Yes, this. I've had to deal with inter-datum conversions in the past, and as far as I can tell, though after a
very brief review, GCJ-02 is just another datum, albeit one not well documented. The "random offset" described in various documents may be easily explained by the datum being optimized for accuracy in a relatively small area, and thus errors increase as you expand your mapped area.
(For those who don't speak GIS, datums are essentially a way of mapping an orange peel to a table; because any attempt to dedimensionalize points on a sphere must by necessity introduce errors, datums are designed to either minimize that error in all locations -- which is what WGS84 does -- or to gain greater accuracy by focusing on a smaller part of the globe, such as the "state plane" datums. When you see misaligned points on maps that derive data from various sources, it's usually a good idea to make sure that multiple unconverted datums aren't the source of the problem.)