I saw a good explanation on Reddit by Redditor JustMeSR:
"It's a library. Think of it as a pre-made tool for working on something (in this case AI). Just like when you want to repair something you don't start by creating your own screwdrivers, hammers, etc. You use the ones available. (Tenserflow represents the tools in this example).
As for AI. It's just a subset of CS that uses graphs, statistics and probability to get some results. It "learns" by getting examples of right results. (Let's say you want computer to "learn" to predict which team wins a certain match. You start by giving it previous matches, info about them and the result -> who won. After giving it info about any match it will be able to predict who won based on what it "learned" from your examples).