AI is a broad term referring to more than one concept, each with its own definition.
Different Types of AI
The lowest levels are extremely simple and common, such are an artificial chess opponent. They work well because the programs internal model of reality is a 8x8 grid with only a few rules. The program chooses a preferred action by running simulations in it's internal model of reality.
What is often meant by AI is a "General Intelligence" that can come up with a creative solution to any directive, based on it's internal comprehension of the world. There is no existing example of this as of yet. The problem is that its internal model of reality needs to provide for every possible action. And it's possible actions and reactions may not be limited to a finite number of discretely distinct moves as in a game of chess. (At least if it works on the basis of conventional programming) Even ignoring the computational power needed to run such a broad and inefficient program, it would also require some stupendous amount of labour to program this internal model of reality in the first place.
I think in Sci-Fi, when people say AI, they mean a computer program that has a kind of awareness of itself and the world around it and can come up with creative and unexpected courses of actions in order to achieve its objective. Often the AI is NOT sentient, which is why it does not understand that its actions are morally wrong, or that its solution defeats the underlying intention of its assigned directive. The Horror lay in the concept that an amoral entity has more processing power than human kind.
