ADC filter?
Ah! Then your answer is this:
If the source impedance is varying (R depends on T!), then your filter won't be consistent. That's bad!
How to solve?
Add a resistor in series, from the divider, to the ADC. Make this resistor value several times larger than the thermistor's highest value. Place the cap-to-GND after this resistor.
If you have a 10k + 10k(thermistor) divider, the maximum Thevenin resistance will be between 1 and 10k, for most temperatures you'd measure. Add a 47k in series from this point, and follow it with, say, a 0.1uF cap, to get a time constant of 4.8 to 5.7 milliseconds.
If you can't tolerate a large series resistance (maybe this application is low current, and the thermistor is huge, like 1Meg, and the ADC can't reliably read from a source resistance of 5Meg -- actually it probably will, despite what the datasheet says, but at that point, it depends), then the alternative is to buffer the signal. You'd still use a capacitor on the thermistor, accepting that it won't do your final filtering job -- its job is only to set a maximum bandwidth. This goes into an op-amp voltage follower, which is followed by another filter, which will have a constant cutoff frequency. (You may also be able to use the op-amp as a Sallen-Key active filter, and get even sharper cutoff. Not an issue for temperature inputs, but a valuable approach for faster analog inputs.)
This becomes very important at high frequencies, for example for a function generator output circuit: the DAC output needs to be AA filtered, but you can't connect the filter to the outside world because its impedance isn't constant (it depends on frequency -- that's why it's a filter, after all!). You'd place the buffer (and you most likely need a lot of gain, anyway) after the filter, so the filter response remains constant, no matter what gain setting or load you have attached.
Tim