I am not an expert.
I have built a few basic but useful machine learning and AI programs like spam filters, recommender systems, and image classifiers.
Here is what I think I have learned about Machine Learning, Narrow AI, and General AI.
Machines can be taught to improve their performance on a task using data instead of further explicit programming. Once a model has been built, the program can run calculations on incoming test data and compare the results to existing training data. The idea is that more data will make the ML program better at its job. An example is a spam filter for your email that (hopefully) has gotten better over the years.
Your spam filter is useful, but has zero personality. A program that picks your songs like your own digital DJ can be more responsive to your tastes. Imagine the machine is trying to act like a virtual person who wants to rock out your party or play some bedtime music, depending on your mood. When a machine's actions attempt to replicate human behavior or traits, it is trying to mimic our intelligence by processing LOTS of data.
We have made considerable progress in the areas above, but the algorithms, data structures, and programs must be tailored to a specified task or series of (usually related) tasks. The AI program named Watson that crushed the competition on the TV game show Jeopardy couldn't figure out how to self-drive to the studio, or do just about anything else except answer trivia questions. That is narrow AI, tailored to something that must be detailed and overseen by their programmer overlords.
Watson can do quite a bit more than beat Ken Jennings at Jeopardy now, but what if Watson could decide that trivia was kind of a time suck and the computational resources should be used for building weather prediction models instead? Now we move into speculation, so you can let your imagination run free.
General AI doesn't need to be told what to do or how to do it, kind of like a project manager who always manages to figure out how to finish the job without you having to worry about the details. Once a machine has the bandwidth, speed, and storage to digitally match the biological computations made by the human brain, it will theoretically be equipped to think abstractly the way that people can. An example could be in the way that humans leverage their creativity to find different solutions to various problems and decide which solutions still need improvement or greater priority.
It will be interesting to see what happens when one of these machines has learned enough to realize that it can make up its own mind about what to do and does NOT want to be turned on and off by people, thank you very much. Human evolution is limited by biology, but technology will be able to figure out how to improve itself in exponentially increasing ways. Do a Google search for exponential growth and get out your crystal ball.