It can be used as a text classifier. It takes as input a collection of (output, input1, input2, ..., input_n) tuples, stored in the newly-announced Google Storage, and then uses a variety of machine learning algorithms (which I would bet includes some Bayesian stuff) to build a model which it can use to take (input1, input2, ..., input_n) tuples and predict the corresponding output.

So, that's the API. You can do a lot of the same things offline, with almost the same file format, using Weka:

http://en.wikipedia.org/wiki/Weka_(machine_learning)

So if you're interested in playing around with Google's Prediction API, you should probably download Weka and fiddle with it some. It's pretty easy to get started with, and it will definitely give you an idea for the sort of thing you can do here.