A system created by another sentient being (e.g. humans) that is capable of cognitive analysis and unique production or output. Currently there are two primary approaches to AI: Atomic Components of Thought or ACT, and connectionist (or neural network) models.
ACT models of cognition use elemental tasks and then build larger sequences of actions composed of these elements. The atomic components can be either actions or declaration of certain facts. The components are strung together according to circumstance, and most implementations of this structure are capable of incorporating new knowledge into the architecture. While there are many advantages of this method, it is severely restricted in its capability to generate multiple solutions or action sequences, and is heavily dependent on the a priori information programmed into the architecture.
On the other hand are connectionist models that use parallel processing to generate a system of solutions that simultaneously satisfy multiple constraints. Connectionist models involve three layers: input, hidden and output layers. The hidden layer is driven by a differential equation that in most cases is trying to reduce the error between the output calculated by the hidden layer and the target output provided by the programmer. The advantage of connectionist models is that they are capable of taking inputs that they have never seen before, and generating outputs consistent with their experience. If the input is totally unrelated to the network's experience, the generated output will be meaningless. However, if the input is part of a set with which the network is familiar, and there are sufficient hidden layer units to properly represent the situation, then the network can output highly accurate responses even if it has never seen that particular problem before.
Connectionist models have many disadvantages as well, though most are specific to a particular implementation. In the example above, the network requires target outputs to determine the error of its guess, if there were no programmer providing these targets then these models would be useless. There are connectionist models that do not require target outputs, the network simply compares different input sets and looks for a set of relationships. However, these networks have great difficulty dealing with non-uniform sets, ie problems that do not have a central tendency or definite relationship.
Overall artificial intelligence is a far way from being in everyday use. While quite capable of modeling human behavior, both the ACT and connectionist architectures are difficult to translate in to actual practical devices. While the US Air Force and certain IT solutions providers have begun to implement these models in more advanced programs, we are far from designing the HAL 9000.