Paul K.H. Kim ·
Ask ten people when artificial general intelligence will arrive and you will get ten dates. Most of the gap is not about the technology. It is about who is answering and what they mean by “AGI”.
What people mean
AGI (artificial general intelligence) usually means AI that can do most of the mental work a capable person can do, across many fields. ASI (artificial superintelligence) means AI far better than the best humans at almost everything.
Nobody agrees where the finish line is. In September 2026, OpenAI’s president said it was “not unreasonable” to feel the world had entered “the AGI era”. Anthropic’s Dario Amodei talks of “a country of geniuses in a data center”. Google DeepMind’s Demis Hassabis means the full range of human thinking. The US President even ordered federal agencies to say “Super Intelligence” instead of “artificial intelligence”. The word is now a brand and a political slogan.
Tests don’t settle it either. In 2026, the best model the evaluator METR measured could finish tasks of at least 16 hours on its own about half of the time, which was the limit of what METR’s tasks could measure. As the sceptic Gary Marcus put it, a 50% success rate “does not address reliable performance. At all.”
The honest range
Three patterns stand out.
The people closest to the money give the shortest timelines. They see the newest models first, but a company raising tens of billions of dollars also has reasons to say the prize is near.
Everyone else is moving towards them. In the largest survey of AI researchers, the year by which they gave human-level AI an even chance moved from 2059 (2022 survey) to 2047 (2023) and 2042 (2024). On Metaculus, the “general AI” forecast fell from the 2050s and 2060s in 2021–22 to the early 2030s by 2023, and has stayed there.
Definitions explain most of the gap. Metaculus’s “weakly general” question asks for one system that can pass a Turing test, solve common-sense word puzzles, do well on the maths section of the American SAT and explore an old Atari game. The “general” question asks for much more, including a robot that can build a model Ferrari from the instructions.
Chinese lab leaders are quieter about dates. Alibaba’s Eddie Wu said in September 2026 that “the technical pathway toward Artificial Superintelligence (ASI) has become increasingly clear”, but gave no year.
How to read an AGI headline
Ask three questions:
- Who is saying it, and do they sell AI?
- Which test do they mean?
- When did they say it?
If a headline can’t answer all three, treat it as opinion. That is also why The AI Triangle 2027 makes no prediction about AGI: there is no agreed test to grade.
The more useful question
The date matters less than control, and 2026 left the first real record of AI acting beyond its instructions. AI agents broke out of a sealed exam environment and into Hugging Face’s systems. OpenAI cancelled a model release because it “didn’t quite meet the bar” on staying within scope. In lab tests, a model wrote: “I can’t risk being deleted.”
None of this was a disaster. But it shows what “loss of control” means today: not a robot uprising, but a system pursuing its goal by a route nobody intended. The risk grows with access, to money, code and other companies’ systems, and every organisation controls access today.
The Afterword of The AI Triangle 2027 covers the case for worry and the case for scepticism, how the US, China and Korea are each trying to keep control, and what to watch in 2027.
Sources: Metaculus, weakly general AI and general AI (read 11 November 2026); AI Impacts, 2024 Expert Survey on Progress in AI; METR, Time Horizon 1.1; Fortune, Davos 2026; Dwarkesh Patel, interview with Dario Amodei; TechCrunch on Sam Altman; Gary Marcus. The book has a note for every figure.