That's actually a pretty complex argument.
Porting the drivers and such to a microkernel architecture in full (L4, Minix, Hurd) would isolate parts of the code and require strict API adherence (and ABI, but ABI amounts to your IPC protocol). That reduces the scope of bugs, in the long run; and it minimizes short-term porting bugs. The cost is essentially a large amount of man-power.
So you have the likelihood of finding a lot of bugs, eliminating a lot of bugs in the process, and creating new bugs, all at odds with each other, and each with different short- and long-term implications (you'll create new bugs in the short-term, but fewer than e.g. porting everything to BSD; and you'll eliminate and produce fewer bugs in the long-term); along with the enormous cost of simply organizing the change (everything has to be broken down and fixed around boundaries first).
The single short- and long-term advantage of keeping the Linux kernel architecture is it's a hell of a lot less work to not rearchitect an OS kernel.