Ask ten AI experts when superintelligence will arrive and you will get ten answers, from “before 2030” to “never”. Here is a fair map of the territory — and why the disagreement is not a sign that anyone is being silly.
What we are predicting
Be precise about the finish line, because vagueness causes most of the confusion:
- AGI — a machine roughly as capable as a person across most tasks.
- Superintelligence (ASI) — a machine far beyond the best humans at almost everything.
Many predictions you read are about AGI. Whether ASI follows quickly is a separate question. Here is the difference in detail.
The four camps
1. “This decade.” Some leaders of frontier AI companies say AI could match or beat expert humans at most economically useful tasks within a few years. Their evidence: the pace of improvement from 2020 to today, and the fact that AI now helps write the code and design the chips for the next generation. Worth noting: these people also have products to sell, which doesn’t make them wrong but is relevant.
2. “Mid-century.” When researchers are surveyed in large numbers, the median estimate for human-level AI has tended to land between the 2040s and 2060s — and that median has moved earlier in recent surveys. But the spread is enormous; the average of wildly different guesses tells you less than it appears to.
3. “Much later, or never.” Some scientists and philosophers argue that today’s approach — learning statistical patterns from vast data — lacks essential ingredients: grounded understanding, continual learning, perhaps embodiment. On this view, impressive benchmark scores are being mistaken for the real thing, and we may be decades away or facing a wall.
4. “Unknowable.” Almost everyone agrees on this much: the track record of AI prediction is poor in both directions. Go was “ten years away” in 2015 and fell in 2016. Fully self-driving cars were “a few years away” in 2015 and mostly still are.
Why sensible people disagree
- No precedent. Weather forecasts work because there are millions of past days to learn from. There are zero past examples of building a new kind of mind.
- No agreed finish line. Does passing exams count? Doing a job for a year? Different definitions give different dates.
- Different weights on the same clues. Rapid progress on benchmarks impresses some experts and leaves others unmoved, because benchmarks are not the same as capability in the world.
- Incentives. Companies, researchers seeking funding, and commentators seeking attention all have reasons to lean one way or another. That doesn’t settle who is right, but it is worth remembering.
Does AGI mean ASI is close behind?
This is where the intelligence explosion argument comes in. If an AGI could improve its own design, each generation would build a smarter one faster — and ASI could follow AGI within months. Sceptics reply that the physical world sets speed limits: chips take years to manufacture, experiments take real time, and some problems get harder rather than easier. The honest answer is that this, too, is unsettled.
How to read the next headline
- Check which thing is being predicted: narrow AI, AGI or ASI.
- Ask who is predicting and what they gain from your belief.
- Look for what the system can actually do, not what someone says about it.
- Remember that “soon” has been wrong before — and so has “never”.
We keep the dated facts on our timeline page in a single file and update it as things change. The predictions section sits deliberately in the fog.