In one line: narrow AI is brilliant within its training distribution and unreliable outside it; AGI would generalise to genuinely new problems the way a person does. ChatGPT, Claude and Gemini are all narrow AI — however impressive.
Side by side
| Today's AI (narrow) | AGI (general) | |
|---|---|---|
| Scope | Specific trained tasks | Virtually any task |
| Novel problems | Fails (see ARC-AGI) | Solves like a human would |
| Learning | Fixed after training | Learns from experience |
| Examples | ChatGPT, Claude, Gemini | None yet |
| Status | Here and widely used | Debated, undefined, not achieved |
The catch: "narrow" still means superhuman
Narrow doesn't mean weak. Today's AI writes faster than any human, codes across dozens of languages and summarises a 200-page document in seconds. It's narrow only in that it can't step outside what it learned. That combination — superhuman inside the box, near-helpless outside it — is the whole story.
Why the distinction matters for business
Treat narrow AI as narrow AI and it's one of the best tools you'll ever deploy. Treat it as general intelligence — assume it can reason about anything, trust it unchecked on novel problems — and it will fail you exactly when stakes are highest. The distinction is the difference between using AI well and getting burned.
Is AGI close?
Hotly disputed, and largely a fight over definitions and incentives. Some executives say it's here; leading scientists say scaling LLMs never gets there. The clearest evidence — near-zero scores on novel-reasoning tests — sits firmly on the sceptical side. The full picture is in AGI explained honestly and can AI become intelligent.
The practical takeaway is in our opinion piece: if it quacks, it's a duck — for business, what AI reliably does matters more than what it "is".