The honest answer first
No. Superintelligence is not here yet.
Today’s AI can write, draw, talk and code well enough to surprise people every day. It is also capable of mistakes that a seven-year-old would spot, like miscounting the letters in a word or inventing a book that doesn’t exist. It is not better than the best humans at almost everything. By the definition on the first lesson, that means it is not superintelligence.
But “not yet” is not the same as “never”, and the story of how we got here is worth knowing.
The road so far
The timeline scene follows the signposts. Here are the ones that matter most.
1950 — a question. A British mathematician, Alan Turing, asked whether a machine could ever chat so well that you couldn’t tell it from a person. At the time, computers filled whole rooms and could barely add up.
1956 — a name. A small group of scientists met at Dartmouth College in the USA and coined the phrase “artificial intelligence”. They thought big progress would take one summer. It took rather longer.
1997 — chess falls. IBM’s Deep Blue beat the world champion, Garry Kasparov. Newspapers asked if machines were now smarter than us. They weren’t: Deep Blue could do nothing but chess. It was the perfect example of narrow AI.
2012 — learning from examples takes off. A neural network won an image-recognition contest by a huge margin. Suddenly the learning-from-examples approach that had been dismissed for decades was clearly the future.
2016 — Go falls. The board game Go has more possible positions than there are atoms in the universe. Experts said computers were ten years away from beating a champion. DeepMind’s AlphaGo did it that year. This is a good moment to remember: predictions have been wrong before.
2020 — proteins. AlphaFold predicted the 3D shapes of proteins, a puzzle biologists had worked on for fifty years. AI was now helping with real science.
2022 — chatbots for everyone. ChatGPT launched, and for the first time millions of ordinary people had a conversation with an AI. Within a couple of years, AI could see pictures, hear speech, talk back and write working computer code.
Today. AI is everywhere and still not superintelligent. It is, however, far closer to the human step than most experts expected in 2012.
Into the fog
After the “Today” signpost, the road in the scene disappears into fog. That is deliberate. Everything past this point is a prediction, not a fact.
Here are the main groups of predictions, presented as fairly as we can.
“Soon — this decade.” Some leaders of AI companies believe that AI matching or beating expert humans at most tasks could arrive within a few years. They point to how fast things improved between 2020 and today and argue that the ingredients are already in place. Critics note that these people also have companies to promote.
“Mid-century.” When large groups of AI researchers are surveyed, their average guess for human-level AI usually lands somewhere between the 2040s and the 2060s. But the spread is enormous — some say the 2020s, some say after 2100. When experts disagree that much, the average tells you less than it seems.
“Much later, or never.” Some scientists and philosophers argue that today’s AI is missing something essential — real understanding, the ability to keep learning, or a body that experiences the world — and that we may be much further away than the excitement suggests. A few think superintelligence may be impossible.
“Nobody knows.” Almost everyone agrees on this one. AI has arrived faster than predicted (Go) and slower than predicted (self-driving cars have been “a few years away” since about 2015). The track record of confident predictions is poor in both directions.
Why can’t the experts just agree?
It feels strange that clever people looking at the same facts reach such different answers. But it makes sense.
Nothing like this has happened before. When you predict tomorrow’s weather, you have millions of past days to compare it with. When you predict the arrival of a kind of mind that has never existed, you have zero past examples. Everyone is reasoning from clues, and reasonable people weigh clues differently.
Experts also disagree about what counts. Is a chatbot that passes exams “human-level”? Some say yes; others say passing an exam is not the same as doing a job. If you can’t agree on the finish line, you can’t agree on when the runner will cross it.
How to think about it yourself
You don’t have to pick a side. But you can be a good thinker about it.
- When you read a prediction, ask who is making it and what they might gain from you believing it.
- Notice whether they are talking about narrow AI, AGI or superintelligence. These are very different things.
- Remember that “a few years” has meant different things before.
- Keep an eye on what AI can actually do, not just what people say about it.
Dates that need updating
The facts on this page are kept in one file so they can be updated as things change. If you are reading this and the “Today” signpost feels out of date, tell a grown-up — or a teacher — that the site needs a new milestone. That is how honest timelines work: they get longer.