I have not investigated masking artifacts in auto_sharpen, so I would be very happy to have a look at your map if you can share it.   I am a particularly interested because (see below) I would have expected that any part of your map that is far from any atoms in your model would be flat as the masking procedure should select only parts of the map that are near your model.  If this is not happening I may need to fix something.

When you use auto_sharpen with a model, by default the map will be masked around the model (mask_atoms=True).  This means that there is the possibility of some model bias, as masking removes density that is outside the model.  Auto-sharpen uses phenix.map_box to do this masking, with a default of 3 A around all atoms in the model for this mask (mask_atoms_atom_radius=3) , plus a "soft" boundary of an additional 3 A beyond that (soft_mask=True).

If you want to minimize model bias, you can set mask_atoms=False instead, or increase the radius for masking.  You can also use sharpening without the model (I recommend trying this as it is almost as good as sharpening with a model).

In model-based local sharpening, auto-sharpen carries out sharpening in one box of density at a time within the map. The boxes are chosen only in regions that contain atoms in your model, and the density for a point in the final map is the weighted average of densities for all boxes that overlap that point.  For regions where there are no atoms in your model, the density from overall (global) sharpening is used.

So to answer your key question, there should normally be minimal model bias, but it can certainly be present.  The way to determine if there is model bias is a very simple experiment:  Take your model, remove something from that model that is a small part of the model but that is very clear in the density (for a high-res map, remove a few side chains with good density for example, for a low-res map remove a helix).  Run the procedure again with this slightly modified model.  If the density changes substantially in the region where you manipulated the model...you have model bias.

All the best,

Tom T