There might be a bit of analog jiu-jitsu contributing to the mix, but, with that caveat, the linear data recorded at the sensor with a shot at, say f/8 @ 1/400 @ ISO 100 ("sunny f/16") will be identical to the same shot at f/8 @ 1/400 @ ISO 50. At some point in the processing chain, however, the values in the ISO 50 file will get halved (by simple arithmetic, not electronic amplification), resulting in an image one stop darker than the ISO 100 file. That will reduce noise overall. However, the sensor is still saturating at the exact same point. The net effect is that a pixel that the sensor recorded as, say 256, is being rendered as 128...and that there's nothing in the original data that gets mapped between 128 and 256.
Thanks to the gamma curve that gets applied after linear processing, the end result is that there's no data in the last stop. It therefore gets rendered as pure white -- and, thus, a loss of a stop.
You can do the exact same thing yourself, assuming your RAW processing software is capable of linear exposure adjustments.
It's potentially useful in scenes with low dynamic range, or in scenes where you don't care about highlights but do care about shadows.
The common term amongst photographers who do that sort of thing is, "ETTR." I generally strongly caution against doing that, as it's very easy to blow out the highlights, and there's so much wonderful and delicate color to be found in the highlights that is so easy to clobber. But there are certainly situations in which it can be useful.
In general, the meter in most cameras underexposes the linear data by one to two stops, and the processing pipeline applies an equal and opposite amount of digital overexposure to compensate. This is generally a very good thing, because sensors clip so readily and so unforgivingly and modern sensors have so little noise. But, yes, if you're very careful, you can make use of that "extra" headroom.