AI researchers are working towards creating artificial general intelligence (AGI) but there is no clear road map for how to do this and there is no clear way to tell if we have created AGI. This post discusses ways that we might use to test for AGI.
Perhaps we can use the Turing test to test for AGI? When I studied AI at university I was taught about the Turing test as being a way to test for artificial intelligence. The Turing test is a practical way to test whether a machine has intelligent behaviour indistinguishable from that of a human. In the Turing test, the computer is hidden in a closed room and a human is able to have a text-based conversation with the computer. If the human cannot tell whether the computer is human or not then the test is passed.
However, I don't think the Turing test will be enough to test for AGI. This is because the Turing test does not test for some of the qualities of AGI such as the ability to learn.
The qualities of AGI can be found by comparing AGI with the AI that we have currently. At the moment we have many task-specific artificial intelligence programs which can equal or surpass human performance on those specific tasks (examples include game playing, route finding, product recommendation, medical decision support, speech recognition, machine translation, text prediction, face recognition, theorem proving, equation solving, spam filters, information retrieval and the list goes on).
Artificial general intelligence is something that we are working towards and it is not task-specific. It includes the ability to learn as an integral feature. If a human is faced with a new task he or she is able to transfer skills from other tasks and trouble-shoot the situation to find a way forward. An AGI will be able to do this and so AGI will need to have the ability to learn how to do new tasks.
AGI will also have the ability to extract useful concepts from sensory data and internal states. It should be able to use the concepts to reason.
So to test for AGI we need to adapt the Turing test so that it includes a test for learning and for the ability to extract concepts and reason using the concepts.
Perhaps the Turing test could include giving the computer an IQ test and test on tasks it has not previously encountered?
Some authors have argued that if the computer is able to converse using natural language then it will have AGI. Indeed a human needs an IQ of around 70 to be able to use language. But I think that even if the computer can use natural language this is not enough to demonstrate AGI because of the need to demonstrate the ability to cope with previously unseen tasks and situations.
The Turing test, even if imperfect, at least provides something that can actually be measured. As such, it is a pragmatic attempt to answer a difficult philosophical question.