There's nothing intelligent about it. It's just fancy pattern-matching
The problem is that there is no clear-cut definition or dividing line. I've seen long online debates about this, and there are no good lines in the sand yet. All attempts failed key tests offered up, or were too subjective to evaluate well.
For one, we still don't know enough about how the human brain works such that we cannot say what distinguishes things called "AI" from something as powerful as the human brain. For all we know, the human brain is merely "fancy pattern matching" at a level of fanciness we don't understand yet.
Some call pattern-matching AI "lossy statistical analysis for the sake of speed/cost".
I suspect human brains also (typically) use abstract modelling of various sorts where symbols or some kind of ID's with attributes/links/factors are stand-in's for actual people and things to simplify certain cognitive processes. Thus, the human brain may merely be "fancy pattern matching" coordinated with "fancy modelling": statistics + modeling.
Various known AI techniques use pattern matching and others use modelling, BUT nobody has found a way to coordinate them together in a general-purpose way to reinforce each other (triangulate). It's as if we got all the key parts, but don't know how to put them together right. We don't know how to build central governors to coordinate AI "organs" for common goals.