Narrow AI describes systems built to perform specific, well-defined tasks well. By the definition used here, every AI system that exists today is narrow, including the most capable current language models. Strong AI, closer in meaning to Artificial General Intelligence, describes a hypothetical system with genuinely general intelligence comparable to a human's across essentially any domain. The distinction matters because conflating the two leads to real confusion about what today's AI systems actually are and aren't.
What makes a system narrow, even when it's very capable
A system is narrow if its capability, however impressive, is scoped to the kinds of tasks it was designed or trained for, rather than being a general capability that transfers seamlessly to any new domain. A chess engine that plays at a superhuman level is narrow: it has no capability at all outside chess. Modern large language models are far broader, capable across a wide range of language, reasoning, and coding tasks. Many researchers still class them as narrow in the strict sense: they don't set their own goals, they don't learn continuously the way a human does, and their broad but bounded capability isn't the fully general intelligence "strong AI" describes. (Some researchers argue these systems already show early forms of general intelligence, so the line is debated.)
Why today's most impressive AI systems are still narrow
It's easy to look at a highly capable language model and assume its breadth of apparent capability means it's approaching general intelligence. But breadth of task coverage isn't the same as generality in the strong-AI sense. A system that's remarkably capable across many trained-for tasks is still different from one with the flexible, self-directed, continuously learning intelligence "strong AI" refers to.
Why this distinction matters practically, not just philosophically
Treating a narrow AI system as if it has strong-AI-level understanding leads to real mistakes: trusting it in situations outside its trained scope, assuming its apparent understanding transfers to domains it was never designed for, or building a safety and oversight model around general reliability the system doesn't have. Recognizing that even the most capable current systems are bounded keeps deployment decisions grounded in what a system can verifiably do.
Where the field genuinely stands
Every AI system currently in production or research use is narrow by this definition. Some are remarkably broad in the range of tasks they handle well, but they're narrow nonetheless. Strong AI remains a hypothetical future capability with no reliable timeline, as covered in more depth in our piece on what AGI actually is. Treat any claim that a current system has achieved, or is close to, strong AI with real skepticism.
How we approach this
We build and deploy narrow AI systems, deliberately scoped to specific, well-understood tasks with tested, verified capability. We don't assume that broad apparent capability implies a general reliability the system doesn't have.