Narrow AI, General AI, and the Landscape Today
You'll learn to
- -Distinguish narrow AI from artificial general intelligence (AGI)
- -Place today's large language models correctly within that distinction
Every AI system that exists in production today, image classifiers, recommendation engines, large language models included, is narrow AI. It performs a specific class of tasks well, often superhumanly well, but it does not have general, flexible intelligence across arbitrary domains the way a human does.
Narrow AI
A narrow AI system is built and evaluated for a bounded task or set of tasks. A model that plays chess at a superhuman level cannot fold laundry. A large language model that writes fluent code and prose still fails at tasks a five-year-old finds trivial, like reliably counting objects in an image without a specific tool for it, or truly understanding a fact it was never trained on. Fluency at one class of task does not imply general competence.
Artificial General Intelligence (AGI)
AGI describes a hypothetical system with human-level flexibility across essentially any intellectual task, able to learn genuinely new domains the way a person can. Whether AGI is close, far away, or reachable at all with current techniques is a genuinely contested question among researchers. This course does not take a position on it, but you should recognize the term when marketing or press coverage uses it.
- -Narrow AI is excellent at a bounded task, but does not generalize outside it. Every deployed system today falls here.
- -AGI is a hypothetical, human-level-flexible system, not yet demonstrated, and actively debated.
Be skeptical of any product claim that implies general intelligence. A model that writes excellent marketing copy and a model that reliably reasons about a novel legal contract may share an architecture, but neither implies the other. Always ask what the system was actually evaluated on.
A vendor pitches a single AI model they claim handles customer support, legal contract review, and medical triage equally well. How would you evaluate that claim?
"If it is a large, well-known model, I would trust that it can probably handle all three."
"I would be skeptical by default, since that claim implies something close to general intelligence, and every AI system in production today, including large, capable models, is narrow AI: strong within a bounded task domain, not automatically strong outside it. Fluency at customer support conversation does not imply reliable performance on medical triage, where errors are far more costly. I would ask specifically what the model was evaluated on for each of those three domains independently, rather than accepting fluency in one as evidence for the others."
What best describes "narrow AI"?