How close are we to AGI?
Nobody knows, and estimates vary widely. As of 2026, prediction-market forecasters put a roughly 50% chance of AGI by around 2033, while a large survey of AI researchers still puts the median near 2047. Progress is fast on tests, but machines remain far behind humans on open-ended, adaptive tasks.
Why — the first-principles explanation
How 'close' we are depends entirely on how you define AGI and what evidence you trust. There is no official finish line, so two honest experts can look at the same models and give timelines decades apart.
One camp watches benchmark scores climb. On the static ARC-AGI-2 puzzle test, the best model reached about 85% by mid-2026, near the human panel level. Systems now write code, pass professional exams, and use tools. To this camp, the curve is steep and AGI could be a handful of years away.
The other camp watches what machines still can't do. On ARC-AGI-3, an interactive test released in 2026, top models scored under 1% while people scored 100%. That gap is about learning on the fly, forming goals, and remembering across a long task, the everyday glue of human intelligence. To this camp, we're missing core ingredients, and scores on narrow tests don't prove generality.
Forecasters try to average these views. In 2026, Metaculus community forecasters cluster around a median in the early 2030s for a first general system, while the big academic survey of thousands of AI researchers lands closer to 2047. The honest summary: timelines have shortened sharply over the past five years, but the range is still wide, and confident single-date predictions are marketing, not measurement.
An example that makes it click
Think of climbing a mountain in fog. You can see you're gaining altitude fast, each week the trail is higher than the last, so it feels like the summit is near. But you can't actually see the peak through the clouds, so you don't know if it's one ridge away or ten.
That's AGI forecasting. The 'altitude' (benchmark scores) is rising quickly, which excites optimists. But because the summit (real general intelligence) is hidden and undefined, careful forecasters give a wide range, roughly the early 2030s to mid-century, instead of a single date.
Key facts
- There is no agreed definition of AGI, so timeline estimates differ by decades depending on the bar used.
- As of 2026, Metaculus community forecasters cluster around a median in the early-to-mid 2030s for a first publicly announced general AI system.
- The largest academic survey of AI researchers still places the median for 'high-level machine intelligence' near 2047.
- Expert timelines have compressed dramatically since ~2020, when many surveys pointed to roughly 50 years out.
- Machines reached ~85% on static ARC-AGI-2 by mid-2026 but under 1% on interactive ARC-AGI-3, showing progress is uneven.
▶ The 60-second explainer (script)
How close are we to AGI? The honest answer: nobody knows, and the estimates are all over the map. Here's why. There's no agreed definition of artificial general intelligence, so experts are aiming at different targets. If you watch benchmark scores, progress looks incredibly fast, top models now hit about 85 percent on hard reasoning puzzles. But if you watch what machines still can't do, like learning a brand-new game on the fly, they score under one percent where humans score one hundred. So forecasters give a range. In 2026, prediction markets cluster around the early 2030s, while a big survey of AI researchers still says roughly 2047. What everyone agrees on is that timelines have gotten a lot shorter over the last five years. Anyone who gives you one confident date is selling something.
What authoritative sources say
People also ask
Why do experts disagree so much?
Because AGI has no standard definition. Some measure it by benchmark scores, some by economic impact, some by human parity, so the same evidence yields very different timelines.
Are timelines getting shorter?
Yes. Over the past five years, many expert and forecaster estimates have shifted from mid-century toward the 2030s, driven by rapid gains in large language models.
What's still missing for AGI?
Robust on-the-fly learning, long-horizon memory, self-set goals, and reliable reasoning on tasks outside training data, exactly what interactive tests like ARC-AGI-3 expose.
Could AGI never happen?
Some researchers think current methods will plateau before reaching true generality. Most now expect it eventually, but 'eventually' still ranges from years to many decades.