Consider what it means to have synchronous detection. Suppose there is a source at some particular frequency, and has I+Q modulation. Suppose there's a synchronous I+Q detector, as well. (i.e., a pair of mixers, driven by the local oscillator at 0 and 90 degrees phase shifts).
If the source and detector reference frequencies are equal, the frequency offset between them is zero, and the source and detector will be perfectly in phase (relatively speaking). If the phase is also correct, the input and output I and Q signals will match one-to-one. (If not, only a linear adjustment need be made -- the analog equivalent of multiplying by a complex number.)
If the frequency is off a little, then the phase rotates at the difference frequency. Any I+Q information sent from the source will still be present, but they will also be rotating. If you have a fancy decoder that can "spin its head around" and divide out the phase shift as it goes, you can still recover the information. But you can't read it without that added consideration: the raw I+Q won't be identical to the source.
Suppose we do the opposite: instead of sending independent data through the I+Q channels, we send I(t) = cos(2*pi*f*t) and Q(t) = sin(2*pi*f*t). Now the output phase essentially rotates at a frequency f, which would fool our detector into thinking it's the wrong frequency -- the I and Q are rotating precisely as if the carrier and detector frequencies are different. And indeed, this is no accident, because this is a perfectly roundabout way to shift a frequency!
For generating something like FM broadcast, this isn't very handy, because you need the trouble of an I+Q modulator set, with, in essence, a VFD (variable frequency drive -- normally used to control induction motors at up to perhaps 400Hz) of quite wide bandwidth (+/-75kHz deviation!) and high linearity. But if you have practically unlimited digital resources (say, an FPGA or a DSP chip to synthesize the signals), such a signal is absolutely no problem!
There's also the relation between phase and frequency modulation. Sure, you can take a sine wave and just shift it instantaneously to get something that always looks like a sine wave of constant frequency, cut up into pieces; but in any bandlimited phase modulation (which means, absolutely every real physical signal!), you necessarily have to momentarily "slow down" or "speed up" the carrier. Which means you're frequency modulating it, but only transiently. In fact, frequency is the derivative of phase, and phase is the integral of frequency. This ideal integrator is the reason why PLLs achieve perfect (integer ratio) frequency: even if the phase is unstable or noisy over a small range, if it never fully wraps around (skipping a beat, so to speak), the frequency must always be perfectly locked.
On a theoretical level, you can consider all of these methods as adjusting different parameters of a basic transformation. AM might be likened to the Fourier transform, FM to the Hilbert transform.
Tim