Human brains are poor RNG. People are bad at generating random values in the privacy of their heads. They just cannot think randomly; though they can convince themselves that they do.
Physical process, on the other hand, are rather good sources of entropy. Take your mouse movements. A few dozen times per second, the mouse measures how far it has moved since the last tick, and sends that information to the server. When your hand shakes, it tends to do so somewhat regularly, but biology is such that each elementary move will be subject to some jitter, which happens to be substantially bigger than the precision of the mouse; even with a lot of training, it is very hard for a human hand to do the exact same move repeatedly (otherwise there would be a lot more people like Yehudi Menuhin). So the bottom line is that mouse movement measures contain some entropy. (Remember that "entropy" is here defined as "that which the attacker does not know"; the mouse certainly knows how much it has moved, since it is that mouse that actually sends the values on which the RNG are built.)
The other half of the answer is aggregation. A mouse-based RNG will use hundreds or even thousands of measures, accumulate them all and condensate them into an appropriate seed that will concentrate all that entropy. This is simple enough: simply feed all the values to a cryptographic hash function, e.g. SHA-256, and you will get a 256-bit seed that has all the source entropy, wherever it was hiding in the measured mouse movements. Hash functions are good for that; they reduce the size but keep the entropy (up to the hash function output size, but 256 bits is more than enough for all purposes).
An attacker may guess that the user will do circles, but will have a hard time getting all the individual movements right, especially since psychology won't help him: the human user himself has no idea how his hand movements are turned into numbers. Since we are talking about hundreds of numbers, the number of possible combinations (i.e. "entropy") raises exponentially. Contrast that with a human user thinking about a new password: the user will choose letters following some inner "witty" train of thought, that the attacker can guess more or less brutally (e.g. if the letters are all the first letters of some words in a sentence from a book, the attacker can automatically try all sentences from all books he can find in electronic format); and, more importantly, the human user won't be bothered to produce more than a dozen or so of "seemingly random" characters.
In passwords, length does not make strength -- but lack of length can be quite effective at preventing strength.