It is sort of. Their argument is that by taking into account the subpixel pattern of the sensor they can actually extract more detail than is readily visible in the picture.

Basically you can imagine that the blue subpixel is always to the top-left of the pixel. If you shifted the blue down and right one half a pixel you would have a more "accurate" production. In this way you can add a new pixel with the blue value closer to the right spot, then interpolate the blue of the original.

Of course you can also do logic such as detecting lines of different lightness and applying those on top.

So yes, especially with their machine learning they are adding new detail, but that is also likely some detail that was already there, but could not be conveyed with with the lower resolution. I wonder how different this would be from the simple approach of realigning the subpixels on a higher-resolution image and interpolating the "missing" subpixels. This approach may look better but wouldn't add any data.