Applying gain by increasing the ISO setting generally only serves to reduce read noise by amplifying the analogue signal before readout and digitization. Signal to noise ratio in the highlight remains unchanged. The increase in image brightness that also results would be irrelevant (besides any reduction in quantization noise) to the AF sensor, as it compares the pattern of brightness between two strips, the absolute value wont matter.
Read noise usually accumulates due to the transfer of the signal from a pixel on the sensor to an ADC chip, which is quite a long journey in some cases. The design of the AF sensor, with single strips of pixels facilitates a very short transfer with an ADC likely to be right on the sensor itself. This means read noise will be very low* and thus applying gain will be of no use.
The signal to noise ratio is pretty much determined by the amount of incoming light. When the signal to noise ratio is too low, phase detection simply wont work as the noise will be different across the two strips of pixels than make up the AF point, and the brightness values wont "match" even when in phase.
This is why you can't have AF sensors that focus in near darkness - reliability drops significantly. An AF sensor that refuses to focus is better than one that reports focus lock but then instructs the lens to focus at some totally random distance.
***** Performance characteristics of the AF sensor will be similar to the new Sony EXMOR image sensors (found in the Pentax K5 and Nikon D800 and others) which feature advanced noise cancellation and parallel ADCs and are regarded as being "ISO less" in that the noise obtained when pushing an ISO100 image two stops is the same as a native ISO400 exposure.
