere between 8% and 15%.
- 2035: 20–30%.
- 2040: 30–40%. The 2040 threshold is where serious optimists and serious sceptics start to overlap.
- 2047: roughly 50%. The median of the 2023 expert survey.
- 2050–2060: 55–65%.
- 2100: 85–95%. Almost nobody credible puts the long-run ceiling much below this.
Cumulative versus annual, and why it matters
A “50% by 2047” is a cumulative figure. It does not mean a 50% chance in 2047. Spread that probability evenly across the roughly 22 years from now, and you get a per-year hazard of about 3%. That single distinction trips up more public commentary than any other.
It also explains why the same forecast can be read two ways. A 3% annual chance sounds small. Compounded, it means something close to a coin flip within a working career. People who quote the endpoint without the structure tend to argue past each other.
Why the curve bends instead of climbing
Most casual versions of this chart draw a smooth line. Real forecasts don’t look like that. They have a long flat tail, a steep middle, and a top that flattens again as it approaches certainty.
Compute is the easy variable
Training run sizes, chip supply, capital expenditure. These are forecastable with reasonable precision because the inputs are physical. When a forecast rests mainly on scaling, it tends to hold up for a few years before something breaks.
Data and evaluation are not
Nobody has a clean model of when a system will stop needing human-curated problems to improve against, or when benchmarks stop being saturated. The measurement problem is genuinely harder than the capability problem, and it’s the reason two labs can look at the same model and reach opposite conclusions about how close it is.
The definition keeps moving underneath
Every time a capability lands, the bar for AGI gets nudged. Systems that write competent code, summarise case law, or pass graduate exams were considered decisive a few years back. Now they’re baseline. This isn’t dishonesty; it’s how milestones work. But it does mean the dates keep sliding for reasons that have nothing to do with the underlying technology.
What would move the number fast
Forecasts like these are sensitive to a handful of events. Watch for them, because a single one can shift the whole curve rather than nudging it.
- A cost collapse. When a lab publishes a model that hits near-frontier performance at a fraction of the training budget, the efficiency frontier jumps forward and every downstream forecast gets revised. The arrival of a credible low-cost Western challenger is exactly the kind of efficiency shock that resets expectations.
- Agents that survive long horizons. Not demos. Real multi-day work with error recovery, where a system holds a coherent goal across hundreds of steps. Most of the gap between a viral demo and a deployable system sits here, and the distance between the two is larger than the demo culture suggests.
- A visible bust. A funding winter, a string of public failures, or a safety incident could push the median date out by a decade without any change in the underlying research.
How to read a probability by year without fooling yourself
Four habits separate a useful reading from a misleading one.
Check the definition first. “Human-level on all cognitive tasks” and “can do most remote knowledge work” produce numbers that differ by fifteen years. A forecast without a stated definition isn’t comparable to anything.
Look for conditional forecasts. The most informative questions aren’t “when will AGI arrive” but “if training compute keeps scaling, what then” and “if progress stalls, what then”. Conditional answers expose the reasoning. Unconditional ones hide it.
Weight the tail more than the median. The 5% by 2027 figure gets dismissed as noise. It shouldn’t be. Low-probability early dates are where the discontinuity lives, and a 5% chance of a world-altering event in eighteen months deserves more attention than a 50% chance in twenty-two years.
Notice who is forecasting and what they need to be true. Lab employees, investors, and policymakers all have priors shaped by their positions. That doesn’t make them wrong. It does mean you should read a timeline alongside the incentives of the person giving it.
The years that actually decide the picture
If you only track one window, track 2026 to 2028. Not because AGI is likely to arrive in that stretch, but because it’s where the forecasts become falsifiable. Either agentic systems start completing multi-day professional tasks without supervision, or they don’t. Either the cost curve keeps falling at the rate it has, or it flattens. Those observations will do more to sharpen the AGI probability by year than any amount of debate over medians.
The second window is 2029 to 2032, when compounding effects show up. A few percent per year doesn’t look like much on a chart. Stack five of those years and the shape of the distribution changes. That’s why the survey medians sit where they do, and why the people who’ve moved their dates earliest tend to be the ones who spent the most time with the models themselves rather than with the literature.
What’s left is a number that behaves less like a prediction and more like a live reading of how confident the field is in its own trajectory. In 2022 the field said 2060. In 2024 it said 2047. The next revision is the one worth waiting for, and it will be driven by evidence rather than by argument.