What is general AI?
General AI (or AGI, artificial general intelligence) is a hypothetical system that could learn and perform any intellectual task a human can, transferring skill across domains without retraining. It does not exist as of 2026-07. Every deployed AI today — ChatGPT, Gemini, Copilot — is narrow AI: broad-seeming, but trained and bounded.
Why — the first-principles explanation
The distinction turns on one word: transfer. A human who learns chess picks up something portable — planning ahead, thinking about what an opponent wants. That skill leaks usefully into business negotiation and even parenting, without anyone retraining the human. General AI names a system with that property across the board: give it a genuinely new domain, and it figures the domain out using what it already knows, the way a smart person handles a job they've never done.
Narrow AI has no such portability, even when it looks like it does. This is where modern chatbots create honest confusion. A system that discusses medieval history, writes Python, and drafts a wedding toast certainly seems general. But it isn't doing one flexible thing — it's doing an enormous number of narrow things, because its training data covered medieval history and Python and wedding toasts. The breadth lives in the data, not in a general capacity to learn. That's why the same model that writes elegant code can fail at a simple puzzle in an unfamiliar format: there's no underlying reasoner to fall back on when pattern-matching runs out.
There's a deeper structural reason today's systems aren't general: they don't learn from experience the way you do. When a model finishes training, its weights freeze. Your conversation doesn't change it. It cannot practice, notice it was wrong yesterday, and be better tomorrow — the improvement loop that defines general intelligence isn't running. Each conversation starts from the same frozen snapshot. A system that can't accumulate skill from its own mistakes is not general, regardless of how much it knows.
On timelines, the honest answer is that nobody knows, and confident predictions in either direction have a poor track record. Surveyed AI researchers give estimates spanning decades, with enormous disagreement — and the field lacks an agreed definition of AGI or an agreed test for it, which means "when will we get AGI?" isn't yet a well-formed empirical question. What we do have is measurement of current systems, and it counsels humility: METR's July 2025 randomized trial found experienced developers using state-of-the-art AI tools on real tasks were 19% slower, while believing they were 20% faster. Systems that can't reliably speed up an expert at their own job are a long way from doing any job a human can. That gap between perceived and actual capability is worth remembering every time someone claims AGI has quietly arrived.
An example that makes it click
Imagine two students facing a pop quiz on a subject neither has studied.
The first has memorized the answers to every question that has ever appeared on any exam, in any subject, going back a century. Ask him anything standard and he's flawless — better than any teacher. Then the quiz asks something genuinely new, phrased in a way that's never appeared before. He has nothing. Not a wrong answer arrived at by reasoning: nothing to reason with. That's narrow AI, and today's chatbots are this student with an astonishingly large memory.
The second student has memorized far less but can think. She's never seen the question either. So she works it out — pulls an idea from a different class, tries it, notices it fails, tries something else, gets there. Slowly, imperfectly, but she gets there. That's general intelligence.
Here's the giveaway: come back a week later. The second student is better, because she remembers being stuck and what fixed it. The first is identical — frozen, exactly as he was. That's the real gap. Not how much they know. Whether they can grow.
Key facts
- General AI (AGI) does not exist as of 2026-07; every deployed AI system is narrow AI trained for a bounded set of tasks.
- The defining property of general AI is transfer: applying knowledge learned in one domain to a genuinely new one without retraining.
- Modern large language models appear general because their training data is broad, not because they possess a general learning capacity — the breadth is in the data.
- Current models have frozen weights after training: they do not learn from their own mistakes between conversations, so they cannot accumulate skill through experience.
- METR's July 2025 randomized controlled trial found 16 experienced developers using state-of-the-art AI tools on 246 real tasks were 19% slower, while believing they were 20% faster.
- There is no agreed technical definition of AGI and no agreed test for it, which is a major reason expert timeline predictions vary by decades.
- NIST published the AI Risk Management Framework (AI RMF 1.0) on January 26, 2023, adding a generative AI profile (NIST-AI-600-1) in July 2024 to address risks unique to generative systems.
▶ The 60-second explainer (script)
What is general AI? It's a system that could do any intellectual task a human can — learning a brand-new domain the way a smart person handles a job they've never done. And it does not exist. Not in 2026. Everything you've used — ChatGPT, Gemini, Copilot — is narrow AI. The key word is transfer. A person who learns chess picks up something portable: planning ahead, thinking about what an opponent wants. That leaks into negotiation, into parenting, with no retraining. General AI would have that everywhere. Now, why the confusion? Because a chatbot that discusses medieval history, writes Python, and drafts a wedding toast sure looks general. But it's not doing one flexible thing — it's doing a million narrow things, because its training data covered all of them. The breadth is in the data, not in a general ability to learn. That's why the same model that writes beautiful code can fail a simple puzzle in an unfamiliar format. There's no reasoner underneath when the pattern runs out. And here's the structural tell: these models don't learn from experience. When training ends, the weights freeze. Your conversation doesn't change it. It can't be wrong today and better tomorrow. When will AGI arrive? Nobody knows, and there isn't even an agreed test for it. What we can measure isn't flattering: METR found expert developers using top AI tools were 19% slower on real work — while believing they were 20% faster. That's the distance we're talking about.
What authoritative sources say
People also ask
Does general AI exist yet?
No. As of 2026-07, every AI system in use is narrow AI. There is also no agreed definition or test for AGI, so claims that it has 'arrived' are arguments about wording as much as capability.
Isn't ChatGPT already general — it can do almost anything?
It's broad, not general. Its range comes from training data covering enormous ground, not from an ability to learn a new domain from scratch. It also can't learn from its own mistakes between sessions.
What's the difference between general AI and superintelligence?
General AI would match human ability across tasks. Superintelligence (ASI) would substantially exceed it across essentially all domains. Both are hypothetical; AGI is generally treated as the nearer milestone.
When will AGI arrive?
Nobody knows. Surveyed researchers give estimates spanning decades with huge disagreement, and without an agreed definition or benchmark the question isn't yet empirically well-formed. Treat confident dates — near or far — with skepticism.
Is general AI the same as strong AI?
The terms are often used interchangeably. Strictly, 'strong AI' originally raised whether a machine could genuinely understand or be conscious, while AGI concerns capability across tasks — a system could in principle be generally capable without being conscious.